Anthropic’s answer has been to assemble one of the largest and most heterogeneous AI compute portfolios in the United States — Measured AI estimates Anthropic’s U.S. compute footprint and capacity pipeline at more than 15 GW, anchored by 15 data center campuses, procured through five distinct contracting models, running four separate silicon ecosystems, and financed by an increasingly elaborate stack of project debt, guarantees, vendor-backed hardware financing, and strategic supplier investment.

The defining feature of the portfolio is what it lacks: ownership. Anthropic holds leases, colocation and managed services agreements, and long- and short-term compute capacity contracts. It owns none of the buildings, none of the substations, and — in nearly all cases — not even the chips.

That heterogeneity is the story. The instruments range from 20-year direct data center leases with former Bitcoin miners in Kentucky and Texas, to 10- to 15-year colocation and managed services agreements with a neocloud across multiple states, long-term dedicated compute capacity from AWS, a six-year reservation of finished NVIDIA compute at a microgrid campus in West Virginia, and merchant compute bought by the month from SpaceX in Tennessee, terminable by either company on 90 days’ notice.

Each model makes a different trade-off between duration and reversibility, balance sheet commitment and speed to market — and Anthropic is using all five at once.

This article maps Anthropic’s AI infrastructure underpinning the Claude machine — campus by campus, chip by chip, dollar by dollar.

The full Excel dataset behind this analysis — 15 campuses and more than 400 data points — is available to paid subscribers at the end of the article.

The place to start is not the campuses. It is the question they raise.

Subscribe now

I. Before the Campuses: What Has Actually Run Claude

Anthropic reached a $65+ billion revenue run rate before most of its dedicated data center campuses existed. Claude’s early growth ran on rented cloud infrastructure; the purpose-built campuses came later, in response to that demand. The buildout can be divided into four distinct eras:

Cloud-First Era (2023–2024)

Claude initially ran on rented accelerator capacity inside Amazon Web Services (AWS) and Google Cloud, making the underlying infrastructure largely invisible at the individual data center level. AWS became Anthropic’s primary cloud provider in September 2023. In January 2024, Google first listed Anthropic as an AI Hypercomputer customer, saying the company was “building and serving models on it.” In April 2024, Amazon CEO Andy Jassy wrote that Anthropic planned to use Trainium and Inferentia chips to “build, train, and deploy its future FMs [foundation models].”

Dedicated-Hyperscale Era (2025)

In February 2025, Amazon announced Project Rainier, a massive cluster of AWS Trainium2 chips dedicated to Anthropic. In October, Anthropic also secured access to as many as 1 million Google TPUs. Rainier came online at the end of October 2025 with nearly 500,000 Trainium2 chips spread across multiple U.S. data centers and states, including Indiana and Mississippi. By year-end 2025, Anthropic was using more than 1 million Trainium2 chips across its broader AWS footprint, while an internal OpenAI memo estimated Anthropic’s total compute capacity at 1.4 GW.

Capacity-Scramble Era (Late 2025–2026)

In November 2025, Anthropic announced a $50 billion U.S. infrastructure program with Fluidstack, committed to additional capacity on Microsoft Azure, and placed a $10 billion order for Google’s Ironwood TPUs through Broadcom. In March 2026, NVIDIA CEO Jensen Huang described Anthropic as “compute limited, factory limited.”

The focus then shifted to securing capacity as quickly as possible across a growing and increasingly diverse set of compute providers:

• April 10, 2026: Anthropic signed a multibillion-dollar, multiyear agreement with CoreWeave to provide NVIDIA-based compute for Claude, with capacity coming online in phases during 2026 across undisclosed data center sites.

• May 3, 2026: Anthropic signed cloud services agreements with SpaceX for NVIDIA GPU capacity at Colossus 1 and Colossus 2, at a combined $1.25 billion per month.

• May 7, 2026: Anthropic committed $1.8 billion over seven years to Akamai for dedicated cloud infrastructure capacity, including rented NVIDIA GPU compute.

• July 5, 2026: Anthropic’s first TPU-based compute capacity under its colocation and managed services agreements with Fluidstack came online at Lake Mariner in New York.

• August 26, 2026: Anthropic added a fifth type of arrangement, agreeing to spend approximately $45 billion over six years on NVIDIA Vera Rubin compute at neocloud Nscale’s Monarch campus in Point Pleasant, West Virginia.

Purpose-Built Era (2027–2028)

Most of Anthropic’s developer-led data center campuses — across Kentucky, Texas, Louisiana, Indiana, New York, and West Virginia — are scheduled to deliver capacity in 2027 and 2028, generally under contracts running from 10 to 25 years. Anthropic’s hardware cycle points to the same timeline: Google’s next-generation TPUs, co-designed by Broadcom, begin arriving in 2027, AMD’s first gigawatt of capacity is scheduled for the first half of 2027, and NVIDIA’s roadmap transition from Grace Blackwell to Vera Rubin is already beginning in the second half of 2026 and will continue through much of 2027.

These campuses represent the next stage of Anthropic’s infrastructure strategy: long-term, purpose-built capacity added after demand had already outgrown what Anthropic could secure through conventional cloud rentals.

The next question is how Anthropic actually secured that capacity.

Subscribe now

II. The Footprint: 15 Campuses, Contracted 5 Different Ways

Measured AI estimates that Anthropic’s U.S. compute footprint and capacity pipeline exceeds 15 GW.

The estimate combines Anthropic’s 1.4 GW year-end 2025 capacity baseline, approximately 7.9 GW of critical IT capacity associated with the 15 campuses mapped in this analysis, roughly 1 GW of Microsoft Azure capacity, approximately 3 GW of additional Google/Broadcom capacity not yet allocated to identified campuses, roughly 1.8 GW of unallocated AMD MI450 capacity, and approximately 250 MW of additional capacity under its CoreWeave and Akamai agreements, where the underlying data center locations have not been disclosed.

The 15 identified campuses account for ~7.9 GW. The remaining capacity reflects the existing-capacity baseline and commitments not yet allocated to identified campuses; these amounts are included in the pipeline total but excluded from the campus count to avoid double counting.

What organizes Anthropic’s capacity is procurement: five contracting models, namely direct data center leases, colocation and managed services agreements, long-term dedicated compute capacity, long-term neocloud compute, and short-term merchant compute agreements.

Direct Data Center Leases: Anthropic Cuts Out the Middleman

Anthropic’s direct leasing model is the simplest: Anthropic leases data center capacity directly from the developer, with no intermediary. The developer provides the space, power, and cooling, while Anthropic is responsible for supplying the chips and other compute hardware. Collectively, direct data center leases comprise ~1.1 GW of critical IT capacity for Anthropic:

These are the only data center campuses where Anthropic is directly named as the tenant, and they carry some of the longest base term commitments in its portfolio.

TeraWulf’s Justified Data Campus — Hawesville, Hancock County, Kentucky

The lease runs 20 years and $19 billion in total contract value (TCV), equivalent to $197/kW/month, with two optional 5-year extensions. If fully extended to 30 years, the contract could total $33 billion.

Source: TeraWulf.

TeraWulf bought Hawesville — a decommissioned Century Aluminum smelter with five 161 kV transmission circuits and an energized substation — for $200 million in cash in February 2026, and leased it five months later, after Anthropic won a competitive process for the site.

TeraWulf expects to deliver Justified’s 401 MW in phases, with initial capacity in the second half of 2027 and full delivery in early 2028; nothing is energized today.

Riot’s Rockdale Campus — Rockdale, Milam County, Texas

The lease runs 20 years and $9.1 billion, equivalent to $198/kW/month, also with two optional 5-year extensions. If fully extended to 30 years, the contract could total $16.1 billion.

Source: Riot Platforms.

Riot targets an initial 96 MW at Rockdale in December 2027 and the full 191 MW by June 2028, with land cleared and graded but no capacity built.

Nexus Hubbard — Hubbard, Hill County, Texas

Four separate leases cover approximately 500 MW of capacity, with initial delivery expected in Q4 2027. Google guarantees Anthropic’s lease and power obligations in exchange for approximately 20% project equity, with the guarantee becoming effective only after completion. Anthropic’s initial 500 MW commitment is the project’s first phase, requiring more than $5 billion of financing. The campus is capable of ultimately reaching about 7.7 GW of data center capacity.

Source: Nexus Data Centers.

Construction at Hubbard began in early 2025 — roughly 14 months before Anthropic signed, in March 2026. The developer built speculatively and found its tenant, Anthropic, afterward. Nexus is building its own on-site gas generation, beginning at 1.6 GW with plans to reach 7.2 GW, to run the campus entirely off-grid, skipping ERCOT’s interconnection queue.

Colocation and Managed Services: Fluidstack Holds the Lease, Google Holds the Risk

Anthropic’s colocation and managed services model is the most distinctive and the one that has scaled across counterparties. Here, the developer (e.g., TeraWulf, Hut 8, Cipher Digital) owns the campus and signs a data center lease with Fluidstack as tenant. Google signs a separate backstop with the developer, known as a “Recognition Agreement” — Google guarantees the rent — usually taking warrants in the developer as consideration. Anthropic signs a colocation and managed services agreement with Fluidstack. Anthropic is not a party to the lease and guarantees nothing.

Across five data center campuses, colocation and managed services comprise ~1.4 GW of critical IT capacity for Anthropic:

TeraWulf’s Lake Mariner Data Campus — Barker, Niagara County, New York

The agreement runs 10 years and $6.7 billion, equivalent to $155/kW/month, with two optional 5-year extensions. If fully extended, the contract could run for 20 years. The $6.7 billion TCV figure reflects the original Fluidstack agreements; contracted capacity was subsequently increased to 378 MW, and the revised aggregate contract value has not been disclosed.

The Lake Mariner site comprises a repurposed 700 MW coal plant with dual 345 kV transmission feeds.

Source: TeraWulf.

Fluidstack signed three leases with TeraWulf at Lake Mariner: CB-3 at 42 MW (delivered early July 2026), CB-4 and CB-5 at 168 MW each, with first data halls expected in late September 2026 and early January 2027, respectively. TeraWulf also funded ~$150 million of Fluidstack’s fit-out, to be recovered as more than $300 million of rent over the initial 10-year term.

Hut 8’s River Bend Campus — St. Francisville, West Feliciana Parish, Louisiana

The agreement runs 15 years and $7 billion, equivalent to $159/kW/month, with three optional 5-year extensions. If fully extended, the contract could run for 30 years.

Source: Hut 8.

Hut 8 targets River Bend’s first 65 MW data hall in Q2 2027, with three more 60 MW halls following at roughly 60-day intervals across the balance of the year.

More broadly, Hut 8 frames River Bend as the first tranche of a December 2025 partnership with Anthropic and Fluidstack covering up to 2.295 GW. Only the 245 MW is contracted with Anthropic: the 1 GW right of first offer on the next phase belongs to Fluidstack as tenant, and the remaining 1.05 GW reflects Hut 8 and Anthropic’s intention to jointly evaluate other Hut 8 sites.

Cipher Digital’s Barber Lake Data Center — Colorado City, Mitchell County, Texas

The agreement runs 10 years and $3.8 billion, equivalent to $154/kW/month, with two optional 5-year extensions. If fully extended, the contract could run for 20 years.

Source: Cipher Digital.

Anthropic’s 207 MW at Barber Lake was assembled in two steps: a September 2025 lease for 168 MW of critical IT, then a November amendment adding 39 MW — the site’s last unallocated interconnection. Cipher Digital targets delivery of Barber Lake’s 168 MW first phase on September 30, 2026, with the remaining 39 MW of the second phase due by January 31, 2027.

Fluidstack’s Abernathy HPC Campus — Abernathy, Hale County, Texas

The agreement runs 25 years and $9.5 billion, equivalent to $188/kW/month, with options to shorten the term to 20 or 15 years. TeraWulf, the original majority JV owner, sold its 50.1% stake in Abernathy to an investor group led by Fluidstack for ~$530 million in July 2026.

Source: Aligned Data Centers.

Abernathy was originally guided by TeraWulf to a second-half 2026 delivery and Q4 2026 lease commencement. Aligned Data Centers owns the broader ~120-acre land parcel at Abernathy. Of this total, Aligned has ground-leased 29.3 acres to the Fluidstack-owned project company, on which the 638,000 square foot data center will be built.

Meridian Arc — New Lebanon, Sullivan County, Indiana

The agreement runs 15 years and approximately $13.3 billion, equivalent to $172/kW/month, with three optional 5-year extensions. If fully extended, the contract could run for 30 years. The campus, developed by the Meridian Compute JV, comprises two buildings, 245 MW and 185 MW, on ~140 acres.

Source: Board of Commissioners of Sullivan County, Indiana.

Meridian Arc’s first 65 MW data hall is planned for July 2027, with lease commencement phased from August 2027 and full commencement running from late 2027 into early 2028.

The Meridian Compute joint venture that owns the campus is a partnership between Fluidstack and Next Frontier, Coatue’s data center venture. Coatue co-led the February 2026 round valuing Anthropic at $380 billion and has advised it on its data center strategy. This is an example of an Anthropic investor and adviser sitting on the sponsor side of an Anthropic data center campus.

Subscribe now

Long-Term Dedicated Compute at AWS: Amazon Owns Everything

Anthropic’s dedicated hyperscaler compute capacity model with AWS is a portfolio-level commitment of more than $100 billion to use AWS’s AI infrastructure. Amazon owns the buildings, the substations, and the chips — Trainium2 — and Anthropic buys AWS cloud services under a 10-year portfolio agreement. Collectively, AWS’s dedicated compute comprises ~4.45 GW of capacity for Anthropic:

Source: Amazon.

Long-Term Neocloud Compute at Monarch: Six Years of Vera Rubin

Anthropic’s newest contracting model arrived on August 26, 2026. Under a long-term neocloud compute agreement, Anthropic will spend approximately $45 billion over six years on 460 MW of NVIDIA Vera Rubin compute at Nscale’s Monarch Compute Campus in Point Pleasant, Mason County, West Virginia, with delivery beginning in late 2027. Nscale — a UK-based neocloud backed by Aker ASA, NVIDIA, Dell, and Nokia — finances, builds, owns, and operates both the data center and the chips, and sells Anthropic the finished compute. At the headline $45 billion figure, it is one of the largest individual infrastructure commitments in Anthropic’s portfolio to date.

Source: Nscale.

Anthropic takes the first of three planned buildings — roughly a third of a campus designed for 1.35 GW in its first phase across more than 2,250 acres, with the remaining buildings expected to begin providing capacity in 2028.

Source: West Virginia Department of Environmental Protection (WVDEP).

Nscale plans more than $20 billion of campus investment, an expansion path beyond 8 GW, and a grid-independent, on-site microgrid. Planning documents put the full development cost at approximately $71 billion, $47 billion of it GPUs — a chip bill that lands on Nscale, not Anthropic.

On a straight-line, fully ramped basis, the agreement implies $1,359/kW/month. That is a finished-compute price — it buys installed Vera Rubin systems, networking, facilities, and operations — and it is not comparable to the roughly $154–$198/kW/month Anthropic pays for powered shell and colocation capacity elsewhere in the portfolio, where the chips are financed separately.

Structurally, the agreement sits apart from the rest of the portfolio: a six-year committed reservation of dedicated capacity is not the rolling, 90-day-terminable merchant compute Anthropic buys from SpaceX, and a single-campus compute contract with a neocloud is not the 10-year, $100+ billion portfolio relationship it maintains with AWS — while, unlike at the Fluidstack campuses, Anthropic holds no colocation agreement and leases no racks.

Short-Term Merchant Compute at Colossus: Reversible in 90 Days

Anthropic’s agreements with SpaceX are the most flexible — and least durable — part of its infrastructure portfolio. Signed on May 3, 2026, they give Anthropic access to roughly 500 MW of compute capacity across Colossus 1 and Colossus 2 for $1.25 billion per month through May 2029.

Source: SpaceX.

Anthropic’s package includes at least 325,000 NVIDIA GPUs, along with CPUs, storage, and networking, all purchased and owned by SpaceX. That GPU fleet comprises NVIDIA H100, H200, and B200 GPUs at Colossus 1, and NVIDIA B200 GPUs at Colossus 2.

Source: SpaceX.

The headline three-year value is approximately $45 billion, although actual payments could be lower because of discounted ramp pricing and early termination rights. The SpaceX deal represents finished compute rather than Anthropic’s own dedicated data center capacity. The NVIDIA GPUs reside in SpaceX’s Colossus data centers spanning Memphis, Shelby County, Tennessee and Southaven, DeSoto County, Mississippi.

Source: SpaceX.

After an initial three-month period, either Anthropic or SpaceX can terminate on 90 days’ notice, effectively turning the arrangement into a rolling quarterly commitment. The capacity is also not dedicated exclusively to Anthropic: SpaceX can reallocate it to its own workloads, including Grok; other customers, such as Cursor (owned by SpaceX), share the clusters. That makes Colossus the fastest capacity in Anthropic’s portfolio to secure, but also the easiest to lose.

Five contracting models, one buyer. Across all five runs a second organizing layer: the chips.

III. Four Kinds of Chips, One Claude

Anthropic is deploying Claude across four silicon ecosystems: Amazon Trainium2 at the four AWS campuses; Google TPUs, including Ironwood, at Justified, Nexus Hubbard, and the five Fluidstack sites; NVIDIA H100/H200/B200 GPUs at Colossus, with Vera Rubin coming to Nscale’s Monarch in late 2027; and AMD MI450 GPUs at Rockdale.

Who Owns the Silicon

Who owns or finances the chips depends on the contracting model:

1. At the direct-lease campuses, Anthropic is responsible for the compute hardware installed inside the developer’s shell.

2. Chip financing and ownership at the five Fluidstack sites runs through Broadcom’s new AI XPV Platform. An Apollo-managed fund, WarehouseCo Intermediate Holdings LP, carries the purchase obligation under the $35 billion initial tranche, with Athene guaranteeing 15% of the outstanding obligation. Anthropic leases the racks, while Fluidstack operates the infrastructure at the five campuses. Broadcom provides residual-value support on roughly $30 billion of senior financing: if Anthropic defaults and the racks cannot be sold for enough to repay the senior investors, Broadcom covers the shortfall.

3. Amazon owns the Trainium fleet at its AWS campuses.

4. Nscale owns and operates the Vera Rubin systems at Monarch and sells Anthropic the resulting compute capacity under the six-year agreement.

5. SpaceX owns the GPUs and other hardware at Colossus and sells Anthropic the resulting compute capacity.

The Supplier-Investor Flywheel

Anthropic’s ownership-light model does not make it independent of the companies supplying its infrastructure. Increasingly, the same counterparties providing its chips and compute capacity are also investing in Anthropic, guaranteeing infrastructure built for it, financing the hardware inside those facilities, or co-engineering the platforms on which Claude runs.

The guarantees, in particular, solve a more fundamental financing problem. Anthropic, on its own, lacks the balance sheet and credit rating that lenders would need to finance data center projects of this multi-gigawatt scale. By backstopping Anthropic’s lease obligations, Google effectively places its own investment-grade credit behind the tenant, giving developers and their lenders greater assurance that the facilities will generate the contracted revenue.

Together, these four chip ecosystems reduce Anthropic’s technical dependence on any single platform, but vendor diversification is not the same as vendor independence. Anthropic owns little of the physical stack, yet it has enormous future purchase commitments to counterparties that simultaneously hold equity, provide guarantees, finance equipment, or help design the systems Claude runs on. The model is asset-light, but it is neither capital-light nor counterparty-light.

The campuses, chips, and financing structures explain what Anthropic has secured and who ultimately owns it. The next question is when those megawatts actually arrive — because the gap between contracted and energized capacity is the portfolio’s live risk.

Subscribe now

IV. The Race From Announced Megawatts to Energized Megawatts

Currently, only a fraction of Anthropic’s identified campus capacity is actually serving Claude. Project Rainier is operational in Indiana, Mississippi, and Pennsylvania; Anthropic is buying compute from Colossus 1 and Colossus 2; and 42 MW at TeraWulf’s Lake Mariner CB-3 in New York came online in early July. Nearly everything else in the portfolio is still under construction or awaiting delivery.

The Delivery Calendar: 1.4 GW in 2025, 15 GW by 2028

The next wave begins in late 2026. Cipher Digital’s Barber Lake in Texas targets its first delivery on September 30, followed by its second phase in January 2027. TeraWulf’s Lake Mariner CB-4 is expected in late September, with CB-5 now pushed to early January 2027. Fluidstack’s Abernathy in Texas is expected to begin delivering in the second half of 2026.

The larger purpose-built campuses arrive later: Hut 8’s Louisiana site begins in Q2 2027; Meridian Arc in Indiana in July 2027; TeraWulf’s Kentucky site in the second half of 2027; Nscale’s Monarch campus in West Virginia in late 2027; and Riot’s Texas campus in December 2027, reaching its full 191 MW of critical IT capacity by June 2028. Nexus Hubbard is expected to bring its first building online in Q4 2027.

Taken together, those delivery schedules imply a step change in Anthropic’s available compute capacity.

Measured AI estimates year-end capacity rising from 1.4 GW in 2025 to 5.0 GW in 2026, 9.5 GW in 2027, and more than 15 GW by the end of 2028. The 2028 figure reflects capacity already identified or implied by Anthropic’s current commitments; it is not necessarily a ceiling, as additional capacity secured between now and 2028 could push the ultimate figure higher.

Those schedules are themselves a source of risk. Several Fluidstack leases include 180-day remedies for delivery failures, while Meridian Arc still depends on major utility milestones ahead of its first data hall: 300 MW of power by year-end 2026 and another 320 MW by March 2027.

Contracted megawatts are not the same as available compute. Anthropic’s 2026 revenue growth was supported primarily by existing cloud capacity, Project Rainier, and, beginning in May, Colossus. The first capacity delivered through the Fluidstack model did not arrive until Lake Mariner’s 42 MW CB-3 in July 2026 — by which point projects using that model had already raised more than $15 billion of project debt.

The Power Underneath the Portfolio

The delivery schedules reinforce a broader point: data center development is fundamentally a power problem. Anthropic’s 15 identified campuses span PJM, MISO, ERCOT, SPP, NYISO, and the Tennessee Valley — plus, at Nexus Hubbard and Monarch, campuses designed to run entirely off-grid — but they reach the rack through materially different power strategies.

Most rely primarily on conventional utility service. Justified is supplied through Kenergy and Big Rivers in MISO; Rockdale and Barber Lake through Oncor in ERCOT; Abernathy through Xcel Energy’s Southwestern Public Service in SPP; River Bend through Entergy Louisiana in MISO; and the four identified AWS campuses through Indiana Michigan Power, Entergy Mississippi, and PPL.

More revealing is how developers secured that power in the first place. Justified occupies a former aluminum smelter with five incoming transmission lines and an energized substation. Lake Mariner reuses the site and transmission infrastructure of a retired 700 MW coal plant and also benefits from a 90 MW allocation from the New York Power Authority. Barber Lake had approvals for a 300 MW interconnection before its Fluidstack leases were completed. Meridian Arc’s development includes its own dedicated electrical substation.

These sites illustrate why developers target locations with existing or pre-secured electrical infrastructure, although that infrastructure does not necessarily mean the full contracted load can be served today.

At the other end of the spectrum, some campuses are increasingly building around the grid. Hubbard plans an initial 1.6 GW of on-site gas generation, while Colossus 1 combines utility power with behind-the-meter generation and batteries; Colossus 2 is designed around gigawatt-scale on-site gas generation. Monarch in West Virginia goes the furthest: Nscale is building the campus around a fully grid-independent, on-site microgrid, with an initial 2 GW of power capacity expected by the first half of 2028.

How Much of the Pipeline Has Power Today

Of the ~2.5 GW that remains undelivered across Anthropic’s eight leased and colocation campuses, only ~30% sits behind power that is already approved and energized. The remaining ~70% still depends on infrastructure that has not been completed. About 1.25 GW of that outstanding capacity depends on utility work — transmission, substations, or switchyards — while the 500 MW at Nexus Hubbard depends on an on-site gas plant that still has to be built. Monarch’s 460 MW sits outside that count entirely — and entirely behind on-site generation that has not yet been built.

The operating AWS campuses are not fully de-risked either. New Carlisle already serves a substantial ~1 GW load, but further expansion still depends on new transmission infrastructure and permits; Canton has more than 500 MW sitting behind a substation scheduled for 2027; Ridgeland’s later phases require additional substations and generation; and at Susquehanna, PPL’s dedicated 500/230 kV lines are not scheduled for completion until May 2028. In other words, these energized campuses still have much of their ultimate capacity waiting on the grid.

The implication is simple: Anthropic’s data center portfolio is also a portfolio of power rights, interconnections, and generation infrastructure. The lease can be signed, the building financed, and the chips ordered, but none of it becomes compute until electricity reaches the rack. A signed 2028 lease is not compute available today.

Subscribe now

The dataset behind this article is available below as an Excel download for paid subscribers. It tracks all 15 identified campuses across more than 400 data points, including location and coordinates, power market and utility, developer and compute intermediary, contracting structure and terms, critical IT capacity, chip type, site acreage, buildings and square footage, contract value and pricing, and announced capital expenditure, where disclosed. This flagship analysis is free and stays free; the underlying workbook is available to paid subscribers.

Measured AI provides institutional-grade analysis of the physical infrastructure powering AI data centers. For access to our full research library, regulatory intelligence, and weekly briefings, visit MeasuredAI.com.

Source: Talen Energy.

The rejection did not stop the campus. It re-architected it. Seven months later, on June 11, 2025, Talen Energy — the plant’s majority owner and the campus’s original developer — announced that the behind-the-meter arrangement was gone.

Instead, Susquehanna would sell its output into the PJM grid, while Talen would serve Amazon as a retail supplier over the wires of PPL, the local electric utility. The agreement would run at a full contracted capacity of 1.92 GW through 2042 — twice the 960 MW that the co-location design had ultimately aimed for — and Talen estimated it could be worth about $18 billion over the life of the contract.

All of it sits on one piece of ground: roughly 1,200 acres beside the Susquehanna River in Salem Township, Luzerne County, in northeastern Pennsylvania — PPL’s territory, inside the PJM grid.

Susquehanna both defined the co-location era and became its clearest counterexample. This is the campus that made nuclear-adjacent AI data centers a national regulatory fight — the test case that FERC’s vocabulary, PJM’s large-load policy, and Pennsylvania’s tariff framework were all built in response to. Then it abandoned co-location and got bigger by doing it.

This primer traces the Susquehanna campus from Talen’s original nuclear co-location strategy and $650 million sale to Amazon through Amazon’s data center build-out, the FERC fight, the 1.92 GW power deal that followed, the resulting grid and tariff consequences, and the local land-use and environmental constraints shaping the campus and its surroundings.

Subscribe now

I. A Power Company Built This Campus to Sell It

Before Susquehanna was Amazon’s campus, it was a power company’s product. Talen did not attract a data center to its nuclear plant; it built one — assembled the land, erected the first shell, ran a crypto mine as the anchor tenant, and marketed the package to hyperscalers as the fastest route to nuclear-scale power. The AI demand wave was the exit, and Talen took it.

Building the Campus Before a Buyer Arrived

The assembly started early, and the decision earlier still. Talen launched the Cumulus venture in October 2019 and began work in the spring of 2020 — three and a half years of building, begun with no buyer under contract. Susquehanna Data, LLC — a Talen vehicle — had begun obtaining the necessary permits by mid-2021. In March 2022, Cumulus Data, LLC — the Talen subsidiary that would own the buildings — mortgaged 13 parcels to Orion Energy Partners, an infrastructure investment firm, to fund the build.

Two months later, Talen itself filed for bankruptcy in a cash crunch brought on by surging gas prices — and kept assembling the campus through Chapter 11, repaying its first-lien debt in full and converting its unsecured notes into equity before emerging in May 2023 owned by its former creditors.

The anchor tenant came first: Nautilus Cryptomine, a bitcoin-mining joint venture with TeraWulf rated at up to 200 MW gross, with 150 MW submetered from March 2023. The first data center shell went up beside it — 48 MW of what Talen called “processing load,” expandable to 65 MW with more cooling.

Source: TeraWulf.

In 2023, Talen positioned the site as a 1,200-acre campus directly connected to the Susquehanna nuclear plant, with four substations providing up to 956 MW of potential redundant capacity, a fiber network with entries from three directions, and the first data center ready for fit-out.

Source: Talen Energy.

Substations #1 and #2 draw from separate nuclear reactors — the redundancy the whole design rested on — while #3 and #4 serve the data centers and the former Nautilus Cryptomine.

The $650 Million Sale — and What the Price Bought

Talen announced the sale on March 4, 2024: substantially all the assets of the Cumulus Data campus sold to Amazon Data Services, Inc. for $650 million — $350 million at closing, $300 million held in escrow against zoning milestones. Of the $650 million purchase price, the agreement allocated just $90 million to the real property — land, buildings, and easements — and the remaining $560 million to contracts, government permits, and equipment. The returns were significant for Talen: a $324 million gain on sale and a greater-than-2.5x multiple of invested capital by management’s account. That price came out of an auction among four of the world’s largest companies by market value, with Amazon the winner.

The sale had in fact closed on March 1, 2024, with the deeds recorded on March 13. Nautilus Cryptomine initially remained on the property under its existing ground lease. Mining ceased on September 30, 2024, and the following month Talen bought TeraWulf’s 25% interest in the venture for $85 million in cash. The transaction gave Talen full ownership of Nautilus, while TeraWulf received the mining fleet as the in-kind portion of the consideration. Talen stopped using the building in June 2025 under the revised power arrangement, and the ground lease was terminated that September, bringing the site’s crypto-mining era to an end.

One distinction matters throughout what follows: Amazon Data Services, Inc. owns and permits the campus, while Amazon Energy, LLC is the wholesale power buyer at FERC.

Both entities point back to the same asset — the ground beside the plant — and that ground had been readied for gigawatt-scale loads long before Amazon arrived.

II. The Ground: A Nuclear Site’s Second Act

What Amazon really bought was a site next to a nuclear plant. That location matters in two ways: the campus sits on land once intended for a nuclear expansion that was never built, and the purchase agreement leaves open the possibility of another reactor being built there in the future.

Source: Talen Energy.

The Plant and the Reactor That Never Got Built

Susquehanna Steam Electric Station is two separately operable nuclear units of roughly 1.25 GW each — 2.494 GW of summer-rated capacity between them, commissioned in 1982 and 1984 and licensed to 2042 and 2044. Talen holds 90%; Allegheny Electric Cooperative holds the other 10%. Two units, not one machine: that is the fact governing everything downstream, from the megawatt ladder at FERC to the contracted capacity Talen supplies today.

The site’s industrial zoning dates back to Bell Bend, a nuclear project that was never built. In 2011, Salem Township created the I-3 Special Industrial District for PPL’s proposed 1.6 GW nuclear plant on 2,055 adjoining acres. Nuclear Regulatory Commission staff recommended issuing the license, but the Army Corps withdrew Bell Bend’s wetlands permit application in February 2016 over problems with its alternatives analysis, and Talen withdrew the reactor application that August.

Rendering of the proposed Bell Bend nuclear plant beside Susquehanna’s two existing units; Bell Bend was never built. Source: U.S. Nuclear Regulatory Commission (NRC).

The reactor disappeared; the heavy-industrial zoning remained and was later inherited by the data center.

Amazon’s purchase agreement also preserves a path for another reactor. Under a 2021 agreement, Allegheny Electric Cooperative retained an option to participate in a nuclear reactor project pursued by the buyer on the property; Talen was required to buy out that right by March 1, 2025, or pay $500,000, and the seller’s liability cap would rise by $30 million if the cooperative invoked it. Amazon and Talen have since said they will explore small modular reactors (SMRs) at the site.

What Amazon Is Building: 20 Data Centers on 1,200 Acres

Amazon’s hyperscale data center campus, known as PHL 100, uses 1125 Electron Avenue, Berwick, Pennsylvania, as its primary address. The campus is located at the intersection of Cumulus Boulevard and Salem Boulevard in Salem Township, within the Berwick area of Luzerne County in northeastern Pennsylvania.

When fully built out, Amazon’s campus is expected to include 20 data center buildings across approximately 1,200 acres.

Source: Pennsylvania Public Utility Commission (PUC).

Amazon began construction in October 2025. As of August 2026, three data center buildings adjacent to substations #3 and #4 were occupied and operational. Immediately to their south, at least two additional buildings were in structural/shell construction, while sitework and civil works were underway for three more, including preparation of their building pads.

Talen previously envisioned the campus in phases: Phase 1 would comprise three data centers totaling 200–240 MW, while Phase 2 would add three more immediately to the south, bringing the campus to six data centers and 400–480 MW. A further expansion to the west would have brought the campus to about 12 data centers and up to 950 MW.

The campus is expected to support approximately 1,000 permanent jobs at full build-out. By the end of 2025, Amazon had already invested $2.7 billion in Salem Township, Luzerne County, through capital and operating expenditures, supporting an estimated 4,750 annual full-time-equivalent jobs across the local economy and contributing $849 million to county GDP.

The Fiber Network: Existing Connections, New Routes

Fiber was already part of the campus before Amazon arrived. FirstLight and Lumen were both publicly associated with the original Cumulus data center build: FirstLight said in 2021 that it was serving the Berwick campus with fiber, while Lumen described itself as a network partner in the project. By early 2023, Cumulus said multiple fiber routes were in service.

FirstLight continues to market lit and dark fiber routes directly from the Berwick site. Its network map below shows how that connection ties into its broader Pennsylvania backbone.

Source: FirstLight.

The newer build-out is regional. Lightpath announced a 130-mile Eastern Pennsylvania network in August 2025, with plans to expand beyond 400 miles, explicitly citing AI data center growth and the Susquehanna region’s power infrastructure. By July 2026, it had signed a second anchor customer to a 10-year deal for a multi-gigawatt campus. Separately, FirstLight completed a 280-mile Eastern Pennsylvania network for an unnamed major hyperscaler in 2025, using its existing statewide backbone.

Permitting has moved more slowly. Lightpath’s water-crossing applications covering Berwick, Salem Township, and other Luzerne County municipalities had not passed Pennsylvania Department of Environmental Protection completeness review by mid-August 2026. Lightpath markets the network as fully underground, but preliminary drawings show its Susquehanna River crossing attached to the SR-93 bridge, about five miles from the Amazon campus. The crossing had railroad approval but was still awaiting Pennsylvania Public Utility Commission (PUC) approval, past Lightpath’s original mid-2026 completion target.

The drawings also provide a physical measure of the build. At the SR-93 crossing, two six-inch conduits are designed to carry separate 1,728-strand and 432-strand cables — 2,160 individual fiber strands in aggregate — with additional internal duct space.

Whatever workload lands there, the design premise was always the plant next door — and scaling that premise is what set off the fight.

Subscribe now

III. The Fight: Co-Location’s Test Case, Decided on the Narrowest Ground

The fight that made Susquehanna famous was never over whether a data center could run behind the plant’s meter — that was already allowed. It was over the terms of scaling it up, and the ruling that ended it decided almost nothing while changing almost everything.

The Ladder: 300 MW Allowed, 480 MW Asked, 960 MW Intended

The arrangement had a three-step permission structure. A FERC-accepted interconnection agreement already authorized 300 MW of co-located load, specifically 150 MW behind each of the plant’s two nuclear units. And that standing allowance is what the campus actually ran on throughout the fight.

Source: Federal Energy Regulatory Commission (FERC).

The June 2024 amendment, filed by PJM itself, asked for 480 MW: a ceiling set by physics as much as commerce, because the system studies found that any co-located load addition in excess of 480 MW would result in generation deliverability violations on the transmission system.

Above that sat the stated ambition — 960 MW, the full campus. The commercial architecture underneath ran in 120 MW annual steps, with a one-time Amazon option either to cap at 480 MW or to step up to 960 MW, over an 18-year initial term with two 10-year extensions. Of the 300 MW allowed, only 148 MW was ever taken, all from Unit 1 as a derate, a corresponding reduction in the unit’s output available to the grid. Unit 2 was never derated.

Exelon and American Electric Power (AEP) protested, and their headline claim set the terms of the national debate: co-location at the 480 MW scale could shift “as much as $140 million per year” onto other customers — 98% of it transmission charges, priced off PPL’s retail tariff. Talen called it “a headline-grabbing allegation that is entirely fabricated,” arguing the figure represented at most the utility’s lost opportunity costs.

The Rejection: Six Provisions, Three Commissioners, Narrow Grounds

On November 1, 2024, with two of its five commissioners recused, FERC rejected six non-standard provisions in the proposed co-located data center arrangement because they did not meet the required “necessary” standard, while declining to resolve broader questions about transmission costs, load status, and ancillary services.

Phillips, the lone dissenter, argued that the first-of-its-kind arrangement should have been approved with reporting requirements, emphasizing its reliability implications and the importance of electricity access for U.S. AI leadership. The broader concern was precedent: FERC appeared to keep the ruling narrow precisely because its treatment could shape similar data center arrangements across the country.

The Aftermath: No Precedent for the Parties, New Rules for Everyone

Susquehanna Nuclear, a Talen Energy subsidiary, challenged FERC’s November 2024 decision, but ultimately failed: its rehearing request was denied, and it later voluntarily dismissed its appeal in federal appeals court in January 2026.

More consequentially, FERC shifted to a broader rulemaking process, finding in December 2025 that PJM’s tariff did not clearly or consistently address co-located large loads and ordering PJM to develop new rules. By June 2026, FERC had expanded the framework beyond co-location through a new category called Eligible Load, while PJM separately adopted a 50 MW threshold for new large-load procedures amid growing capacity concerns.

The result was that the Susquehanna case itself produced little direct precedent, but it helped trigger the broader regulatory framework now governing large data center loads and behind-the-meter generation.

IV. The Reversal: Twice the Megawatts, None Behind the Meter

Losing the amendment did not shrink Amazon’s power supply. It doubled it — because what capped the deal at 960 MW was never the plant’s output; it was the co-location architecture.

Source: Nuclear Energy Institute (NEI).

The New Architecture: Everything to the Grid, Amazon Over the Wires

Talen’s June 11, 2025 announcement replaced the behind-the-meter co-location structure outright: Susquehanna sells its full output into the PJM grid; Talen Energy Marketing serves Amazon as its retail electric generation supplier; PPL delivers.

The power purchase agreement (PPA) provides for a full contracted capacity of 1.92 GW through 2042 — with the ramp reaching full contracted capacity no later than 2032, the old 480 MW cap option surrendered, and the power deliverable to other Amazon sites in Pennsylvania.

What Amazon holds under that structure is contracted capacity served as grid load rather than a direct draw on the plant. The nuclear attribution was contractual from the start: alongside the March 2024 sale, Susquehanna Nuclear signed a separate Amazon Environmental Attribute Purchase Agreement, under which Talen would receive additional revenue from Amazon for the carbon-free energy the plant sells to the grid.

The mechanical reason the deal could double is that co-location had held the second of the plant’s two roughly 1.25 GW units in reserve as redundancy for the first. Talen could still offer that unit into the energy and capacity markets; what it could not do was commit it to Amazon. Going front-of-the-meter converted a reserve into a saleable asset — “we also unlock our second nuclear unit previously held in reserve as redundancy,” as Talen put it at the announcement, “which now allows us to expand the PPA to nearly 2 gigawatts.”

The contract is worth about $18 billion over its roughly 17-year term. Because contracted capacity ramps over time, annual revenue rises with it; at the full 1.92 GW contracted capacity, the deal is worth roughly $1.4 billion a year — an implied price of about $80/MWh.

The Ramp: 1.92 GW of Contracted Capacity vs. 1.44 GW of Campus Load

The figures below track two different measures: Talen’s PPA ramp, which can supply up to 1.92 GW across multiple Amazon Web Services (AWS) sites in Pennsylvania through 2042, and PPL’s projected load for the Susquehanna/Berwick campus itself. PPL expects that facility to reach 1.44 GW by 2030. The two series are not directly comparable: one is Amazon’s broader contracted capacity, the other the physical load planned for this campus.

The transition from behind-the-meter to front-of-the-meter service happened in spring 2026. In April, Talen changed the facility configuration and moved to the revised AWS power purchase agreement during Unit 1’s refueling outage. Separately, FERC had already approved the restoration of 148 MW of Unit 1 capacity interconnection rights, effective June 1, 2026.

The $18 Billion Contract: Who Bears What Risk

That $18 billion is Talen’s valuation of nominal cash flows under the minimum contractual ramp. Faster delivery toward the full 1.92 GW would provide additional upside, but Talen has not disclosed the contract’s value under the maximum ramp scenario.

The agreement also does not require Amazon to buy the full amount of power. Minimum commitments rise through 2032, and if Amazon falls short, it must compensate Talen for the difference between the contract price and the market price. But that payment is capped: Talen discloses that if Amazon took no power at all, its maximum obligation would equal about 50% to 65% of the contract’s full financial impact. Talen therefore retains the residual market-price risk beyond that cap, and it posts letters of credit to Amazon under the PPA. It nonetheless expects the agreement to make its earnings more predictable.

One unresolved issue is plant availability. Talen owns 2.245 GW of Susquehanna, and the AWS PPA reaches its full contracted capacity of 1.92 GW — but the units undergo refueling outages in alternating years, typically in the spring, and Talen calls the outage “our largest recurring maintenance project.” Unit 2 was down from March 25 to June 4, 2025; Unit 1 from March 23 to May 4, 2026.

Source: Talen Energy.

With one unit offline, Talen’s 90% share of the running unit is about 1.12 GW, roughly 800 MW short of the full contracted capacity — a shortfall Talen would have to cover from its other generation or from the market. The gap is not binding yet, because the ramp does not reach full capacity until 2032 — but it widens with every step of the ramp.

Talen’s disclosures do not explain how the 1.92 GW is reconciled with those outages or with the plant’s PJM Capacity Performance obligations.

2042 Is Not an Arbitrary Date

The PPA runs through 2042, the same year Susquehanna Unit 1’s operating license expires; Unit 2’s license expires in 2044. That timing is deliberate. The original behind-the-meter agreement had an 18-year term with two 10-year extensions tied to future license renewals. Amazon aligned the new contract with Unit 1’s currently licensed life; anything beyond 2042 would depend on a renewal.

The federal nuclear production tax credit also expires in 2032, when the PPA is scheduled to reach full contracted capacity. The credit can provide up to $15/MWh when power revenues fall below an inflation-adjusted threshold, but Talen received none in 2025 because market prices stayed above it. Talen now presents the AWS contract as reducing its reliance on the credit, which provides no benefit when power prices are strong.

Source: Talen Energy.

The agreement also offers less expansion potential than the headline announcement suggests. Amazon and Talen have agreed to explore SMRs and possible uprates, but management has said further output increases from the existing plant are likely to be modest, and SMRs have been framed as a longer-term possibility rather than a current capital commitment. The PPA is capped at 1.92 GW, so any additional Amazon demand in Pennsylvania would fall outside the current contract.

Subscribe now

V. The Grid Consequences: New Wires, New Load, and New Rules

Co-location’s pitch was that the campus would barely touch the shared system. Ending it moved the campus into everyone’s systems at once — PJM’s transmission plan, its load forecast, and a PPL rate class that did not exist until this docket forced one.

The Wires: $19.4 Million for the Campus, $567 Million Around It

On April 1, 2025, PPL presented Amazon’s Susquehanna data center campus to PJM as a new large-load connection at Berwick, served by dedicated 500 kV and 230 kV transmission lines from PPL’s Susquehanna switchyards.

Source: PPL Electric Utilities.

Amazon requested service beginning in April 2026, but PPL projected completion for May 30, 2028, leaving a roughly 25-month gap. That mismatch did not mean the campus would lose power, however, because it had been drawing electricity through the plant’s own switchyards since early 2024.

The project also involves three separate layers of construction, each on a different track and each with its own cost responsibility:

At $19.4 million, the campus interconnection is well below comparable large-load projects in PPL’s 2025 Local Plan, which range from $73.5 million to $242 million — largely because it connects at a substation built for the nuclear plant.

PPL’s connection diagram below shows what that money buys: three new “customer yards” — two at 230 kV, one at 500 kV — tapped off the plant’s existing switchyards.

Source: PPL Electric Utilities.

The broader regional upgrades are far more expensive: PJM plans $567.4 million of transmission work, including $213.6 million at the Susquehanna substation, with about 60% of the cost assigned to the PPL zone and 40% to other areas. Capacity at the site is also tightening: a year later, PPL said the Susquehanna T10 and Glen Brook 230 kV yards could not be expanded enough to serve data center developer QTS’s planned Salem Township campus, which is seeking about 1.5 GW by 2033.

PPL’s Data Center Load Forecast Surges

PPL’s territory had essentially no data center load before 2024, with summer peak demand remaining roughly flat for more than a decade. For its 2026 forecast, PPL submitted more than 15 GW of proposed large-load growth, counting every project with a signed authorization at 100% probability; PJM cut that to about 6.3 GW, but the increase was still enough to make PPL one of the fastest-growing zones in the PJM system.

The Susquehanna campus is part of that shift: PJM records specifically reference its load, and PPL’s monthly adjustment begins in April 2026, the same month the campus requested grid service. Even after PJM’s reduction, PPL’s forecast peak rises to 14.4 GW by 2046.

Source: PPL Electric Utilities.

LP-6: The Rate Class Invented for This Load

PPL created a new rate class, LP-6, for large data centers after consumer and industrial groups pushed for stronger protections in its 2025–2026 rate case. Effective July 1, 2026, the tariff applies to large data center customers that began service on or after October 1, 2025. Its key requirements include:

• Load Threshold: At least 50 MW of peak demand at a single facility, or 75 MW across facilities within 10 miles.

• Contract Term: An electric service agreement with a minimum term of 10 years.

• Minimum-Load Requirement: Customers initially pay for at least 80% of committed load, falling to 50% in Years 6–10.

• Interruptible-Service Option: Customers choosing interruptible service face lower minimum-load requirements of 60% initially and 30% in Years 6–10.

• Universal-Service Costs: Beginning January 1, 2027, the class will be allocated $11 million per year in universal-service costs.

LP-6 also requires customers to provide financial security for grid upgrades, pay upfront for directly assigned interconnection work, and pay an exit fee if they leave early. For the Susquehanna campus, that exit exposure could reach the high hundreds of millions of dollars: at its planned 1.44 GW load, five remaining years at LP-6’s 50% minimum-load requirement would come to roughly $806 million of transmission charges at PPL’s current $18.663/kW-month rate.

The Statewide Model Tariff: An 80% Floor and Four Years’ Notice

Pennsylvania’s PUC adopted a statewide model tariff for large electricity users on April 30, 2026 — a template for utilities to write into their own tariffs, not a rule that binds customers on its own. Its terms set the benchmark against which Pennsylvania’s utility-specific tariffs, LP-6 among them, are now read. Key terms include:

• Large-Load Threshold: At least 50 MW of demand at a single site or 100 MW of combined demand across multiple sites.

• Minimum Billing Level: Billing demand of at least 80% of contracted capacity — or 80% of the highest demand in the past 11 months, whichever is greater.

• Notice Period: After the initial contract term, four years’ notice to reduce or terminate service.

Amazon pushed the other way on notice. Its energy-strategy principal, Michael Fradette, sought a shorter 18–24-month period; the PUC moved the opposite way, tightening its draft’s three years to four in the final order.

Behind-the-meter generation is excluded when determining whether a customer meets the large-load threshold. Amazon also indicated that shifting or curtailing workloads should not be treated as the customer’s responsibility, reinforcing the expectation that hyperscale data centers will operate as large, firm loads that utilities must plan around.

VI. The Ground Rules: $300 Million Riding on a Single Meeting

Those obligations are statewide and prospective. The constraint that actually governed this campus was local, and it had $300 million attached to it: Talen’s escrow turned Salem Township’s zoning decisions into a term of the Amazon sale.

The Zoning Deal Behind the $300 Million Escrow

Talen disclosed that the $300 million portion of Amazon’s purchase price held in escrow pending Salem Township zoning approvals was released in stages: $200 million after zoning was approved for at least 540 MW of the campus and the remaining $100 million after approval reached 960 MW. The township approved zoning for the full 960 MW on May 28, 2024, and the escrow milestones were completed that August.

The Data Center Overlay: Locked-In Rights and a Growing Footprint

Salem Township later reaffirmed that rezoning through Ordinance No. 01 of 2025, adopted at a special meeting on September 29, 2025. The ordinance created the SDC-O Special Data Center Overlay District and provides that an approved data center master plan can lock in development rights for five years, protecting the project from later Salem Township zoning changes. That master plan, shown below, divides the assembled campus into Premises A through K.

Source: Pennsylvania Public Utility Commission (PUC).

Buildings are permitted by right and can reach 100 feet in height, compared with the underlying industrial district’s 45-foot limit. The ordinance also caps noise at 70 dBA outside the overlay while exempting several power-related operations, including substations and generator testing.

Salem Township’s special data center overlay has expanded by 72.5%, from 2,308 acres to 3,981 acres, with another expansion — presented to the Planning Commission in July 2026 as 4,466 acres across 279 property owners — set for an August 26, 2026 hearing.

That is the campus’s legal ground. Its physical limits sit in other agencies’ files.

Subscribe now

VII. The Local Ceilings: Small Numbers, Negotiated Down

For a campus contracted in gigawatts, the numbers that physically govern it are strikingly small — 99 tons of NOx, 0.129 million gallons of water a day (MGD), 0.252 MGD of sewage — and each was set by negotiation rather than by demand: Amazon capped its own NOx to stay under federal major-source review, cut its own water request by more than half, and took sewer approval one phase at a time.

Air: A 515 MW Diesel Fleet Under a 99-Ton Cap

Pennsylvania has approved 109 diesel emergency generators at the campus — 283.5 MW of backup power — and an application for 86 more is pending.

Source: Caterpillar.

At full approval, the 195-unit fleet would stand at roughly 515 MW, nearly all of it 2.75 MW CAT 3516E gensets alongside a handful of smaller house and auxiliary units.

A rolling 12-month NOx limit of 99 tons a year keeps the site below the federal major-source threshold — a cap Amazon proposed itself.

The generators are regulated under Pennsylvania’s Best Available Technology standard, with no federal Best Available Control Technology determination, and the state has not yet issued the State-Only Synthetic Minor operating permit. The campus is also permitted for 204 diesel tanks holding 1.3 million gallons, although less than 30% of that storage capacity has been installed to date.

Water: 0.129 MGD, Primarily Sourced from Another County

The Susquehanna River Basin Commission authorized the campus to consume up to 0.129 MGD, measured as an average over any 30 consecutive days — a cap on the hottest month rather than the year. That figure was reduced at Amazon’s request from the 0.277 MGD recommended by Commission staff. It is also a long way from Amazon’s earlier plans: the campus was approved for 0.060 MGD in September 2024, and Amazon then sought 0.550 MGD for a fully built campus by 2030, primarily for evaporative cooling and construction. The current 0.129 MGD limit is less than a quarter of that projected full build-out requirement.

The approval allows five supply sources, but the primary one is Pennsylvania-American Water’s Berwick system, which draws from wells in neighboring Columbia County. The campus has no separate water-withdrawal approval, and its on-site well is operated below the 0.100 MGD threshold that would require one.

The cap counts only water evaporated and lost to the basin. Pennsylvania-American’s line-extension agreement provides for delivering up to 1,042 gallons per minute, about 1.5 MGD, under average-demand conditions. Water that is not evaporated is to be reused, sent to the Berwick Area Joint Sewer Authority, or — under a permit still pending — discharged to the Susquehanna River.

The Susquehanna River, which runs directly along the south and east perimeter of the campus, is not an approved campus water source for Amazon. The river withdrawal belongs to the nuclear plant: the approval, held by Susquehanna Nuclear, LLC, authorizes up to 76 million gallons of peak-day withdrawal and 53 million gallons of peak-day consumptive use to run the power station.

Amazon states that its data center campus will use water-based cooling less than 2% of the year. The Susquehanna River Basin Commission said in mid-2026 that this was the only data center water-use project it had approved anywhere in the basin.

Sewer and Wetlands: Phased Approvals and a Pending Cooling-Water Outfall

The Berwick Area Joint Sewer Authority has approved the Amazon campus in stages, beginning in June 2024 with Phase 1 and a peak sewage flow of 0.252 MGD. The authority later approved a campus agreement and signed a reimbursement agreement, while state approvals moved forward for sewage planning, a sewer extension, and an on-site pump station.

Federal wetlands review has been more limited: a 2021 determination identified 19.8 acres of federally regulated waters and wetlands but did not authorize any construction or discharge. State permits have allowed more extensive work, including stormwater controls and a 2025 water-obstruction permit covering 10 impacts and two stormwater outfalls to a Susquehanna tributary. A June 2026 amendment, still pending, would expand the permit to 16 impacts and add a cooling-water outfall into the Susquehanna River — the first proposed permit that would allow the campus to discharge cooling water directly to the river. Amazon’s separate industrial-waste discharge permit remains unissued after five deficiency letters.

VIII. From Salem Township to a Statewide Data Center Build-Out

Amazon’s Susquehanna campus did not stay singular. The combination that drew Amazon here — industrial land beside high-voltage infrastructure — has pulled other developers onto the ground around it, and Amazon has since repeated the model across Pennsylvania.

Source: Talen Energy.

Salem Township’s Expanding Data Center Cluster

Amazon’s campus is no longer alone on this ground. Salem Township’s solicitor now counts four data center projects, not one. The largest belongs to QTS, which has assembled roughly 1,700 acres across 180 lots and holds a land-development plan the township has approved — a campus seeking about 1.5 GW by 2033, off the same switchyards PPL has already said are constrained.

The result is an emerging data center cluster around the Susquehanna site, with multiple developers now pursuing the same combination of industrial land and access to high-voltage infrastructure.

What Talen proved was buildable, others are now building.

Amazon’s Broader Data Center Expansion in Pennsylvania

Amazon’s Salem Township campus is just one component of the hyperscaler’s broader AI data center expansion across Pennsylvania. On June 9, 2025, Amazon and Governor Josh Shapiro’s administration announced plans to invest at least $20 billion in data centers statewide.

Salem Township in Luzerne County and Falls Township in Bucks County were the first locations identified, with Amazon later adding Kline Township in Schuylkill County and Banks Township in Carbon County as additional areas of interest. A potential development at Homer City would further extend Amazon’s data center footprint into western Pennsylvania.

Subscribe now

Measured AI provides institutional-grade analysis of the physical infrastructure powering AI data centers. For access to our full research library, regulatory intelligence, and weekly briefings, visit MeasuredAI.com.

Source: Related Digital.

Historically, the commercial core of a utility’s contract with a data center stayed sealed by default; regulators approved the deals, and the public saw that an agreement existed but almost never its economics. Over the past 18 months, that default has been breaking: contracts and tariffs filed in Indiana, Michigan, Minnesota, and Ohio put minimum-bill floors, exit-fee formulas, and collateral tests for gigawatt-scale AI data center campuses on the public record.

What became visible is not pricing but the risk-allocation architecture: the revenue floor, the exit liability, and the collateral securing it.

I. The Number That Used to Be a Redaction Box

On February 17, 2026, DTE Energy filed the power contracts for Michigan’s largest-ever economic development project. Included in these filings was a Primary Supply Agreement between DTE Electric and Green Chile Ventures LLC, an Oracle subsidiary, which documented a month-by-month schedule of minimum monthly charges and termination payments running from December 2027 to February 2045. It shows that at full ramp, the 1.38 GW Saline Township data center operated by Oracle carries a minimum monthly charge of $19 million and a termination payment of $2.26 billion. The termination figure equals exactly 119 months (roughly 9.9 years) of minimum charges.

That level of visibility is unusual. Because special contracts for data centers are typically confidential, it has historically been difficult to determine whether utilities, with approval from public utility commissions, were shifting costs to ordinary ratepayers, such as residential customers and small businesses. Regulatory scrutiny was further limited because regulators routinely granted utilities’ requests to keep these filings confidential, while decisions approving special contracts often included little substantive analysis of the underlying economics. The Michigan filing therefore offered an unusually detailed look inside a class of agreements that had largely remained hidden from public view.

But Michigan was not an isolated disclosure. Between February 2025 and April 2026, four notable regulatory records put the commercial skeleton of the hyperscaler-utility power deal on the public record:

1. A contested Indiana tariff that Amazon, Google, and Microsoft litigated line by line

2. The Michigan special contracts behind the Oracle campus

3. A pair of Minnesota petitions carrying redacted Google contracts whose formulas are public even where every megawatt is sealed

4. A fully public Ohio tariff now under appeal at the state supreme court

The contract type is not new — public data center electric service agreements (ESAs) predate 2025. What is new is the frequent disclosure of economically meaningful terms across jurisdictions, in a pattern that is selective but mappable. Three terms organize everything that follows:

1. Minimum bill — the revenue floor

2. Exit fee — the price of leaving

3. Collateral — who guarantees payment

Those three terms are the key negotiation points between hyperscalers and utilities.

Subscribe now

II. Confidential by Default — and Why the Default Broke

Disclosure of a utility special contract has never been binary. There are three levels, and historically the public record usually stopped at the first:

The confidentiality norm was deliberate. Regulators routinely allowed utilities to shield customer-specific load forecasts, infrastructure costs, construction schedules, negotiated rates, and other commercial terms as competitively sensitive information. The commissions could review the economics; the public generally could not. That distinction mattered because the sealed terms were often the terms that determined who bore the risk if the project arrived late, consumed less electricity than promised, or disappeared altogether.

Fifty Years of Take-or-Pay

Nor are data-center-style take-or-pay contracts themselves new. Minnesota Power was negotiating individual electric service agreements with minimum-purchase obligations in the 1970s. Decades later, Google’s 2019 “Honeycrisp” agreement with Xcel showed the modern disclosure pattern almost perfectly: the executed contract was public, but its 10-year incremental-cost schedule published the structure while replacing the economic values with trade-secret markers. In a 2022 Xcel proceeding involving QTS Data Centers, the customer’s site-selection rationale, contract term, and high-level load forecast became public, while the rate and renewal economics remained sealed.

The contract form was not new. What changed was the visibility of the economics.

That change happened for two related reasons.

1. Scale Turned a Contract into Energy Policy

A 20 MW or 50 MW customer could fit within an existing resource plan without turning its bilateral contract into statewide energy policy. A 1 GW campus is different. Utilities may have to build generation, transmission, substations, or storage years before the customer’s full load materializes. At that scale, the ESA becomes the instrument that decides who bears the cost if the forecast is wrong.

What happens if a 1 GW project ultimately uses 300 MW? Who pays for the infrastructure built around the original commitment? How long must the customer keep paying? How big is the check if it walks away?

Those are no longer merely questions of confidential pricing. They are questions of stranded-cost allocation.

2. Regulators Forced the Terms into the Framework

Legislatures and commissions began forcing large-load risk protections into the regulatory framework itself.

Minnesota’s 2025 “Service to Very Large Customers” statute requires the Public Utilities Commission, when evaluating a large-customer tariff or ESA, to consider whether the agreement contains protections necessary to prevent other utility customers from bearing stranded costs associated with serving the large customer.

The same conversion was occurring elsewhere without a new statute. Michigan conditionally approved DTE’s Oracle contracts and ordered the utility to file a generally applicable large-load tariff. Indiana and Ohio moved the economics into contested tariff proceedings where Amazon, Google, Microsoft, consumer advocates, industrial customers, and utilities litigated minimum bills, exit rights, and collateral in public.

That is the shift that matters for the records that follow. The agreements themselves are not unprecedented. What is unprecedented is that economically meaningful pieces of the risk allocation are appearing across multiple jurisdictions at once — sometimes in customer contracts, sometimes in commission testimony, and increasingly in tariffs of general applicability.

The Disclosure Line Moves by State

Indiana published the tariff while sealing the executed customer agreement. Michigan exposed hard dollars while redacting parts of the underlying schedules. Minnesota published formulas while blacking out the quantities needed to calculate them. Ohio put the entire large-load tariff in public view.

Four proceedings, one architecture. The next question is what each one reveals.

III. The Four Records: One Architecture, Four Disclosure Regimes

1. Amazon — Indiana Michigan Power, IURC Cause No. 46097

The instrument in force is a settled large-load tariff for customers over 150 MW, approved with modifications on February 19, 2025, after Amazon, Google, Microsoft, the Data Center Coalition, and Indiana’s consumer advocates fought I&M’s proposal through a fully contested record to an all-party settlement. What this record uniquely contributes is the fight itself: testimony, workpapers, and a settlement that show exactly where each side gave ground.

The load driving that fight is Amazon’s $11 billion New Carlisle AI data center campus in St. Joseph County, Indiana, whose anticipated demand of 2.25 GW would make it the largest customer in I&M’s service territory.

Source: Amazon.

The executed AWS contract, by contrast, sits behind protective orders shielding Amazon’s “energization needs, financial obligations, credit support, and load ramp” — Indiana’s public numbers are tariff-level, not customer-level.

2. Oracle — DTE Electric, MPSC Case No. U-21990

DTE signed its Primary Supply Agreement and a companion Energy Storage Agreement — a distinct instrument, not an ESA — with Oracle’s Green Chile Ventures LLC on October 22, 2025, and filed the documents publicly on February 17, 2026. The Commission’s approval, granted December 18, 2025, was conditional — and it is now on appeal at the Michigan Court of Appeals, brought by environmental and renewable-energy groups and the state Attorney General. The Primary Supply Agreement covers 1.38 GW and runs through February 28, 2045 — roughly 19 years from signing, with two 10-year renewal options.

Source: Related Digital.

This record contributes hard dollars. And it is already generalizing: one condition of approval required DTE to file a standard large-load tariff, and the resulting proposal (filed March 2026) carries an 80% minimum billing demand, a termination fee of 15 years of the customer’s minimum billing demand, and collateral equal to the full termination fee — on exit and collateral, stricter than the bespoke deal it grew from.

3. Google — the Minnesota pair, MPUC Dockets E015/M-26-159 and E002/M-26-170

Xcel Energy petitioned in April 2026 for approval of its Pine Island ESA with Echo Zone LLC, a Google subsidiary; Minnesota Power had petitioned one month earlier for a 15-year ESA with Harmony Group, LLC, a subsidiary of Alphabet, Inc. Neither campus is hypothetical:

• Project Skyway, in the City of Pine Island in Goodhue County, is a 482-acre development where Google is the anchor tenant and the initial phase alone carries more than $700 million of construction

• Project Loon, in the City of Hermantown in northern Minnesota, has a development agreement providing for up to four data center buildings and 1.2 million square feet, its first phase committed at no less than $650 million

Both petitions are pending with initial comments closing September 11, 2026, for Hermantown and September 28 for Pine Island. Same customer, same statute, two utilities, one month apart — direct comparability. What the pair contributes is the formulas: every economic mechanism is public; every megawatt and dollar is blacked out.

4. AEP Ohio Schedule DCT, PUCO Case No. 24-0508-EL-ATA

The data center tariff has been in effect since July 23, 2025 — and it got there over the hyperscalers’ objection: the Commission rejected an October 2024 settlement signed by Microsoft, Amazon, Google, Meta, and the Data Center Coalition, and instead adopted, as modified, AEP Ohio’s rival stipulation with commission staff and consumer and industrial groups. The Ohio Manufacturers’ Association Energy Group has appealed to the Ohio Supreme Court, where briefing closed in March 2026, but the August 2026 oral argument was canceled on AEP Ohio’s own motion, with nothing calendared since. Schedule DCT has already governed for more than a year with no argument date set.

The objectors had campuses on the line: Central Ohio is the densest hyperscaler cluster in the Midwest, where Google has invested more than $20 billion since 2019 and Meta’s New Albany Prometheus cluster has passed 3 GW of capacity.

Source: Google and Meta Platforms.

What the tariff contributes is the generalization: a fully public instrument, litigated by the hyperscalers themselves, that converts bespoke terms into standard ones. A tariff has no customer to protect, so nothing is redacted.

The pattern across the four: the disclosure boundary follows the legal regime, not the size of the deal.

Subscribe now

IV. The Minimum Bill: What Not Showing Up Costs Every Month

Start with the settled fact: across these four records, the world’s most sophisticated energy buyers are committing to pay for 75–85% of their contract capacity whether or not the load shows up — against standard-tariff norms of 50–60%, and against Amazon’s own disclosed history of 0–75%. In dollars, the minimum bill runs to nine figures a year: $336.3 million for a 1 GW customer under Indiana’s settled tariff, and $228.4 million for Oracle at full ramp in Michigan. And the percentage alone misleads, because what it multiplies differs in every record.

The floor exists because the utility finances lumpy infrastructure against a customer that might not arrive. As AEP Ohio’s regulatory vice president put it, investments like new 765 kV transmission lines are “lumpy” — completed all at once at the start of the load ramp, so the wires exist for Year 12’s load before Year 1’s revenue does. The economically important question in every one of these records is therefore what the customer pays when it under-uses.

Indiana Cut the Percentage and Raised the Bill

I&M opened at a 90% minimum on contract capacity, against the existing 60%. Amazon called the number “unprecedented” — no other state commission, it told the regulator, had approved so one-sided a minimum — and disclosed its own book to prove it: AWS’s energy-strategy principal, Michael Fradette, testified that the company’s executed agreements run “from 0% up to 75% minimum demand obligations.” The settlement landed at 80% — of the greater of contract capacity or the customer’s highest billing demand in the past 11 months — on a 12-year minimum term.

But the percentage tells less than the rate stack: the settlement also rebuilt the base the percentage multiplies, adding a $10.959 per kW embedded-capacity charge to produce an aggregate minimum demand charge of $35.034 per kW — an annual minimum roughly $76 million higher than under I&M’s original 90% proposal. For a 1 GW customer, the settled minimum is $336.3 million a year against an expected full bill of $492.6 million — I&M’s workpaper illustration at an 85% load factor.

The Indiana lesson generalizes: the floor percentage went down and the revenue went up. And what is not disclosed is any customer’s committed capacity — the executed agreement that would turn formula into bill is sealed. What the record does show is the forecast: at the 2.25 GW the consumer advocates attributed to New Carlisle, the settled tariff implies roughly $757 million a year in minimum charges.

A forecast is not a commitment. The floor exists because the utility cannot count on that demand until a customer is contractually bound to pay for it.

Michigan Wrote the Same Floor into a Bilateral Contract

The DTE Primary Supply Agreement with Oracle applies the same 80% figure to a different base. When Oracle’s monthly on-peak billing demand falls to or below 80% of its committed-capacity ramp, DTE bills the floor on the lesser of that 80% or the maximum capacity DTE was actually capable of delivering that month — a genuine customer protection: Oracle does not pay the floor on capacity the utility could not have provided.

The Commission’s own disclosure shows how the deal compares with DTE’s standard terms: an 80% minimum billing demand, versus 50–60% in DTE’s general large-customer tariff. And the floor is inescapable by design — the load ramp’s timing can be delayed, but the maximum load cannot be reduced during the term.

Michigan’s disclosure stops at the next level of detail: the formula and the full-ramp dollars are public, while the ramp schedule itself — the base the floor multiplies in every month before full ramp — is redacted.

Minnesota Published the Formula and Sealed the Base

The Pine Island ESA’s minimum-bill clause is public in full: Google pays the greater of its actual on-peak demand or 75% of Contract Capacity, times the applicable demand charge, plus the customer charge, the tariff charges, and the Clean Energy Accelerator charges in full. Xcel’s petition states the commitment plainly: a minimum monthly bill of at least 75% of contracted demand for the full 15-year term. What the public record cannot do is turn that into dollars — the Contract Capacity, the ramp table, and every megawatt are blacked out.

The Hermantown agreement, filed one month earlier by Minnesota Power for another Alphabet subsidiary, runs the same structure through the utility’s standard Large Power Service schedule and sets the floor 5 percentage points higher: minimum billing demand is the greatest of a 10 MW absolute floor, measured demand, or 80% of Contract Capacity — with every ramp-schedule value redacted.

Procurement figures are the available proxy for the redacted load: Xcel is bringing 1.9 GW of new wind, solar, and storage onto the system to serve Pine Island, while Minnesota Power is procuring 700 MW for Hermantown. Those figures are not equivalent to campus demand, but they are the closest public indication of the scale concealed by the redactions.

Ohio Turned the Floor into a Standard Schedule

Schedule DCT splits the floor into two mechanisms that work together, and only the pair produces a bill:

1. Percentage: The minimum scales with size: 60% at 25 MW, 72.5% at 50 MW, 82.5% at 100 MW, and 85% at 117 MW and above — a deliberately more gradual ladder than AEP Ohio’s originally proposed flat 90%.

2. Base: The load-ramp capacity schedule phases contract capacity in at 50/65/80/90% across Years 1–4, so early-year billing runs against a fraction of the eventual campus.

To calculate the effective floor, multiply the “Percentage” and the “Base” percentages. For a hyperscaler (at 117 MW+), the ramp-year floor is 85% of the ramp value — an effective 42.5% to 76.5% of full capacity. A large customer’s headline 85% is far softer in Year 1 than it sounds.

Arguably the tariff’s most important effect was not revenue at all; it was demand quality. AEP Ohio filed against a speculative queue of more than 30 GW across some 90 sites that had not executed contracts. Once take-or-pay attached, the queue fell to 5.6 GW of signed ESAs under the tariff, with 13 GW across 36 sites paying for formal load studies. Take-or-pay is a demand filter, not just revenue protection.

Subscribe now

V. The Exit Fee: Pricing Abandonment

The floor guarantees revenue while the customer stays; the exit fee prices the day the customer leaves.

Utilities build assets to last 30–40 years but sign customers to 12–15-year contracts, so the exit fee is meant to cover the gap if the customer leaves — and historically, that fee has most often been the number kept under seal.

Four examples are now public:

1. Michigan protects roughly 10 years of minimum charges ($2.26 billion at Oracle’s full ramp)

2. Minnesota’s Pine Island contract caps exposure at eight years of a 75% floor — but the tariff Minnesota approved two months later has no cap at all

3. Ohio pairs a 5-year lock with a 3-year fee

4. Indiana settled between, at one to five years of remaining minimums — more than $1 billion for a 1 GW customer

Every power-supply agreement whose formula is public stops short of sizing the exit fee against the full remaining term. All of the agreements are long enough to finance transmission against, which is the point.

The commercial meaning is blunt: walking away from a gigawatt-scale site is now a balance-sheet event measured in the hundreds of millions to billions of dollars. That disciplines speculative capacity reservations, prices the option of relocating a data center campus to another state, and makes term length, notice periods, and shed rights the levers worth negotiating.

Indiana Settled the Exit Fee Publicly

I&M’s opening proposal put the number in public: a termination fee of $1.27 billion — five years of minimum billing — with no termination right at all in the first five years, a structure that put a large customer at risk for $2.55 billion. Indiana’s consumer advocates wanted a tougher exit fee — eight years of minimum bills — while the Data Center Coalition’s witness Kevin Higgins (a utility-regulatory expert with Energy Strategies) argued for a three-year exit-notice framework instead.

The settlement produced an exit fee equal to the nominal remaining minimum charges for the affected capacity over a one-to-five-year period — in the Commission’s words, more than $1 billion for a 1 GW customer that leaves five or more years early.

The rules for reducing or ending the commitment are just as revealing as the fee itself: a customer may shed up to 20% of contract capacity after Year 5 on at least 3.5 years’ notice, and any deeper reduction — even one I&M agrees to — must go back to the Commission for approval before it takes effect.

Michigan Scheduled the Exit Fee

The DTE-Oracle termination payment is not a flat amount; it follows a declining payment schedule over time. From full ramp (December 2027), it is effectively a cliff — the sum of every remaining monthly minimum through the tenth anniversary of ramp completion, starting at $2.26 billion and declining by one month’s charge each month. Then a plateau at $456.7 million — exactly two years of minimum charges — running from November 2035 through February 2043. Finally, a runway amortizing to zero at February 2045.

And if DTE is the defaulting party, no termination payment is owed by Oracle at all. For scale, DTE told the Commission the $2.26 billion termination payment far exceeds the roughly $500 million of estimated distribution and transmission upgrades.

A second exit liability sits beside the first: under the companion Energy Storage Agreement, DTE will build and own up to 1.38 GW of energy storage, bid it into the market, and charge Oracle the net cost of the fleet, not its gross capex. Oracle’s termination payment is the sum of the remaining subscription charges — roughly $3.9 billion, a figure the Commission’s order describes as an acceleration of remaining payments, not a stranded-cost estimate.

The DTE-Oracle exposure stopped being theoretical on March 27, 2026, when the Commission denied every rehearing petition, the motion to reopen, and the Attorney General’s request for a contested case — and, at the same meeting, approved the first 332 MW of the 1.38 GW fleet.

Minnesota Published the Exit Formula Itself

The Pine Island exit fee is a published formula: Contract Capacity, times the on-peak demand charge in effect at the termination notice, times 75%, times the lesser of the months remaining or eight years (i.e., 96 months) — plus any remaining clean-energy charges. Voluntary termination requires at least two years’ notice; an uncured customer breach triggers the same fee.

Even with every megawatt redacted, a negotiator learns the architecture: Xcel does not accelerate all 15 years; protection caps at eight; and the clean-energy obligation survives on top. Another mechanism prices partial retreat — the same formula run on the capacity being shed (reduced), priced at the demand charge in effect at the reduction notice, again on at least two years’ notice.

Hermantown draws the line differently: Minnesota Power’s termination mechanics are public — the window opens in Year 3 of the 15-year term, requires four years’ notice, and the fee is the utility’s “sole and exclusive remedy,” with none owed after the initial term — but the formula itself is blacked out.

Same statute, same corporate parent, a different disclosure boundary.

Notably, a tariff of general applicability came out stricter than the bespoke deal before it. Google’s Pine Island contract was filed on April 15, 2026, with a 75% floor and an exit fee capped at eight years. On June 12, the same Commission approved Xcel’s general tariff at 80% with no cap at all, and stated that those percentages are minimums, with the terms of individual agreements subject to Commission review and approval. The contract is still pending, with comments closing September 28.

Ohio Standardized the Exit

Under Schedule DCT, a customer may leave only after Year 5 following the load-ramp period, by paying three years of minimum charges — so a customer with a three-year ramp is locked in through Year 8.

The alternative is assignment: up to 25% of contract capacity may be transferred to another Schedule DCT customer (the hyperscalers’ rejected settlement had asked for 50%, with no cost-responsibility qualifications).

The magnitude was quantified early, in an August 2024 critique of the original proposal: an exit fee of more than $120 million for a 500 MW data center, $4 billion or more across 5 GW. And one pattern is too clear to ignore: the framework Higgins lost in Indiana — a minimum no greater than 70%, a three-year exit notice, a 10-year maximum term — is almost precisely the design Ohio adopted a year later. What one state rejects, the state next door may standardize.

Subscribe now

VI. The Collateral: The Credit Test the Data Center Industry Wrote

Strip away the defined terms, and every collateral section in these records asks one question: who stands behind a 15-year, billion-dollar promise to pay?

The answers come down to two basic options. Either an exceptionally strong parent — investment-grade or holding billions in cash — signs on for the term, or the customer posts security measured in months to years of bills: hundreds of millions of dollars at gigawatt scale.

That makes the critical negotiation a question of how much corporate balance-sheet capacity the power commitment consumes — and it is why Ohio’s decision to let a co-signing financial sponsor satisfy the test matters so much to anyone trying to project-finance a data center through an SPV.

Indiana Sized Collateral Off the Bill

The settled tariff requires collateral equal to 24 times the maximum expected monthly non-fuel bill, recalculated annually.

Full exemption requires a credit rating of at least A- from S&P and A3 from Moody’s, plus liquidity greater than 10 times the requirement. Amazon clears that bar comfortably, at AA from S&P and A1 from Moody’s.

An unrated customer that meets the liquidity test is exempt from half the requirement, up to $250 million — a cap on the exemption, not on the collateral.

The origin of that test is the most telling disclosure of all: the Data Center Coalition’s own witness, Higgins, proposed the exemption standard and the acceptable forms — a parent guaranty, a standby letter of credit, or cash. The settlement adopted his test and tightened his “or” into “and.” The industry wrote the credit bar it now clears. What any specific customer actually posts sits behind the same protective orders that seal the AWS agreement.

Michigan Took a Parent Guaranty and Sealed the Mechanism

Oracle Corporation, rated BBB- by S&P and Baa2 by Moody’s, guarantees its subsidiary’s payment obligations, supplemented by letters of credit on a schedule the public filing redacts. The downgrade provision is blacked out entirely — a trigger plainly exists, but its terms are not visible. Failing to maintain the collateral is itself an event of default.

On July 9, 2026, S&P cut Oracle’s credit rating from BBB to BBB-, still investment grade. DTE confirmed on its Q2 2026 earnings call that the Oracle contract carries additional collateral requirements at downgrade triggers, specifics withheld at Oracle’s request.

Minnesota Published the Tests and Sealed the Caps

Google’s Pine Island obligations carry a parent guaranty up to a cap the filing redacts, plus cash or a letter of credit equal to three months of the minimum monthly bill. The guarantor test, by contrast, is fully public: a U.S.-domiciled entity with either no credit rating but a cash balance above $3 billion, or a rating of at least BBB- from S&P or Baa3 from Moody’s. Alphabet Inc., Google’s parent company, is rated AA+ by S&P and Aa2 by Moody’s, well above those minimums. On a defined downgrade, insolvency, or guarantor default, Xcel may demand cash, a letter of credit, or other security up to its remaining credit exposure, capped at the guaranty cap.

Hermantown follows the same pattern but allows more forms of security — cash, a parent guaranty, a letter of credit, or a surety bond. The disclosure is the reverse of Xcel’s: Xcel made its guarantor standards public, while Minnesota Power redacted them, including the minimum asset level and credit rating a parent must meet to qualify. Two other provisions are notable: Google does not have to post security if its guarantor meets the creditworthiness standard, and either party may revisit the security requirements on the third anniversary of the term.

Ohio Standardized the Collateral — and Softened It Once

Schedule DCT requires, at contract signing, a guaranty or collateral equal to 50% of the total minimum charges for the full term — a parent guaranty, standby letter of credit, or cash — declining by one year’s minimum charges for each year the customer is energized and pays on time. The waiver is the Indiana test verbatim, with the same Higgins provenance behind both.

The Commission’s single substantive modification opened the project-finance door: the requirement may be met by the customer’s financial sponsor, so long as the sponsor co-signs the contract. And the actuarial basis for leniency is stated with unusual candor: AEP Ohio is unaware of any customer of 25 MW or more that declared bankruptcy in the last 10 years.

One architecture, three sizing conventions — months of bills (Indiana, Xcel), a fraction of the total commitment (Ohio, Minnesota Power), a negotiated schedule (DTE) — converging, in Indiana and Ohio, on a single waiver test written by the industry’s own witness.

The collateral is not the last backstop, and Michigan’s order says who eats a shortfall. If the Oracle guaranty and letters of credit fall short of DTE’s remaining exposure, the loss stays with the utility and its shareholders: the Commission barred recovery from ratepayers, ordered quarterly reports on Oracle’s load, credit, and financial condition, and reserved the authority to reopen the case if serving the campus begins costing other customers.

Subscribe now

VII. The Disclosure Boundary — and Where It Moves

Distinguish two things the 2025 and 2026 records changed. Public discussion of ESA economics — testimony, briefs, commission findings, dollar totals — is now routine. The contract text itself mostly is not — what stays sealed almost everywhere are the terms that made these agreements sensitive in the first place: energy pricing and discounts, guaranty caps, letter-of-credit schedules, and the ramp calendars that phase load in month by month.

But the direction of travel is unmistakable: the risk architecture is migrating out of bespoke confidential contracts and into public tariffs, and three of these four states are already there.

Ohio’s has been live since July 2025 and is under appeal at the state supreme court. Minnesota’s arrived on June 12, 2026, when the Commission created a Very Large Customer class at 100 MW and above and approved Xcel’s tariffs — followed a month later by a compliance filing carrying the final tariffs and a public form ESA, a template of the contract itself. Minnesota Power withdrew its own large-power proposal in May and refiled in July under the tighter architecture: a Very Large Power class, an 80% minimum bill, a 15-year minimum term, multi-year termination notice, and exit fees. Michigan ordered DTE to generalize the Oracle deal into a standard tariff, and the proposal is already on file.

That end state is no longer a forecast. Minimum demand, term, credit support, exit protection, and stranded-cost allocation are becoming observable and comparable across utilities through public tariffs — while the ESAs that remain confidential shrink to what the tariffs still cannot show: whose load, how big, arriving when, and where.

Measured AI provides institutional-grade analysis of the physical infrastructure powering AI data centers. For access to our full research library, regulatory intelligence, and weekly briefings, visit MeasuredAI.com.

Source: U.S. Department of Energy.

The site’s own Department of Energy (DOE)-sponsored virtual museum recalls that the switchyards alone delivered “enough to light up New York City at the time the plant was constructed in the 1950s,” and at the campus’s 2026 groundbreaking, the old plant was described simply as “the largest power draw in the world when it was in operation.” Enrichment ended in 2001. Roughly 2,400 workers are still tearing the plant down and burying its debris on site.

On March 20, 2026, ground broke on that same reservation for the PORTS Technology Campus — 10 GW of new power generation, 9.2 GW of it natural gas that will be owned by the United States government and financed from Tokyo under a trade agreement that values the build at $33.3 billion, feeding an AI data center campus that the same 10 GW is sized to yield: 8 GW of AI compute capacity.

Source: U.S. Department of Energy.

The anchor tenant is OpenAI, on a 20-year lease — and on August 17, 2026, NVIDIA guaranteed the power and the rent alike, capping its own exposure at $105 billion. And three months before the shovels, with no ceremony at all, the nuclear startup Oklo had bought 205.9 acres of the same former reservation: the land for a 1.2 GW nuclear reactor campus whose power Meta has already agreed to prepay.

The two projects share no owner and no contract — only a fence line, a common federal origin for their land, and a premise: compute follows firm power.

In the AI buildout, the scarce input is no longer land, building shells, or even chips; it is firm energy on a date certain, and the federal cleanup complex has turned out to be the one landlord in America holding power-ready land at gigawatt scale. Pike County, Ohio, is the industry’s power problem compressed into a single place — with Washington as the landlord and Tokyo as the financier.

This primer walks the two projects side by side: the scorecard and why two AI giants converged here; the Cold War power record being recycled; the gas campus built backwards — Japan’s money, Washington’s gas, and a chipmaker’s guarantee; the wires and the tariff that explain both projects’ shape; the reactor field Meta prepaid; and what one reservation says about the industry’s two theories of power.

Subscribe now

I. Two Campuses, Side by Side

Start with the scorecard: two projects on one former federal reservation, compared on who builds, who owns the generation, whose money is behind it, and when the power arrives.

The table is the article: same former reservation, same federal-land lineage, and opposite answers on ownership, technology, and time — with financing the one dimension where they are similar.

Why Two AI Giants Converged Here

Neither OpenAI’s ecosystem nor Meta arrived at Piketon for the land. Ohio has cheaper farmland closer to Columbus, and the AI industry has never lacked for shells. What it lacks is firm power on a schedule, and in the Ohio territory served by American Electric Power (AEP), that scarcity is now codified: the utility spent a year and a half refusing new central-Ohio data center connections, then locked in minimum-demand terms over the hyperscalers’ objection, and its own load studies push reliable service for new data center load toward the next decade (Section IV). Any AI company that wants Ohio gigawatts this decade has to structure around the queue, not through it.

The two camps structured around these queues, and they split along the key question in industrial strategy: make your own power, or buy someone else’s. OpenAI’s ecosystem — SoftBank, SB Energy, NVIDIA, and the government partners they assembled — chose to make it: lease federal land through the door DOE opened for AI, and burn gas now, on Washington’s books rather than their own. Meta chose to buy it: prepay a nuclear developer that had quietly assembled recycled federal land, and lock in a claim on nuclear power that arrives later — a supply-side hedge layered on top of grid load that still pays the tariff, and on top of gas plants already rising at the load those reactors will serve (Section V). One is about speed at national-industrial scale; the other is about optionality at portfolio scale.

What they agree on is the premise: whoever solves power first gets to train and run AI models first. And they agree on the landlord, because the federal cleanup complex is the rare counterparty that can offer cleared land, existing transmission corridors, water rights, and a political mandate in one package.

Source: U.S. Department of Energy.

The reason the reservation is power-ready is the half-century it spent as the power problem.

II. The Site That Ran on Two Gigawatts

The remediation story is actually the siting story. What DOE’s cleanup program spent two decades producing — cleared pads, live corridors, water infrastructure, and a landlord with statutory authority to lease — is precisely the inventory the AI industry now finds scarce, and PORTS’s Cold War power record is the template for the scale of what is being rebuilt.

The First Take-or-Pay

The Portsmouth Gaseous Diffusion Plant enriched uranium from 1954 to 2001 in three process buildings covering roughly 90 acres under roof, employing 25,000 construction workers at the peak and about 800 in operations.

Source: U.S. Department of Energy.

Powering it required inventing a utility. OVEC and its subsidiary were organized on October 1, 1952, by investor-owned utilities “for the purpose of providing the large electric power requirements projected for the uranium enrichment facilities then under construction by the Atomic Energy Commission (AEC) near Portsmouth, Ohio.” Two dedicated coal plants — Kyger Creek in Cheshire, Ohio, and Clifty Creek in Madison, Indiana, about 2.26 GW of combined capability — came online in 1955, tied to the site by high-voltage lines built for the purpose.

Source: U.S. Department of Energy.

The commercial structure looks strikingly familiar. The AEC’s 25-year contract bound the government to 1.8 GW of firm power at a monthly demand charge of $1.361 per kilowatt, owed “whether energy is used or not” — a take-or-pay minimum, terminable only on two years’ notice. The same 1956 opinion blessed a further 150 megawatts for the Portsmouth area — 1.95 GW of continuous delivery in all. On site, a switchyard complex the plant called X-533 — one of two forming a system with a rating of up to 2.2 GW — received power over four 330 kV OVEC lines and stepped the voltage down to run the enrichment machinery filling the three process buildings.

Source: U.S. Department of Energy.

And in 1969, AEP built the Don Marquis substation to provide an interconnection network between OVEC and several other entities. Hold that name: the substation built to interconnect the enrichment plant’s dedicated power supply is the same station AEP will tap to deliver the AI campus’s first phase (Section IV).

The arrangement outlived its purpose by decades. DOE gave OVEC notice of cancellation in September 2000; enrichment ceased in June 2001; the power agreement terminated in April 2003, and a settlement unwound the rest. But the physical legacy never left: OVEC and Ohio Power transmission easements across the site were extended 50 years, to 2067, and the property today resembles a small city with existing high-voltage power lines that the new infrastructure can tap — making it clear why the developers came.

Cleanup as Land Factory

The other half of the inventory is cleared ground, and it is produced on a federal demolition schedule.

Source: U.S. Department of Energy.

Above-grade demolition of the X-326 process building — the first of the three — finished in May 2024, and its 56-acre pad is where the first data center buildings go.

Source: U.S. Department of Energy.

The X-333 building began coming down in April 2025 under a cleanup task order that runs to September 30, 2031; X-330’s demolition is forecast for 2027.

Source: U.S. Department of Energy.

The waste goes into an on-site disposal facility of 10 cells, the first of them already at 99% of working capacity. The schedule runs through the leased ground itself. In 2026, DOE’s contractor is slated to excavate the soils under the old X-600 coal-fired steam plant and the X-760 chemical-engineering building and to take down four remaining support structures inside the 189 acres. The sequencing fact worth holding: the AI campus will be going up amid an active CERCLA (Superfund) demolition, for years.

Source: U.S. Department of Energy.

The lease is the mechanism that converts cleanup into tenancy. The Phase II construction area is gated “unless and until DOE has completed its remedial obligations.” The developer may fund cleanup work itself and take “a dollar-for-dollar credit via an in-kind rental payment” — the contractual form of the reported $200 million to accelerate cleanup. Interconnection is the tenant’s job at no expense or cost to DOE. The lease conveys rights to the site’s X-608B and X-6609 well fields and raw-water lines, reserves a national-security review to DOE, and runs 10 years with four 5-year renewals. The rent structure is on the record; the dollar amounts are withheld.

The Door Was Already Open

The campus arrived through a bespoke public-private partnership — in DOE’s words, “the first DOE site to reveal plans for an AI data center made possible with private investment.” The door it came through had been open since 1995, when the Southern Ohio Diversification Initiative (SODI) was chartered as the site’s DOE community-reuse organization. An 80-acre land transfer in 2018 set the precedent; more than 350 acres have followed, with an industrial pool of roughly 1,000 acres identified. The 2025–26 land run was the same pipeline at full throttle, running both mechanisms at once: DOE cleared the 189-acre Batch 1 lease to SB Energy’s data center entity on November 17, 2025, and SODI sold Oklo its parcel a month later — lease on one flank, fee simple on the other.

Source: U.S. Department of Energy.

An $80 million community benefits fund accompanies the project — SB Energy’s original $40 million, doubled by an incremental $40 million OpenAI committed on signing — and the site’s advisory board minuted that “no economic development organization or local government could have anticipated the sheer magnitude of the currently proposed investment.” The jobs claims have come in two sizes: SB Energy claims 35,000 construction and 2,500 operating jobs, while the government’s announcement counts more than 10,000 construction jobs over four years and more than 2,000 permanent.

Subscribe now

III. A Campus Built Backwards: Turbines Ordered, Power Guaranteed

Every verifiable dollar at PORTS is power and land, and every filed permit is for the data center’s ground, not yet its generation.

The Stack: Japan’s Money, Washington’s Gas

The deal architecture is a public-private partnership among DOE, the Department of Commerce, SoftBank’s SB Energy (SoftBank Group holds a 90% stake), and AEP Ohio. The $33.3 billion of Japanese money — “Deal 1” under the U.S.–Japan Strategic Trade & Investment Agreement, announced February 17, 2026 — underwrites the 9.2 GW gas fleet, which the United States government will own and SB Energy will operate.

In Tokyo’s own releases, the $33.3 billion (about ¥5.2 trillion) is an estimated project total, to be financed in tranches as Washington selects and calls for them. The first — roughly $1.9 billion, signed April 17, 2026 — pairs a Japan Bank for International Cooperation loan of up to $630 million with $1.26 billion of private-bank lending insured by Japan’s export-credit agency. This money is funneled through a Delaware conduit into a project vehicle that the two governments’ September 2025 memorandum says will be “managed and governed by the United States or its designees.” Profits split 50/50 only until Japan recoups its money, then run 90% to Washington thereafter. About 94% of the headline number has no signed instrument behind it yet.

On the data center side, SB Energy will own and operate the data center itself. Federal financing is not anticipated for the data center complex. A day after the groundbreaking ceremony, the Portsmouth Consortium formed to build it — 21 companies, 12 Japanese and nine American, spanning GE Vernova, Mitsubishi Electric, Hitachi, Toshiba, Bechtel, Kiewit, Kinder Morgan, Carrier, and a bench of banks from Goldman Sachs and J.P. Morgan to MUFG and Mizuho.

170 Turbines Against a 2028 Clock

The physical program is moving at speed. SB Energy has already ordered the gas turbines from GE Vernova — close to 170 of them, a roughly $10 billion order — and reserved slots at other manufacturers, with first delivery expected by roughly March 2027 and the rest coming online by roughly 2030.

Source: GE Vernova.

The count is the architecture: 9.2 GW across some 170 machines averages about 54 MW per unit, an aeroderivative fleet — jet-engine-derived turbines that ship in months rather than the years a large combined-cycle block takes, buying speed at the cost of burning more gas per megawatt-hour. The premium shows in the math: divide the $33.3 billion of Japanese-funded generation by the 9.2 GW it buys and the fleet prices out at roughly $3.6 million per megawatt — two to three times what a combined-cycle gas plant cost two years ago.

As developer, SB Energy said it would go vertical on the first data center this summer; the 800 MW first phase — a $30–40 billion data center build all-in, chips-included — targets power in early 2028, one increment of a full 8 GW campus put at roughly $500 billion. Tim Walsh, DOE’s assistant secretary for environmental management, quipped: “This is referred to as the A Plant … It will soon be referred to as the AI Plant.” SB Energy has also mentioned an additional 800 MW for the site’s second data center.

Rendering of future PORTS Technology Campus. Source: SoftBank Group.

The project became the first AI data center “covered project” under FAST-41, the federal permitting dashboard — a federal flagship whose own initiation notice states that “No DOE permitting support is being requested at this time.” A U.S. Army Corps of Engineers individual permit application went in on May 27, 2026, and a combined Ohio EPA water-quality application followed on July 1 — with tree-clearing scheduled for the window between October 1, 2026, and March 31, 2027, and grading running through 2028.

Source: U.S. Department of Energy.

The stated purpose of the project is to build a 10 GW AI data center complex, add new gas generation to balance data center power demand, and invest in grid upgrades.

SB Energy currently owns, has the option to purchase, or has rights to all land within the project area — a private landholding of roughly 2,700 acres, of which the Army Corps’ published permit notice puts 1,318 acres as the ground actually being permitted: 930 acres surveyed in late 2025 plus 388 more at Woodburn Heights delineated in early 2026, a footprint that grew mid-permitting.

The permit filings state what goes on it — “several” data center buildings and gas plants in SB Energy’s own application. The permitted ground sits alongside — not inside — the 189 acres leased from DOE.

SB Energy’s permit application carries the campus’s site plans, and they draw every building on all three properties.

Source: U.S. Army Corps of Engineers.

Counted off those sheets, the preferred alternative is 10 data center buildings at 450 MW each.

That is 4.5 GW of total capacity actually being permitted, against the 10 GW the same application says the project exists to build.

NVIDIA’s filings supply the first outside number that fits: the initial commitment it guarantees is 4.25 GW of IT load, and the guaranty steps up as each of nine data centers is placed in service. SB Energy says the campus comes online in phases beginning in 2028.

A Solar Developer’s 10-Gigawatt Leap

SB Energy comprises a team of 250+ people, 13 projects totaling 5.8 GW, a “pipeline” of 19 GW-IT of data centers and 17 GW of generation, and more than $17 billion raised — a figure NVIDIA is now adding $1.5 billion to, joining existing investors SoftBank Group and OpenAI on the cap table.

That amounts to a solar-and-storage track record now pointed at an AI campus larger than its entire built portfolio — thin enough that Japanese officials are still pushing for the plant’s operation to be put out to tender. To this end, SoftBank is buying part of the missing expertise in parallel: a $4 billion take-private of DigitalBridge, a 30-year digital infrastructure asset manager with a multi-gigawatt portfolio of hyperscale data centers, expected to close in the second half of 2026.

Two Billion Cubic Feet a Day

Of the 21 companies in the Portsmouth Consortium, exactly one moves natural gas: Kinder Morgan, which operates the largest natural-gas network in North America and carries roughly 40% of all the gas the country burns.

Kinder Morgan’s Natural Gas segment U.S. pipeline network map. Source: Kinder Morgan Inc.

It is there because 9.2 GW of gas generation is inert without a fuel line to feed it, and the line that can feed this one does not yet exist. Every other piece of the project assumes gas shows up at the fence: the turbines SB Energy has ordered, the transmission AEP is building, the concrete already pouring. Kinder Morgan is the consortium’s answer to how it gets there.

The magnitude is the whole problem. Sized for full output — and a feed line must be sized for the fleet’s peak draw — 9.2 GW of gas turbines burns on the order of 2 billion cubic feet of gas a day. That is not a plant’s appetite; it is a region’s, arriving at one address in Pike County. It equals roughly half of all the natural gas the entire state of Ohio consumes today, and it is one-fifth of the more than 10 Bcf/d “shadow backlog” of power and data center demand Kinder Morgan discloses it is chasing across the whole of the United States. One campus, one-fifth of the national opportunity. No pipe in the ground near Piketon was built to carry a sliver of it.

Kinder Morgan already runs the region — its Tennessee Gas Pipeline (the green TGP line in the map above) enters Ohio through Scioto County, immediately south of the reservation, and already feeds Piketon’s local distributor. But a mainline built to move gas through Appalachia cannot drop 2 billion cubic feet a day at a single new meter. Serving PORTS means growing the footprint: a new large-diameter lateral to the campus, new compression and parallel looping to manufacture firm capacity, fresh interconnects to pull supply from the Utica and Marcellus fields to the east, and a delivery station sized to feed the whole fleet at once.

Because it crosses state lines, all of it runs through a FERC certificate — a two-to-four-year path that should be underway now to feed the fleet’s ramp through 2030; the 800 MW first phase can lean on existing supply, but 9.2 GW cannot. And with roughly 700 billion cubic feet of storage on its system, Kinder Morgan can do the other half of the job: guarantee the no-notice, cold-snap-proof supply a data center island cannot operate without.

The gas figure has an electricity mirror, and it is just as state-sized. Run flat, 10 GW of campus load draws about 87.6 terawatt-hours a year — roughly 57% of all the electricity Ohio sold in 2024, and 61% of everything the state generated. Piketon is not a large load bolted onto Ohio’s grid so much as a second power system nearly the size of the first, stood up at one address.

Tenant Signed, and the Chipmaker Guaranteed the Power

On August 17, 2026, the question that had hung over the campus since the groundbreaking was answered by NVIDIA. OpenAI is the tenant. SB Energy will build, own, and operate the data center under a 20-year lease. And NVIDIA made itself the credit behind it. The campus picked up a commercial name the same day: PORTS-Pike Technology Campus.

OpenAI had signed the same counterparty’s paper before — in Texas. SB Energy is building OpenAI’s Stargate data center in Milam County, Texas — announced at 1.2 GW, known as “Freebird” — under a lease term of over 15 years and a non-exclusive preferred partnership for the projects that follow.

Source: Greg Brockman on X.

OpenAI and SoftBank each put $500 million into SB Energy in January 2026 alongside that project. In Ohio the same developer landed the same tenant — and a third party underwrote it.

The instrument is a set of residual value guaranties on leases covering roughly 4.25 GW of IT load, each taking effect when its lease commences. NVIDIA’s aggregate payment obligation is cumulatively capped at $105 billion for that initial commitment — nearly three times the $33.3 billion Japan is financing and the $4.2 billion AEP is spending on wires, combined. Nothing is owed until SB Energy delivers: payment is conditioned on it meeting ready-for-service conditions, the first expected in NVIDIA’s fiscal 2029, which opens in late January 2028.

And it is not the rent alone. The filed contract makes NVIDIA’s signature the condition on which the whole stack was assembled: the SB Energy entity acting as landlord agreed to enter the lease, and its power affiliate agreed to enter both a power purchase agreement for retail electricity and a commitment to fund the transmission upgrades needed to deliver it, “if and only if” NVIDIA delivered the guaranty. The guaranteed minimum value NVIDIA stands behind is defined as data center, power, and transmission costs. A chipmaker’s balance sheet is what closed the physical stack.

What triggers it is narrow, and it is about one company’s credit. If OpenAI becomes insolvent and defaults, or simply fails to pay, NVIDIA covers the shortfall between the lease’s guaranteed minimum value and whatever SB Energy recovers by reletting the space or selling it. NVIDIA can then assume the lease itself, push the landlord to find a new tenant, force a sale, let the lease die, or defer the decision for a year while paying the project’s costs. It is not first in line either: the landlord must exhaust the security deposit, a debt service reserve, business-interruption insurance, and OpenAI’s own parent guaranty of the lease before it reaches NVIDIA. And OpenAI has agreed to reimburse NVIDIA for certain losses if it ever does pay — which keeps the guaranty a bridge across OpenAI’s balance sheet rather than a transfer of the loss off it.

The exit clause is the tell. NVIDIA’s obligation ends at the 20th anniversary of the lease, or if OpenAI terminates it, or — the operative one — when OpenAI achieves a satisfactory credit rating. The guaranty exists because OpenAI does not have one. It is scaffolding, written to come down.

NVIDIA also bought the right to double the position. Beyond the initial 4.25 GW it may provide credit support for roughly 3.75 GW more, which together make up the 8 GW of AI compute the campus’s 10 GW of generation is sized to yield. On NVIDIA’s telling, OpenAI is the customer for all of it.

What NVIDIA gets is a closed hardware franchise. The site will exclusively install and deploy NVIDIA reference architecture for at least 20 years, and the project agreements behind it cannot be amended without NVIDIA’s written consent. OpenAI will use the capacity to deploy NVIDIA’s full-stack DSX AI factory platform — GPUs, CPUs, and networking.

NVIDIA is not lending its balance sheet for a fee; it is buying an 8 GW install base, one Jensen Huang says can be “upgraded repeatedly with each new generation delivering more intelligence and better economics.” Each generation deployed at PORTS-Pike could represent roughly 1.5 million NVIDIA GPUs, or $150 billion to $200 billion of NVIDIA revenue — and at the ~$40 billion per gigawatt NVIDIA books for its Vera Rubin generation, that band covers the initial 4.25 GW. Alongside the guaranty NVIDIA is putting $1.5 billion of equity into SB Energy.

The shape of the deal explains the appeal to OpenAI. It controls the equipment under a long-term lease; the land is leased from DOE; the generation is owned by the U.S. government and financed from Tokyo; the transmission is AEP’s; and the rent and the power alike are guaranteed by its chip supplier. That is 8 GW of compute with someone else’s balance sheet under every single layer, because here even the government is a counterparty.

“Someone else,” though, is substantially one group: SoftBank has invested roughly $55 billion in OpenAI to date — heading for $65 billion and about 13% ownership with a final tranche this October — owns 90% of the developer, and now shares the SB Energy cap table with both its tenant and its guarantor. A single Ohio campus at 8 GW of compute would exceed OpenAI’s entire Stargate program as announced in September 2025.

The tenant is settled; the megawatts are not. That gate is written in AEP’s own filings — and it is where both Piketon projects’ shapes were forged.

IV. The Wires Decide the Schedule

AEP’s own filings define the box any grid-served Ohio data center now sits in: a take-or-pay tariff the hyperscalers fought and lost, and a transmission system whose operator says it cannot reliably serve the new load for years yet.

The $4.2 Billion Answer

The campus’s grid framework is a $4.2 billion SB Energy–AEP transmission program, publicly framed as compliance with Ohio’s Ratepayer Protection Pledge. It arrives in stages — an initial 345 kV tie at the end of 2027, the full 765 kV backbone in 2029 — but no stage can carry the whole campus, which is why on-site gas still supplies first power in early 2028 and scales with the load. Its marquee element is Baku–Gavin, a roughly 50-mile 765 kV line from a proposed Baku Substation on Wakefield Mound Road in Pike County, Ohio to the existing Gavin Substation in Cheshire, Gallia County, Ohio.

Source: American Electric Power (AEP).

The build comprises 155-foot lattice towers on a 200-foot right-of-way and costs roughly $1.1 billion. It is now before the Ohio Power Siting Board (OPSB), with construction targeted for February 2027 and service in December 2028. The village of Cheshire is a fitting endpoint: it is the Ohio River town where Kyger Creek was built to feed the enrichment plant 70 years ago.

The near-term feed is smaller and faster. Monza, a greenfield 345 kV ring-bus station — six breakers and a 150 MVAR static compensator, $156.6 million, projected in service December 31, 2027 — hangs off the existing Don Marquis substation to serve a need of 800 MW near Piketon identified by PJM, the regional grid operator.

Source: PJM.

This is the full circle: the substation AEP built in 1969 to interconnect the A-Plant’s dedicated power supply will deliver the AI campus’s first phase.

What the projects do not do is instructive. AEP’s active interconnection queue runs to roughly 190 GW system-wide, with 63 GW of contracted capacity — 90% of it data centers — and the 10 GW Piketon campus counted among the requests. But in PJM’s own generation queue, as of July 2026, no campus-scale Pike County, Ohio, entry exists at all.

The wire story runs through AEP’s transmission planning and the state siting board, not the standard queue both projects were built to avoid.

The Tariff the Hyperscalers Lost

The tariff was built in one Public Utilities Commission of Ohio (PUCO) docket — case 24-0508-EL-ATA — and its outcome explains why Piketon looks the way it does. In March 2023, AEP quietly stopped taking new central-Ohio data center service requests — a moratorium. In May 2024, sitting on an unsigned queue of more than 30 GW, it filed for a data center tariff; two rival settlements followed — Amazon, Google, Microsoft, and Meta behind one; a utility-and-consumer coalition behind the other. The hyperscalers lost: the Commission adopted the coalition’s version, approved final tariffs in July 2025, and denied rehearing that September — resting its reliability case in part on a confidential AEP transmission-planning summary that Meta had entered into the record as an exhibit. Meta helped arm the tariff it fought.

Schedule Data Center Tariff (DCT), as adopted, is two mechanisms:

1. Load Ramp: Contracted capacity steps up to 90% over as many as four years.

2. Minimum-Demand Ladder: Once ramped, a customer owes for 60% to 85% of contract capacity — the 85% cap binding at 117 MW and above — “regardless of the customer choosing to curtail, reduce, suspend, or terminate service.” The term runs the ramp plus eight years (12 years at the maximum ramp), behind exit fees and collateral; even self-help is fenced, with netting of behind-the-meter generation requiring equipment that can instantaneously curtail at least as much load.

The aftermath says who felt bound. The hyperscalers stayed; it was the manufacturers — the Ohio Manufacturers’ Association Energy Group — who took the tariff to the Ohio Supreme Court as “textbook discrimination,” in a still-pending case. Seven decades earlier, on this same ground, the federal government signed the same shape of deal from the other side of the table: a demand charge owed “whether energy is used or not.” The take-or-pay contract has come home to Piketon — this time with the data center paying it.

The Gate: Grid in 2031, Campus in 2028

Then there is the timeline itself. AEP’s November 2025 load study stated that none of the 13 GW of “New Data Center Load” can be reliably served until the completion of the “Ohio Seven Year Solution” (PJM 2025W1-570) project, a proposed 765 kV / 345 kV transmission expansion anticipated to be in service by approximately Q4 2031.

By February 2026, AEP Ohio alone counted a grand total of 17.9 GW of data center load under signed Electric Service Agreements. Set those two facts against the Piketon campus’s early 2028 first power and the conclusion writes itself: for a campus of this size on this clock, on-site generation plus dedicated transmission is not a preference — it is the only solution that works.

Ohio’s two biggest data center power projects never appear in the tariff docket — the proceeding that prices grid service for Ohio data centers — because both engineered around it: one by generating on site under federal ownership, the other by building supply rather than taking service. But that is a queue-bypass, not an exit from the grid.

PORTS Technology Campus stays wired to AEP, its grid ties phased in from 2027 to 2029 behind the gas, and its self-supply is now explicit: SB Energy says at least 10 GW of new generation results in 8 GW of IT capacity — implying a power usage effectiveness (PUE) ratio of 1.25. The 2 GW gap is cooling, conversion, and losses. None of it is the roughly 20% reserve margin PJM plans around, so the grid still carries the balancing and contingency an island would have to supply itself. Its tariff status is unresolved, not avoided.

What Meta bought is the other half of this story.

Subscribe now

V. Meta’s Answer Is a Reactor Field

Meta bought power certainty the opposite way: it prepaid a first-of-a-kind nuclear reactor company that had already spent two years assembling recycled federal land and a co-located fuel chain. The deal is signed, and it buys output from a design that has never been licensed anywhere, in a state where its licensing has not publicly begun.

Oklo Was Here First

The Piketon nuclear reactor campus was not conjured by Meta’s demand; it was supersized by it. On May 18, 2023 — two and a half years before the AI land rush reached Pike County — Oklo and SODI signed an agreement for two commercial Oklo power plants in Southern Ohio, up to 30 megawatts electric in total. In February 2024, Oklo took an option and right of first refusal on the land.

Only then came Meta: a December 2024 RFP seeking 1–4 GW of new nuclear, delivery starting in the early 2030s. On January 5, 2026, Meta signed a prepayment agreement. Oklo’s CFO connected the dots on the company’s own call: “another important part on Meta was already having a ROFR in place and access to land in Ohio … we’ve leveraged that land access.” Meta’s money scaled a 30 MW toehold forty-fold, to 1.2 GW.

The instrument is a prepayment, not a power purchase agreement (PPA) — a mechanism for Meta to prepay for power and provide funding, with proceeds put toward securing nuclear fuel and advancing the first phase; a future PPA is anticipated. The structure is effectively a non-refundable upfront payment that will be attributed to future power delivery. And the prepayment is the seed of the capital stack, not the whole of it: Oklo assigns the forthcoming PPAs a second job — “to enable project financing with the Office of Energy Dominance Financing and commercial lenders” — routing the build through DOE’s renamed Loan Programs Office.

Note the symmetry across the fence line: Tokyo’s state credit stands behind the gas Washington will own; Washington’s own state lender is penciled in behind the nuclear Meta prepaid. Both theories of power plan to borrow from a government.

The land is also documented. On December 16, 2025, Oklo Power SODI LLC — a Delaware entity registered in Ohio — recorded its deed to 205.9 acres at 3930 US-23 South, Scioto Township, purchased from SODI for $5.15 million. The chain of title runs from “Parcel 2 of the U.S. Department of Energy property”; the deed carries a 40-foot DOE easement, CERCLA environmental covenants, and a defense-nuclear-facility indemnity under §2811. The parcel today is vacant industrial woodland. What Meta prepaid for is the right to what gets built on it.

A Design Licensed Nowhere

The Aurora “powerhouse” is a sodium-cooled fast reactor carrying about 75 MW electric per unit.

Source: Oklo Inc.

It has a licensing record with one essential feature: none of it is in Ohio. Oklo filed the industry’s first custom combined license application in 2020, for a unit at Idaho National Laboratory (INL); the U.S. Nuclear Regulatory Commission (NRC) denied it without prejudice in January 2022. The INL unit, still targeted for 2027, has since stepped out of the NRC’s line entirely — authorized by DOE itself under the Reactor Pilot Program — so the design’s return to the NRC now runs through the Meta campus: “Aurora Ohio … in connection with Meta agreement” heads Oklo’s own list of assets slated for commercial licensing. Oklo’s strategy is product-based licensing — a design “licensed to a product … regardless of its geographical location,” the aviation type-certificate analogy.

The Fuel Runs Ahead of the Reactor

The fuel chain is further along than the reactor. Aurora runs on HALEU — high-assay low-enriched uranium — and Oklo holds two distinct instruments with Centrus, the uranium enricher expanding HALEU production next door at Piketon:

1. Supply Letter of Intent: Explicitly non-binding, covering up to five powerhouses, deliveries beginning in 2029; and

2. Planned De-Conversion Joint Venture: Co-located with Centrus’s enrichment operations and adjacent to Oklo’s planned 1.2 GW power campus.

Source: Centrus Energy Corp.

The federally funded enrichment district those instruments plug into is a story of its own (Section VI).

The published timeline: preconstruction and site characterization in 2026–27, with general construction not starting until 2028–29, first 150 MW targeted around 2030, and 1.2 GW by 2034. Oklo intends to operate each powerhouse with about 35 permanent staff, 9,000 gallons of water per day, and no cost to ratepayers.

But “further along” is relative. At about 75 MW each, 1.2 GW is 16 reactors, not one — and the non-binding supply letter covers only five. The fuel for the rest barely exists: Centrus’s Piketon cascade makes roughly 900 kilograms of HALEU a year, essentially all current U.S. output, with no new capacity before 2029. The chain leads the reactor into a supply that isn’t there yet.

What the Prepay Buys Meta

There is no Meta data center facility at Piketon and none proposed; the campus would inject up to 1.2 GW of clean baseload power directly into the PJM market, with interconnection applications submitted on April 27, 2026, and Oklo pointing to existing transmission corridors and substations in Pike County. The load it serves sits about 75 miles north: Meta’s operations in the region, including its Prometheus AI supercluster in New Albany, Ohio. The connection is a contract and the grid.

Source: Meta Platforms.

What powers that load in the meantime is the wrinkle in the tidy make-or-buy split. Prometheus is already partially operational and has expanded to more than 3 GW of capacity. Additionally, Meta is building generation at the load itself: the Ohio Power Siting Board approved the 200 MW Socrates South plant, built and operated by a Williams subsidiary, to supply the New Albany campus behind the meter. The buyer of reactors is also a builder of gas-fired generation — standing up exactly the self-supply the new tariff’s netting clause was written to fence (Section IV). Both Piketon camps reach for gas first; they differ on where it burns, who owns it, and which hedge arrives in the 2030s.

What the prepay buys is position. Oklo sits inside Meta’s January 2026 nuclear slate — up to approximately 6.6 GWe across Oklo, TerraPower, and Vistra — which itself sits atop Meta’s separate 1.1 GW Clinton PPA from June 2025; in total a portfolio of 7.7 GW of nuclear partnerships.

Meta has paid for power that arrives after the gas rush — including its own — from a supplier whose construction risk is capped, for Meta, at an undisclosed prepayment. Oklo is pre-revenue, with a $105.7 million net loss in fiscal 2025 and $2.5 billion of liquidity as of March 31, 2026. Meta’s downside is a check it has already written and a delivery date it can survive missing; its upside is first claim on firm, carbon-free megawatts in the exact PJM region where a 50- to 60 GW capacity shortfall looms over the next decade, and where Meta just lost the fight over what grid service costs.

Against the gas project across the fence line, the trade is asymmetrical: the government’s campus is faster, taxpayer-owned, and leased to a tenant whose credit had to be guaranteed by its chip supplier; Meta’s is slower and prepaid off its own balance sheet. Which is the better theory of power is the question the whole reservation now exists to answer.

VI. One Reservation, Two Theories of Power

Piketon is the first place where the industry’s two competing answers to the power problem — build gas fast, or buy nuclear early — sit on the same fence line, on land the federal government spent 20 years clearing.

The Cluster Is Bigger Than Two

Pull back and the reservation is quietly becoming a federal energy-industrial park.

Alongside the gas-and-data center lease and the reactor campus: Centrus’s enrichment expansion — a $900 million DOE task order to expand Piketon HALEU production, under contract since June 30, 2026, with Fluor and Geiger Brothers engaged, adjacent to but never part of Oklo’s campus. Uranium enrichment, in other words, is returning to the reservation whose enrichment plant DOE spent two decades demolishing — this time to feed reactors next door. Trillium H2 Power, an early-stage low-carbon hydrogen developer, holds 79.5 acres; Mid-America Conversion Services (a joint venture between AtkinsRéalis, Westinghouse, and Fluor) runs the DUF6 conversion plant. And between the X-326 pad and the Oklo flank sits the “5-Unit Plume,” an active groundwater remediation area — a reminder that this industrial park’s anchor tenant is still the cleanup itself.

Source: U.S. Department of Energy.

A Model for Other DOE Sites

That cluster is the real policy story. The lease runs under the Hall Amendment and DOE’s 10 CFR 770 authority. Its own recitals state the following: DOE support “furthers the Executive mandate to enhance America’s global AI dominance.” The nuclear-weapons complex’s cleanup program — built to shrink federal liability — has become a land-and-power pipeline for AI, and Portsmouth is its proof of concept: a site DOE’s own AI-siting program passed over in July 2025, now framed by DOE as “a model for what could happen at other DOE sites” — implying there are more reservations where this one came from.

Subscribe now

VII. What to Watch

Source: U.S. Department of Energy.

Nearly everything decisive at Piketon is already scheduled — on a permit calendar, a docket, or a delivery window. The dated markers:

• Whether NVIDIA doubles down. The option on a further 3.75 GW is exercisable at NVIDIA’s sole discretion and would take the campus to 8 GW of compute.

• SB Energy goes public. A confidential filing points to an IPO as early as September 2026 at a reported $50 billion-plus — now with a signed 20-year tenant, a $105 billion NVIDIA guaranty behind the leases, and NVIDIA itself on the cap table, but a gas fleet still unbuilt.

• First power, early 2028. First building vertical in summer 2026, first GE Vernova gas turbines by roughly March 2027, Monza 345 kV energized December 31, 2027 — and the 800 MW phase closes the Don Marquis full circle.

• OPSB case 26-0426. AEP’s Baku–Gavin 765 kV application was filed in summer 2026 and is now before the OPSB, with construction targeted for February 2027 and service in December 2028 — the marquee of the $4.2 billion grid program.

• Oklo’s Ohio docket. The real tripwire is the first Ohio-specific NRC filing; site work is slated for 2026–27 and general construction for 2028–29, Centrus HALEU deliveries begin 2029, first 150 MW around 2030.

• X-333 demolition, running to 2031. The measure of how long the AI campus rises inside an active CERCLA cleanup — and, with SODI’s roughly 1,000-acre pool, of how much more land the factory can produce.

Source: U.S. Department of Energy.

Subscribe now

Measured AI provides institutional-grade analysis of the physical infrastructure powering AI data centers. For access to our full research library, regulatory intelligence, and weekly briefings, visit MeasuredAI.com.

A month later the shift in control became a matter of county record. On July 8, the Laramie County Board of County Commissioners told residents that “Crusoe Energy has withdrawn from the project,” that the campus formerly called Project Jade is now Project Tembo, and that the party across the table is Google. The stronger document had already been filed a week earlier: Google’s June 30 site plan narrative states that “Google, through its wholly owned subsidiary Jupiter Star Holdings LLC … will be the developer and owner of the project” — a resubmitted program of four data center buildings inside a campus totaling about 3.08 million square feet, smaller than the five-building plan Crusoe had permitted. Bloomberg also reported in June that Google “had raised concerns about the costs and timetable under Crusoe’s watch.”

What Google took over is one of the largest planned AI data center campuses in the United States: a 715.9-acre data center site inside a 1,375-acre industrial park on former ranch land eight miles south of Cheyenne, in Laramie County, Wyoming.

The project was announced at 1.8 GW with a stated design ceiling of 10 GW. That 1.8 GW load is roughly 4.1 times the all-time peak of the utility that will serve it. Next door, on land the campus’s own power developer now owns, Blackstone-backed Tallgrass is building the Cheyenne Power Hub — a first phase of two Mitsubishi M501JAC turbines totaling roughly 1.15 GW, inside a permitted generation fleet the applicant sizes at up to 2.2 GW, behind a power investment Tallgrass puts above $7 billion.

The structure is what makes Cheyenne an instructive case study. Google’s Project Tembo is a unique model — Google takes only the data center, and everything upstream belongs to someone else — and it is not the model everyone, including the applicant, has described it as.

This primer walks the project layer by layer: the ownership structure and why Google needs it, the handover from Crusoe, the land, the data centers, the power hub and its capacity numbers, the utility deal and who pays for it, the water, and the markers that will test all of it over the next 18 months.

Subscribe now

I. The Google Model — Own the Data Center, Not the Power

Every layer of this campus belongs to a different named company, and the relationships between them reveal what’s really going on.

Google is the data center’s developer and owner — and holds nothing upstream. Tallgrass owns the gas turbines. An American Electric Power (AEP) subsidiary owns the fuel cells. Tallgrass’s Rockies Express Pipeline (REX) moves the gas. Tallgrass also owns the switchyard where every megawatt is sold. Black Hills’ Wyoming utility owns the Robinson Substation that ties the campus to the grid — and Google is paying for it. And — the fact that reorganizes everything else — none of that generation sits behind Google’s meter.

The Structure, Named

“Behind the Meter” — or Not

At the county hearing that approved the site plans in January 2026, Crusoe’s Chief Real Estate Officer, Matt Field, described the power strategy in the phrase the whole industry now uses: “We also have a unique take on power … We’re colloquially viewing it as BYOP, bring your own power.” The air-permit application states the same thing, describing the plant as “a behind-the-meter power generation facility that will deliver electricity to a data center.”

The FERC filing says otherwise, and it is the better authority. Cheyenne Power Hub’s July 2026 waiver application walks through three phases of the project, and in every one of them the generators — the gas fleet and the fuel cells alike — “sell the entirety of the output to BHE,” and Black Hills Energy serves the data center “through a retail sale.” The interconnection is engineered to cap flow from the generation onto the utility’s transmission system at “only 600 MW“ — dedicated supply, not a merchant plant. Black Hills takes title on site and delivers to the data center over dedicated lines that never enter its transmission network, so the 600 MW is the ceiling on what can reach the wider grid, not a limit on what the utility buys. The waiver’s own exhibit draws the arrangement — every owner color-coded, every generator sized, and the cap labeled at the interconnection.

Source: Federal Energy Regulatory Commission (FERC).

The same corporate family told two agencies two different things: “behind-the-meter” to the air regulator as a physical description, and utility-resale to FERC as the commercial architecture. The commercial architecture governs.

That distinction is not pedantry; it is the model. Bring-your-own-generation means a load develops or contracts for on-site power that it takes directly — behind its own meter, netted against its own consumption, with the utility bypassed as the seller. That is what lets it begin operating without waiting for grid service, and it is why the load, not the utility, controls the supply.

Cheyenne keeps every part and moves the meter: the generation is new and dedicated, but the load never owns or nets against it — it is front-of-the-meter, dedicated generation, sold at regulated retail. Wyoming’s answer to the interconnection queue is not to leave the utility; it is to keep the utility as the seller, build the generation anyway, and ring-fence who pays for it. Sections V and VI trace each half of that framework.

Why Google Needs It

The demand side of the project is Google’s, but the frame is Alphabet’s. At fleet level rather than site level, Google’s parent company provides the corporate context this Wyoming project sits inside. Alphabet’s ability to scale its physical infrastructure is increasingly constrained by the availability of power, water, and land. The spending curve behind that statement: capital expenditures of $52.5 billion in 2024, $91.4 billion in 2025, and FY2026 guidance raised twice since the beginning of the year to $195–205 billion — with 2027 guided only as “a significant increase.”

Google’s own energy team describes the approach behind the spending. In November 2025, Google’s global head of data center energy, Amanda Peterson Corio, said the company has “been moving Google from a power customer to an infrastructure partner,” running an “energy first” screen before siting decisions — and, notably, stated a preference for grid-connected load, exploring co-located generation only “in markets where it is regulatorily allowed” and “while still building a path to ultimately interconnect these co-located power and data center assets to the grid.”

The named example was a partner: Intersect Power, a developer that co-locates data centers with dedicated renewable and gas-fired generation. Within months Google bought the company, closing March 10, 2026 at a $5.9 billion valuation.

Cheyenne fits Google’s approach precisely — dedicated generation, sold through a utility, grid relationship intact from day one — but not the ownership: here every megawatt belongs to Tallgrass or AEP. And Google’s cost pledge to “intentionally structure our energy deals to cover 100% of the costs of the power we use” is about to get its most literal test in a Wyoming tariff docket (Section VI).

On August 5, 2026, Google itself stated in the Wyoming regulatory record, via a petition to intervene, its arrangement with its utility counterparty, Black Hills: “Google is considering development of data center facilities in Cheyenne, Wyoming, within the Company’s service territory. Under this proposed development, Google will purchase electricity from the Company and become a large-load customer.”

The structure explains what Google bought. The handover explains how it came to be the buyer.

II. The Handover: Crusoe Out, Google In

Project Tembo’s defining event is a mid-build change of control and one of the clearest cases of a hyperscaler customer displacing a neocloud developer from a flagship AI data center campus while keeping the developer’s power platform intact.

The public sequence ran three days. Crusoe, June 9: a customer-requested “pause.” Black Hills, June 10: no pause — Crusoe “is no longer the development partner,” and the utility “is working directly with the potential large-load customer.” Bloomberg, June 11: Crusoe “pushed aside” after Google raised cost and schedule concerns, with Google “working to finalize a deal with the remaining partners to buy computing from the site.”

The government record fixes the outcome. The data center’s own stormwater permit — held by builder DPR Construction — was terminated on June 8, the day before Crusoe’s statement, in a single sentence with no successor permit. The county’s July 8 memorandum states simply that “Crusoe Energy has withdrawn from the project.”

The record also shows how early the transition began. On May 14 — about a month before the first press report of Crusoe’s exit — the Laramie County planning director told a hearing, “we met with Google last week.” A county letter the following week disclosed that “Laramie County was under a confidentiality agreement with Crusoe only” — the incoming owner, Google, had no NDA with the county at all. And on Black Hills’ August 6 earnings call, CEO Linn Evans dismissed the drama entirely: “the exit of Crusoe has not had any impact on the negotiations. In fact, it’s been important to us from essentially day 1 to ensure that we’re negotiating with the hyperscale end user … we want to do it right, not just fast.”

What Google did next is on the record: a smaller campus, resubmitted in its own name, on ground it does not own.

Subscribe now

III. The Land: From Ranch to Industrial Park

Six decades of grazing land became an approved 1,375-acre industrial park in under six months.

The campus sits in the Switchgrass Industrial Park, about eight miles south of Cheyenne along US 85, on ground that spent six decades as the Lazy D Grazing Association’s ranch land, held since a 1965 deed.

Source: Laramie County Planning & Development.

On December 4, 2025, the ranch changed hands outright: Lazy D transferred the park and several sections around it — including the lots the data center and the power project sit on — to Tallgrass-owned Cheyenne Power Hub, LLC by warranty deed. And the buying has not stopped: on July 24, 2026, Lazy D sold Tallgrass a further full section southeast of the industrial park, roughly 640 acres, plus part of another.

Switchgrass Industrial Park is a 1,375-acre planned unit development across five lots: the data center project area is 715.9 acres (Lots 1–3), and the generation site, known as Cheyenne Power Hub, is another 659 acres (Lots 4–5).

Source: Laramie County Planning & Development. Measured AI annotations.

The entitlement moved fast: a zone change was filed in July 2025 and approved in October 2025 with no conditions. Both site plans — data center and power plant — cleared the Laramie County Board of County Commissioners unanimously on January 6, 2026.

The land was the fast part. What sits on Lots 1 through 3 is what Google inherited from Crusoe — and then cut.

IV. The Data Centers: What Google Took Over — and Shrank

The resubmitted campus is smaller than the one Crusoe permitted. The county memorandum records the program dropping from five data center buildings to four, from 4.0 million square feet of built area to 3.08 million square feet. Of this total, the four data center buildings (DC-1A, DC-2A, DC-3A, and DC-4A) will make up 2.84 million square feet, with the remaining 240,000 square feet made up of office, logistics, and network buildings.

Source: Laramie County Planning & Development.

The memo also records a smaller construction force: peak workforce “reduced from 5,000 to 2,500,” both figures belonging to the data center program. Tallgrass’s Cheyenne Power Hub (discussed next) is counted separately, with its most recent studies putting its peak workforce at 2,740 workers in early 2028. Laramie County’s planning director has described the program as a build-out running to 2031.

The buildings are not the whole build. The campus also carries battery energy storage systems (BESS) — an on-site battery bank that sits behind Google’s meter, unlike every asset that actually makes power.

The site plans name Google as “developer and owner of the project.” What is not in question is the limit of what Google will own: four data center buildings, on ground held by Tallgrass, drawing a load Google will not generate a watt of.

Everything hard about this project sits upstream of that meter.

Subscribe now

V. The Cheyenne Power Hub: Everything Upstream of the Meter

The Cheyenne Power Hub is what Google did not buy. Owned and built by Tallgrass, it occupies 659 acres immediately southwest of the data center tracts and holds two natural gas power generation facilities, two aeroderivative substation yards, three fuel cell yards leased to an American Electric Power subsidiary, a gas meter station at the southeast corner fed by Tallgrass’s own pipeline, and one 230 kV substation yard where every megawatt on the property converges.

Source: Laramie County Planning & Development.

Natural Gas Power Generation Facilities

The two natural gas-fired power generation facilities are the permanent plant that will run the campus long term, owned by Tallgrass. They hold the turbines: two Mitsubishi M501JAC combined-cycle units — announced from Tokyo on May 15, 2026. These combined-cycle turbines will provide approximately 1.15 GW of site-ready power as Phase 1 of the Cheyenne Power Hub, with major components scheduled to begin arriving in July 2026 and total investment that “will exceed $7 billion.”

Source: Mitsubishi Heavy Industries.

More precisely, from the permit’s perspective, the two M501JAC combined-cycle units can produce 1.16 GW at this site — the pair’s 1.33 GW ISO nameplate derated 12.7% by Cheyenne’s ~6,200-foot elevation and ambient conditions to 580 MW each in combined-cycle mode.

Aeroderivative Substation Yards

The two aeroderivative substation yards are the bridge-and-backup fleet. These are the aeroderivatives — jet-engine-derived machines, small, modular, fast-starting, several of them literally trailer-mounted (e.g., the GE TM2500 — see below).

Source: GE Vernova.

The temporary units provide early generation for initial data center loads, plus startup and backup power for the combined-cycle plant, while the permanent units provide base-load generation that supplements and backs up those same units. Together they carry the load before the combined-cycle plant is commissioned, and back it up after.

More Turbines, Less Generation

The air permit issued July 2, 2026, authorizes 42 turbines: the two M501JACs, 16 permanent simple-cycle units, and 24 “temporary” simple-cycle units — a fleet totaling 2.20 GW at site rating and 2.76 GW at ISO rating.

An amendment now pending would swap the fleet again: 56 turbines — the M501JACs plus 30 Baker Hughes NovaLT16s and 24 Mitsubishi FT8s — totaling less capacity: 1.83 GW at site rating and 2.58 GW at ISO rating.

More machines, fewer megawatts, and not yet issued. The timing explains it better than the machines do: the amendment was filed July 21, just weeks after Google took over and cut the campus from five buildings to four. The generation shrank with the buildings — though the load did not, since the utility record still describes a 1.8 GW data center through August. It also re-cuts the supply chain: the issued permit’s simple-cycle fleet is GE-based, while the amendment splits the 56 machines between Baker Hughes (30) and Mitsubishi (26) and leaves GE Vernova out entirely.

The environmental math is major-source scale in every direction. All nine regulated pollutants trigger federal Prevention of Significant Deterioration review; greenhouse-gas potential to emit is 10.4 million tons of CO₂e per year; modeled PM₂.₅ sits at 96% of the annual air-quality standard and cumulative one-hour NO₂ at 99.7% — about as close to the line as a permit can be written. The permit requires an ambient monitoring network within 180 days of startup, run at least three years.

Fuel Cells: Bloom Sells, AEP Owns

The fuel cells are dedicated generation running on the same natural gas as the turbines, converting it to electricity electrochemically instead of burning it — no flame, no rotating machinery, no water in normal operation.

Source: Bloom Energy.

The three fuel cell yards sit on 86.8 acres owned by Tallgrass, through its Cheyenne Power Hub subsidiary, and leased to AEP in January 2026. The fuel cell supply chain has two parties:

1. Bloom Energy manufactures and sells the solid-oxide units — an initial 100 MW order under a framework for up to 1 GW, backed at signing by a $100 million letter of credit. The two Bloom models being purchased are the Energy Server 5 (ES-5) and Energy Server 6.5 (ES-6.5) — with the ES-6.5 making up the bulk (up to 900 MW) and the ES-5 a 72 MW tranche.

Source: Bloom Energy.

2. American Electric Power (AEP), through its unregulated subsidiary, Cheyenne Fuel Cell, LLC, is the buyer, developer, owner, and offtake counterparty of the fuel cells. In January 2026, AEP purchased the fuel cells for approximately $2.65 billion. As of June 30, 2026, AEP carries the whole facility at $4 billion of incremental capex, with an in-service date by 2028. The fuel cell array actually being interconnected comprises approximately 650 MW of capacity — the air permit authorizes a 900 MW not-to-exceed ceiling. AEP has established a 20-year offtake agreement selling 100% of the output of the fuel cell generation facility to Google — a commercial layer sitting over the physical sale, since the FERC filing has the fuel cells’ entire output going to Black Hills for resale at retail.

The offtake’s conditions deadline was extended by mutual agreement in June 2026 — the month Crusoe exited — with a decision milestone in December 2026 intact. AEP’s CFO has said the company can “put the fuel cells back to the hyperscaler at a cost plus” roughly 10% if the deal fails, and that Google may elect to deploy the fuel cells at an alternate location — “could be anywhere in the U.S.” — if the Cheyenne site does not advance.

The commercial clock is set by the federal tax code: the cells must be “installed and ready to go by the end of 2028 to qualify for the investment tax credit.”

The Gas Supply: Tallgrass’s Own Pipeline

The fuel artery is Tallgrass’s own. Rockies Express Pipeline (REX) is a 1,700-mile interstate pipeline that originates in Rio Blanco County, Colorado, to the southwest and runs across Wyoming’s southeast corner, where Google’s Project Tembo sits. Here, Tallgrass is building the Switchgrass Interconnect: 7.2 miles of 36-inch lateral rated at 419,000 dekatherms a day, a $57.3 million project.

Source: Federal Energy Regulatory Commission (FERC).

Construction of the Switchgrass Interconnect began March 2, 2026; as of July 2026, pipe was being strung and bored, with concrete work finishing at the meter station.

The Substation Yard: Where Title Changes Hands

The Substation Yard, known as the CPH Substation, is the largest block on the Cheyenne Power Hub’s site and generates nothing. It sits along the northern edge of the site — half a mile from the utility’s Robinson Substation — and it is the reason the whole arrangement is legal.

Tallgrass’s FERC filing describes what runs in and out: 230 kV lead lines roughly 1,000 feet long from the combined-cycle plant, 2,000 feet from the second gas facility, and 1,000 feet from the fuel cells, all terminating in the CPH Substation, with “approximately twenty 1.5-mile, 34.5 kV lead lines” running out the far side to the data center. Bechtel is a contractor of record on the yard.

Two half-mile 230 kV lines — these owned by Black Hills — connect the yard to Robinson Substation, Black Hills’ new switching station to the north, which terminates transmission at 230 kV and steps nothing down.

Every transformation the campus needs happens at the CPH Substation, on Tallgrass ground, and so does every sale. The waiver Tallgrass is asking FERC for turns on this one asset and the lead lines feeding it.

VI. The Utility Deal: Who Pays, Through Which Meter

Everything in Section V ends at the same commercial destination: Black Hills’ Cheyenne utility buys the output and sells it to Google at retail. That makes the tariff architecture the load-bearing structure of this project, and it is an architecture with a history: Google is stepping into a rate framework a different hyperscaler built.

A Tariff Built Around Microsoft, Inherited by Google

Cheyenne Light, Fuel and Power — Black Hills’ Wyoming electric utility, called Wyoming Electric in Black Hills’ investor disclosure, then serving about 41,000 customers — “was contacted in 2011 and asked to provide a proposal to serve a large data center,” and the framework that emerged is the Large Power Contract Service (LPCS) tariff, filed in October 2015 and litigated with Microsoft as an intervenor telling the Commission that “Approval of the Tariff is essential to Microsoft’s plans for its data center in Cheyenne.” The Commission approved it in July 2016, writing that “the potential new load is large enough to nearly double the load of CLFP existing system” and calling the tariff “the beginning of a controlled experiment to be conducted under our continuing supervision.”

Three design choices from that docket now define Google’s deal a decade later:

1. LPCS customers form their own ring-fenced class — costs “assigned to them individually based upon cost causation principles,” with no allocation of the utility’s existing generation fleet.

2. Capacity is the customer’s problem: the current tariff book requires each LPCS customer to maintain Eligible Capacity Resources of at least 115% of its actual peak monthly load, from a set of options that includes “Company-owned generators built specifically for one customer served under this Rate Schedule” — the exact category a dedicated gas fleet sold to the utility slots into — and a 2024 amendment approved by the Commission made customer-owned on-site backup generation an LPCS obligation.

3. The consequence: the utility’s 2026 resource plan states outright that it “does not address Large Power Contract Service (LPCS) tariff capacity needs, which are handled separately under customer-specific agreements.” The plan forecasts 650 MW of contracted LPCS peak by 2031–2033 — counting only customers with agreements already in place — and provides no capacity resources for any of it.

A campus announced at 1.8 GW is absent from its serving utility’s capacity plan by design.

The Reservation Agreement: $377 Million, Refundable

The money moving today runs through an instrument that barely has a name in the tariff. On April 22, 2026, “Wyoming Electric entered into a generation reservation agreement with a prospective new customer seeking to construct a 1.8 GW data center in Wyoming subject to the LPCS Tariff” — Google — under which the customer advances refundable cash so the utility can make milestone payments to suppliers for long-lead generation equipment the utility itself may own.

The balance has climbed fast: $201 million disclosed in early May, $285 million on the June 30 balance sheet — booked as a current liability, “Refundable advances for construction,” a line that did not exist at year-end — and a July amendment raising the ceiling to $377 million with a maturity of August 31, 2026.

Black Hills calls it “a bridge agreement” while definitive agreements are negotiated; if it matures without renewal, the utility has said it would look to its revolver, short-term loans, and equity program to repay the advances. And the advances sit outside the tariff architecture entirely: none of the large-load tariff instruments mentions contributions in aid of construction, so how customer money for generation equipment is booked, refunded or converted is not established anywhere in the public tariff record. Asked whether the reserved equipment covers the full 1.8 GW, CEO Linn Evans was direct: “No, it is not. And we’re not yet ready to announce what kind of megawatts we would serve.”

The hard commitments to date are narrower and physical: the reservation advances, and the Robinson Substation — a 230 kV interconnection substation with ten line terminals, six reserved for future use — approved in May 2026 at about $20.5 million. It is customer-funded and has been under construction since June 10, 2026.

The Cost Wall: Making Large Loads Pay 100%

The mechanism that decides whether any of this touches ordinary bills is a new tariff. On June 19, 2026, Cheyenne Light asked the Commission to approve a Large Customer Transmission Cost Adjustment Mechanism — LCTCAM — by September 1, effective January 1, 2027. Its design principle, in the utility’s words, is that “cost allocation follows cost causation”: large-load customers — those on the LPCS tariff, or with dedicated transmission facilities of 25 MW or more — “will pay 100 percent of LCTCAM costs” — new transmission plant built for them — plus their load-ratio share of the shared system. Those obligations rest on 10-year demand forecasts that are binding and ratchet only upward, backed by an exit-fee provision and collateral requirements, with the utility earning a return on construction work in progress as it builds.

Black Hills frames the whole architecture to investors as service that “is fully self-supporting and does not shift costs to other customer classes,” and to Wyoming as alignment with Governor Mark Gordon’s executive order titled “Data Centers the Wyoming Way.”

The same cost-shifting protection exists at the federal level: since 2025, the utility’s FERC transmission formula excludes any plant “direct assigned to a large load Cheyenne Light retail customer” from wholesale rates — an exclusion that went from zero to $78.8 million of plant in its first year of operation, with a workpaper that names six plant accounts and no customer.

The docket is where the campus’s economics will be tested, and the parties tell you who has skin in it. Microsoft intervened in the present tense — it “owns and operates data center facilities in Cheyenne … and is a customer” whose transmission-cost responsibility is directly at stake. Google intervened with the hedged petition quoted in Section I. Industrial customers Dyno Nobel and Cheyenne Renewable Diesel — veterans of the 2016 fight — brought consultants; the state’s consumer advocate is reviewing. None has yet stated a position on the merits.

The Utility Being Transformed

The scale mismatch explains why every one of these mechanisms exists. Wyoming Electric’s all-time peak load is 439 MW, set in July 2026 — up 40.7% in three years, driven by data center growth that does not yet include this project. The Google campus’s 1.8 GW ultimate load is roughly four times that entire system.

The utility’s disclosed pipeline has tripled to more than 3 GW of signed NDAs, but its five-year, $4.7 billion capital plan carries only the approximately 600 MW of Microsoft and Meta load already under agreement — this project enters the plan only when definitive agreements sign, and would be financed on top of it.

The structural shift is in the service model: Black Hills describes three ways it serves large loads — market energy with a management fee (Microsoft and Meta today), contracted resources, and utility-owned rate-base generation “recovered through rates specific to customer.”

Project Tembo is positioned to be the first customer served under the third model. The reservation agreement’s long-lead equipment — the piece the utility itself may own — is what points there; the generation in Section V is Tallgrass’s, bought under contract — the second model. Which one the campus lands in is what the definitive agreements will settle.

Electricity is the layer where Google’s money is most visible. Water, the campus’s other critical input, runs even further from its owner.

Subscribe now

VII. Water: Ceilings, Not Consumption

What the record establishes is what the campus may draw, not what it will use — and its most consequential filing is the one still pending.

The campus’s water is groundwater, permitted through the State Engineer’s Office in seven filings whose combined ceiling comes to roughly 620 to 645 million gallons a year (1,895 to 1,974 acre-feet, the spread depending on whether pumping rates or annual volumes are totaled). Nearly all of it sits in two production wells, and only one of those is permanent — the other is a construction well whose right expires December 31, 2030.

Only one permit states cooling as a use: the campus’s permanent industrial well, applied for January 2, 2026 and permitted at 315 million gallons a year (967 acre-feet), restricted to a single completion in either the deep Lance or the Fox Hills formation, with multiple-formation completion expressly prohibited. The campus takes no municipal water.

The applicants’ consistent public position is that the ceiling overstates the draw: the county minutes record that the company “will drill its own well and will utilize closed-loop water resources”; the pre-application file puts the requirement at “approx. 2 million gallons“ reused in a closed system. The Tembo narrative commits to “no ongoing water demand for cooling operations” beyond an initial fill, and at the January 2026 hearing the applicant told the Board each data center fills once with the equivalent of about 20 households’ annual water use and consumes “less than three households” ongoing.

VIII. What to Watch

Source: Laramie County Planning & Development.

Nearly everything unresolved about Project Tembo carries a date, and most of the dates land within a year.

• August 31, 2026 — the $377 million maturity. The generation reservation agreement’s amended term ends, with definitive agreements between Black Hills and Google targeted “by the end of the quarter.” Signature converts the project’s entire commercial base from a refundable bridge into contracts; a further extension keeps the utility fronting supplier payments on customer cash.

• September 1, 2026 — the FERC waiver. Cheyenne Power Hub asked for its waiver order by this date to support a September financing launch, and promised a market-based rate application “shortly.” The generation configuration is “not yet final and may change”; the financing and the fleet harden together.

• December 2026 — the fuel-cell decision. AEP’s amended offtake milestone comes due, with Google holding the option to move the fuel cells elsewhere if Cheyenne “does not advance,” and AEP holding a cost-plus put-back.

• February 15, 2027 — the first metered water. The production wells’ first annual water-use reports are due, replacing the households-versus-ceilings argument with numbers.

• Early 2028 — first service. This is the current target, already moved once from the “Q1 of 2027” that management described in February 2026. Robinson Substation and the Switchgrass lateral are under construction; the M501JAC components were scheduled from mid-2026.

Subscribe now

Measured AI provides institutional-grade analysis of the physical infrastructure powering AI data centers. For access to our full research library, regulatory intelligence, and weekly briefings, visit MeasuredAI.com.

Over two-and-a-half years later, another entity surfaced in New York. On July 27, 2026, a Delaware shell company named after a fried pastry — Sopaipilla Investor LLC — priced roughly $12.5 billion of senior secured notes. The next morning, Meta and BlackRock announced what the money was for: an 80/20 joint venture to build a $14.3 billion, 960 MW, 4-million-square-foot AI data center campus on 1,039 acres in El Paso, Texas.

Source: Meta Platforms.

What makes El Paso, Texas, a case study worth exploring in detail is that, for once, both ledgers are public. Meta engineered one of its largest disclosed campuses so that it would not own it — Meta is the 20% member and the tenant of a venture BlackRock’s funds control — and the rated bond that moved the debt off Meta’s balance sheet also forced the deal’s full economics into view: the lease terms, the guarantees, the rent mechanics, even the sponsors’ return thresholds. That is disclosure the hyperscalers’ self-built campuses rarely produce.

Read against the local record — the covenants El Paso actually signed, the tariff invented for the load, the bridge power plant its utility is buying, the water contract that survives any drought — the two ledgers show, line by line, how the risk of a frontier AI campus gets distributed: bondholders and the electric utility got Meta’s guarantees, and El Paso’s governments got the project SPV, Wurldwide LLC.

This AI data center primer walks through Meta’s El Paso campus ledger by ledger: the bet and the balance-sheet strategy behind it, the deal Wall Street was sold, the campus itself, the gigawatt of power arranged outside ERCOT, the water, and the instruments El Paso signed.

Subscribe now

I. One of Meta’s Largest Data Center Campuses Is One It Won’t Own

Meta’s problem is not demand for compute; it is where to put the cost of the buildings. The company spent $72 billion of capex in 2025 and guided 2026 to $130–145 billion to support its AI efforts and core business.

Meta’s non-cancelable contractual commitments ran from $131 billion at year-end 2025 to $349 billion by June 30, 2026; leases signed but not yet commenced reached $279 billion, plus roughly $68 billion more signed in July 2026 alone.

Long-term debt went from $59 billion to $84 billion in six months, and free cash flow in Q2 2026 was $784 million — against $8.5 billion in the same quarter a year earlier.

A gigawatt-scale AI data center campus like El Paso is what that squeeze produces: a campus Meta needs, pays for, operates, and guarantees, but deliberately does not own.

The Demand Case — “Meta Compute”

The strategic frame is AI. On Meta’s Q2 2026 earnings call, the company described AI usage across its products as still ramping, with a significant share of its compute headed to training models, growing the core business, and delivering personal agents and new products — the demand a campus like El Paso exists to serve.

The El Paso campus belongs to Meta Compute, the internal division building AI infrastructure and pairing “Meta’s infrastructure expertise with capital partnerships that deliver the speed and flexibility its long-term AI ambitions require.” El Paso sits within Meta’s fleet as one of five gigawatt-scale U.S. AI data center campuses — alongside Prometheus in Ohio, Hyperion in Louisiana, Project Domino in Indiana, and the leased QTS Cedar Rapids campus in Iowa — tied together by a backbone network designed to link campuses nearly 1,500 miles apart — not a standalone training cluster but a node in a networked machine.

The Inversion — Tenant, 20% Member, Guarantor

Every gigawatt-scale AI data center campus Measured AI has covered is, underneath, a story about ownership. Amazon owns everything at New Carlisle, Indiana, and points it at one customer. xAI built its own power plant rather than wait for a grid. Microsoft rents an entire gigawatt from a neocloud at Monarch in West Virginia.

El Paso is a different corner of the map: the landlord is an asset manager. Meta holds a 20% interest through Bluebonnet Crossing LLC — which doubles as the campus’s property and asset manager — leases the whole campus back through Roadrunner Leap LLC, and wraps the structure in parent guarantees. BlackRock’s funds, through Global Infrastructure Partners (GIP) and HPS Investment Partners, own the other 80%. Meta supplies the demand, the construction management, and the credit; BlackRock supplies the balance sheet.

Senior secured noteholders and BlackRock-affiliated funds capitalize the note issuer (Sopaipilla Investor LLC), which holds 80% of the joint venture company alongside Meta’s 20%, and the venture itself owns Wurldwide LLC — the landlord on every El Paso instrument. Meta is everywhere except in control: tenant of the whole campus, guarantor of the leases, backstop for overruns and residual value.

Meta’s CFO Susan Li stated the rationale plainly: Meta has “been evolving our capital structure in recent years to include a greater mix of debt as we work to bring down our cost of capital,” and has “broadened our aperture there to include partnerships, like the one we announced with BlackRock.” Pressed by an analyst on how to think about “these off-balance sheet obligations,” she gave the structure’s two selling points — the leases let Meta re-evaluate its compute needs every four years, and “the value of the backstop also goes down over time.”

The Louisiana Template

El Paso is the second run of a structure Meta debuted in October 2025, when it formed a venture with Blue Owl for the $27 billion Hyperion data center campus in Louisiana, financed through Beignet Investor LLC — the deal in which Meta first made use of the residual value guarantee (RVG) in data center construction. The Louisiana disclosures are the template for what El Paso will become in Meta’s filings: a 20% interest, accounted for under the equity method and not consolidated, yet carrying a stated maximum exposure to loss of $46 billion as of June 30, 2026. Meta manages the construction, administers the property, occupies every building, and guarantees the debt service chain; the venture is still not Meta’s, because Blue Owl controls what happens to the campus if Meta ever leaves.

The structure answers a problem every hyperscaler now shares: how to fund the buildings without carrying them. Borrowing through a project entity raises the construction money while keeping the debt, and the data centers themselves, off the parent’s books. El Paso is that answer made concrete, and because its financing was sold as rated public debt, it is the version where every term can be inspected.

II. The Deal: What Wall Street Was Promised

The financing stack is engineered to do two contradictory-sounding things at once: keep the campus off Meta’s balance sheet, and make sure bondholders face nothing but Meta.

The Stack — Sources and Uses

The structure raised $16.8 billion — about $17.5 million for every megawatt of the campus’s 960 MW of IT capacity — and nearly three-quarters of it is borrowed: $12.3 billion of senior secured notes against $4.2 billion of combined equity checks from BlackRock’s funds and Meta. All the leverage sits on BlackRock’s side. Its funds’ $11.5 billion goes into the venture through the note issuer, which borrows $12.3 billion against $1.3 billion of fund equity; Meta funds its 20% with its $2.9 billion contribution alone, unlevered. That is why the minority member writes the bigger equity check. Almost all of it goes into the ground, funding construction of the data centers.

At closing, Meta contributes roughly $2.3 billion of land and construction-in-progress, BlackRock contributes about $4.9 billion in cash, and Meta takes back a one-time distribution of approximately $1 billion to align ownership stakes — a net day-one contribution of about $1.3 billion, with the rest of Meta’s $2.9 billion commitment funded over the course of construction. Meta pays no rent until December 2028 while the notes accrue from 2026, so more than two years of coupon — $1.9 billion — is capitalized into the raise.

The venture’s own economics are disclosed too. Distributions run pro rata to BlackRock’s issuer and to Meta until the issuer clears an 8.25% IRR and a 2.4× multiple on invested capital — after which Meta’s reimbursements for cost overruns, power costs, and repair costs jump the queue ahead of it.

The Leases and the Guarantees

The leases convert a four-year commitment into two-decade-long paper. Meta’s subsidiary, Roadrunner Leap, signs eight leases — six triple-net, one for each pair of the 10 data center buildings, and one for the network building, plus two gross leases for the administration and infrastructure business operations support (IBOS) buildings.

Rent commences December 1, 2028, date-certain, irrespective of construction status — a fixed base rent for the first 13 months, set to a 50% utilization assumption, then reset annually on actual utilization with a floor, escalating 2% a year from 2031. The initial term is four years with four four-year renewals (i.e., Meta’s options extend 16 years total); the notes mature in November 2048.

Construction risk sits with Meta: cost overruns beyond 105% of budget are Meta’s, and it has no termination right even for prolonged construction delay.

Source: Meta Platforms.

Above the leases sits the guarantee stack, all of it running to Meta Platforms, Inc. itself: a payment guarantee of the tenant’s lease obligations, a parent guarantee of the landlord’s power contract with El Paso Electric, a guarantee of the termination fee (the discounted value of remaining rent), and the residual value guarantee — an aggregate threshold of approximately $13 billion, declining over time. Per megawatt, that is $13.5 million guaranteed against $14.9 million built: Meta backstops roughly 91% of an asset it deliberately does not own.

The Bond Market’s Verdict

The notes were rated as Meta credit in a project-finance wrapper. S&P assigned A+, one notch below Meta’s AA- corporate rating, explicitly because the noteholders have no lien on the buildings, the land, or the project contracts — only a pledge of equity interests and accounts. Fitch and KBRA, by contrast, rated the paper AA-, level with Meta’s corporate rating. The contractual rent floor produces a flat debt service coverage ratio of just 1.12× — thin by design, because the real coverage is Meta’s guarantee stack, not the project’s cash flow.

The market took the paper, at a price. The order book peaked around $20 billion — about 1.6× the deal, the weakest subscription of any major AI investment-grade offering of 2026 by Bloomberg’s ranking, against a roughly 4× average — and syndication ran nearly a week without the spread tightening, both rarities in high-grade bonds. The notes priced July 27 at 7.53%, 2.875% over 10-year Treasuries, the widest spread for a bond rated A or better in three years. The launch-day upsize of about $273 million took the deal from the $12.3 billion the agencies rated to the ~$12.55 billion that sold.

That gap is the trade. Meta moves $12.3 billion of construction debt off its own balance sheet, and pays for it in the coupon — a bond the agencies rate as Meta credit, sold about two percentage points wider than comparable A-rated corporate paper.

That is the ledger on paper. What it bought is already rising in the desert.

Subscribe now

III. The Campus: 960 MW and 4 Million Square Feet

The rating file specifies that campus building by building.

Meta’s 960 MW of total IT capacity comprises 10 data center buildings at 90 MW plus one 60 MW network building, across approximately 4 million square feet of built area.

Source: Meta Platforms.

Around the buildings sit three switching stations, six substations, and stormwater retention basins, on 1,039 acres on the north side of Stan Roberts Sr. Avenue, west of U.S. 54, in the desert of Northeast El Paso.

Source: El Paso Electric Company.

Construction is well underway and predates the joint venture by nearly a year. Meta broke ground on October 15, 2025, and by the July 2026 announcement, over 2,300 workers were on site, on the way to a claimed peak above 4,000 workers.

Source: Meta Platforms.

The named builders are Hensel Phelps and JE Dunn — the general contractors — joined by Faith Technologies, the electrical and modular-power specialist, with Stantec as the design firm. Meta expects to begin bringing capacity online in 2028.

All of it depends on the one input the venture cannot build for itself: power.

IV. The Power: A Gigawatt Outside ERCOT

The most consequential fact about El Paso’s power is the grid it is not on. El Paso Electric operates in the Western Interconnection (WECC), not ERCOT — so none of the machinery Texas has built around data center load applies here: no ERCOT interconnection queue, no ERCOT large-load rules, no ERCOT price formation.

Source: ERCOT.

The scale mismatch is the story. El Paso Electric’s 2025 system peak was 2.3 GW; its entire industrial class is 55 customers consuming about 1.07 TWh a year. The campus’s contract commits El Paso Electric to deliver up to 1.0 GW — more than 40% of that system peak — with a right to request 1.4 GW.

Five capacity figures circulate around this Meta data center project, and each measures something different:

The result is that a gigawatt-scale retail load arrived through three instruments: a bilateral contract, a bespoke tariff, and a bridge power plant.

The Contract and the Tariff Invented for It

The electric service agreement — executed March 18, 2026 — runs to 2048, the same year the bonds mature. Its key terms: committed capacity up to 1.0 GW; the 1.4 GW option conditioned on three substations, added transmission and interconnection, and 225 MW of on-site generation capacity; minimum demand of 85% of monthly contract demand; and a Meta parent guarantee. The contracting party is Wurldwide (the landlord) — but El Paso Electric’s counterparty risk is Meta’s credit, exactly as the bondholders’ is.

The retail rate came first, and it came through a side door. El Paso Electric’s Economic Development Rate rider — a mechanism for negotiating rates outside a rate case — names data centers as the only eligible sector exempt from its jobs requirement (two employees per 100 kW for everyone else). What substitutes is a condition that welds the tariff to the local ledger: a data center customer must hold, and annually re-certify, Chapter 380 or 381 incentives from local government.

In August 2024, El Paso Electric applied to create Schedules 33A and 33B for Wurldwide under that rider, describing a customer with “between 100 and 200 MW” of peak load; its modeled scenarios topped out at 179 MW and 1.41 billion kWh a year — by itself more than El Paso Electric’s entire existing industrial class consumes. The Commission approved the settlement in July 2025 for a five-year term, El Paso Electric insisting the rates are “cost based” rather than discounted — and PUCT Commissioner Courtney Hjaltman added a condition from the bench: any revenue shortfall gets booked as a regulatory asset and reckoned with in the next base-rate case. The schedules were always designed to expire into a dedicated data center rate class once the load shows up in a test year.

That rate class now has a docket. On April 7, 2026, El Paso Electric filed Schedule 27 — High Load Factor Large Power Service, a permanent tariff for any customer with peak demand of 75 MW or more and an annual load factor of at least 85%, and named its prototype: Wurldwide, “expected to initiate service in July 2026,” ramping to 1 GW at a load factor above 90%. The filing carries Schedule 33A’s rates forward, terminates 33A on approval — and attaches the electric service agreement. The Commission sent the tariff to a contested hearing, questioning whether a 20-year minimum initial term is appropriate, whether the service agreement should really escape review, and what conditions keep residential and small-commercial customers whole under a tariff that bills its customer “regardless of whether the customer has energized its facilities.”

The Load That Outran Every Forecast

What the tariff record never priced is how fast the load would grow. El Paso Electric’s rate-case load forecast of April 2024 contained no data center line at all — 372 MW of total growth over 10 years. Six months later, its planning studies carried 168 MW of new large load; the September 2025 New Mexico resource plan carried 533 MW at a 97.6% load factor; by January 2026 an internal study showed 1.16 GW of large-load demand by 2029; and by June 2026 El Paso Electric’s own rebuttal exhibit walks Wurldwide alone from 16.4 MW in 2026 to 1 GW in 2029 — then holds it there for the next fourteen years.

El Paso Electric told federal regulators, in a June 2026 forecast, that its summer peak will rise 77.7% and its energy load 132.5% by 2035, with a single 711 MW step between 2027 and 2028. Behind it all sits a pipeline El Paso Electric described to New Mexico stakeholders in April 2025 as 8 GW of data center inquiries, 5.4 GW of them active — against a utility whose peak is 2.3 GW.

The procurement response is already open. El Paso Electric’s 2025 all-source request for proposals seeks 650 MW to 850 MW of accredited capacity for its own load growth through May 2031 — and, separately, up to 1.65 GW more for potential large-load customers, a bucket the utility ties to the surge in service inquiries. None of it is dedicated Meta generation; it is the portfolio a 2.3 GW utility now has to assemble around a gigawatt customer.

Subscribe now

The Gas Bridge Plant El Paso Electric Owns

The physical answer to the 2027 gap — load arriving before the new substations and 345 kV lines can serve it — is McCloud Generation, a natural gas plant built by Enchanted Rock — a Houston company that makes modular natural gas generators and sells them as backup power, as dispatchable capacity, and, as here, as “bridge” power for customers who cannot wait for a grid connection. The plant’s ownership structure is the unusual part: McCloud is not Meta’s plant and not Enchanted Rock’s — El Paso Electric will own the plant and the land, with Enchanted Rock engineering, building, and operating it under contract.

It is 366 MW nameplate, built from 813 Enchanted Rock 450 kW rich-burn natural gas engines in clusters of four or five, arrayed in roughly 168 rows across ~31 acres adjacent to the campus — air-cooled, with a peak water demand of about 4,000 gallons a day for safety showers and potable use, none of it for cooling.

Illustrates the dense, modular natural gas generator configuration expected at McCloud. Source: Enchanted Rock.

The reciprocating internal combustion engines (RICE) are fired exclusively on pipeline-quality natural gas, with no on-site fuel storage: the plant burns what the pipeline delivers, as it delivers it. Enchanted Rock’s dimensioned layout drawing (below) shows what El Paso Electric is buying: an engine farm and a reserved maintenance yard.

Source: Enchanted Rock.

During a bridge period of up to five years, it runs islanded — electrically connected to the customer’s load and to nothing else — delivering 225 MW of continuous power (modeled at a 95% annual load factor) by mid-2027, its capital and operating costs recovered from Wurldwide through a dedicated rate. Grid-connected by roughly 2032, it is rated 343 MW firm. The 225 MW matches the on-site capacity condition in the 1.4 GW option.

The ability to deliver by 2027 was key for El Paso Electric — a schedule only a modular vendor could meet.

Commitment preceded certification: El Paso Electric signed a limited notice to proceed in September 2025 with a first payment of $29.95 million, and Enchanted Rock obtained the air permit in October 2025 — a TCEQ standard permit. It authorized 225 MW by capping natural gas consumption, thereby avoiding Title V and PSD review. Both steps came months before El Paso Electric filed the Certificate of Convenience and Necessity (CCN) application in December 2025. The permit registration names Meta.

The Cost, and the Fight Over It

The cost figures for the McCloud bridge plant are as follows:

$/kW is computed at 343 MW firm capacity.

On the intervenors’ all-in figure, the plant costs $1,609/kW against $1,264/kW for a generic combustion turbine in El Paso Electric’s own resource plan; its modeled heat rate is a flat 12,000 Btu/kWh against roughly 9,300–9,500 for El Paso Electric’s recent units, worth $28–36 million a year in extra natural gas for the same energy, and its non-fuel operations and maintenance (O&M) runs 3.5× that of El Paso Electric’s recent gas fleet.

El Paso Electric concedes every differential and relocates the burden: all of it is Wurldwide’s to pay during the bridge. The catch is the exit. The application says El Paso Electric “will” fold the plant into its system portfolio afterward; its brief retreats to “may be beneficial… no decision has been made” — and a model El Paso Electric produced in discovery shows what folding McCloud’s costs through the standard recovery rider would do: a residential bill increase of 8.43%. El Paso Electric has not proposed that rate; the intervenors are litigating to make sure it never can.

That is why the CCN is contested. Of seven parties, the City of El Paso — whose expert proposes six conditions of its own, a $565.6 million cost cap among them — the state’s residential ratepayer advocate (OPUC), Freeport-McMoRan, and an individual intervenor urge denial; TIEC (representing large industrial customers) would grant only with conditions; Commission Staff makes no affirmative recommendation and two adverse findings; El Paso Electric alone asks for a clean grant. The Sembrando Esperanza Coalition filed 1,210 signed protest forms into the docket. The hearing ran July 8–9, 2026; initial briefs are in; the statutory deadline for a decision is December 9, 2026.

The Wires, and the Money Already Spent

Interconnection is the structural finding of the whole record. In El Paso Electric’s federal tariff, a new generator larger than 20 MW faces cluster windows, application fees, study deposits, site control requirements, and withdrawal penalties. A new load of any size faces Section 31.2, whose entire governing standard is that the customer provide “as much advance notice as reasonably practicable.” No megawatt threshold, no queue, no fee, no deposit, no public posting. A gigawatt of demand entered the system through a provision written for an era that never imagined it, which means the terms were set bilaterally, in contracts, rather than by any published procedure.

The contracts came early, and the money is already flowing. In June 2024 — before any public announcement, and two months before the tariff application — El Paso Electric signed an agreement with “a potential large load customer” to build and expand transmission for “a planned data center,” amended in October 2025 to add the expansion of one existing substation and construction of two new ones. The customer paid El Paso Electric $144.5 million of non-refundable construction advances during 2025 alone. El Paso Electric’s financial filings never name that customer; the electric service agreement with Wurldwide followed the interconnection deal by more than 18 months.

On El Paso Electric’s own Northeast-area expansion map (below — black lines are existing, red lines are new/upgrades), the buildout is unmistakable: McCloud East substation in service January 2026 to serve the Large Load, McCloud Central and McCloud West 345 kV substations in January 2028, a ±200 MVAR STATCOM — six times larger than any other reactive device in El Paso Electric’s 10-year plan — in May 2028, and new 345 kV lines to Pine (May 2029) and Vado (May 2030).

Source: El Paso Electric Company.

The Pine–McCloud Central 345 kV line moves to routing now: a CCN filing in September 2026, targeted approval March 2027, construction from April 2028.

Source: El Paso Electric Company.

The Gas: The One Input Still Dark

Gas is the one input still dark. El Paso Electric designated its gas-supply answer as the docket’s only Highly Sensitive filing; Enchanted Rock told FERC the McCloud gas upgrades “fall outside of FERC’s jurisdiction as intrastate facilities”; and the single named counterparty is Kinder Morgan, whose delivery fees of roughly $8.5 million a year sit among the costs El Paso Electric excludes from its O&M estimate.

What El Paso Electric sealed, Kinder Morgan’s own tariff bulletin board largely answers: in March 2026, its El Paso Natural Gas (EPNG) pipeline ran a binding open season for a “McCloud Lateral” — nearly six miles of 16-inch pipe from its mainline to a new delivery meter station at the plant, offering 167,000 dekatherms a day of firm capacity from January 1, 2027 — and subscribing the whole pipe at the notice’s recommended rate comes to almost exactly the $8.5 million a year in the CCN record.

Source: El Paso Natural Gas.

The floor under all of it is math: run flat out, 225 MW at McCloud’s modeled 12,000 Btu/kWh heat rate burns roughly 65,000 dekatherms — about 60 million cubic feet — of natural gas a day, delivered as it is burned, to a plant with no storage, while the balance of the campus’s gigawatt arrives over El Paso Electric’s wires.

Power the venture had to arrange in public, docket by docket. Water it had settled two and a half years before the bonds priced, on the same December day El Paso’s governments signed everything else.

V. The Water: “Regardless of Any Drought”

The agreement El Paso Water’s board approved on December 4, 2023, for $100 of consideration, reserves 2.5 million gallons per day (MGD) of water capacity across three tiers and up to 1.25 MGD of wastewater for Wurldwide, and designates the campus a “high priority economic development project” whose meters will be approved “regardless of any drought or Water emergency.” The utility cannot reduce the reservation without the customer’s consent, “which consent may be granted or withheld in Customer’s sole discretion.” El Paso Water commits that it “will not discriminate against the data center industry” and will adopt no data-center-specific rates. The customer may terminate at any time, for any reason.

The water utility had been building toward the site for years. In October 2023, two months before the agreement existed, El Paso Water awarded a $4.96 million contract for a 16-inch water main and 30-inch sewer interceptor terminating at Stan Roberts and Dyer — the campus’s corner. The agreement’s own exhibit maps the rest of the program (below): phased interceptors up the Dyer/Railroad corridor and along Stan Roberts itself, an order-of-magnitude upgrade of the NE Dyer/Railroad lift station (0.5 to 5 MGD), and the campus’s wastewater delivered to the Fred Hervey plant — works the agreement obligates El Paso Water to “design, construct and fund.”

Source: El Paso Water.

The utility’s own account holds that the prior data center design abandoned once-through water cooling for an air-cooled system that sprays water only on the hottest days. The rest of the record suggests the campus will rarely need the 2.5 MGD it reserved: Meta’s published description of a closed-loop, liquid-cooled design (liquid at the racks, air at the heat rejection) using “zero water for a majority of the year,” with annual use “similar to a typical golf course in west Texas”; the 0.48 MGD figure El Paso Water modeled for Council in July 2026; and Meta’s standing pledge — the company’s standard policy in high-water-stress watersheds — to restore 200% of consumed water.

That agreement was one of several instruments El Paso signed in the same stretch of weeks. The rest are on the other ledger.

Subscribe now

VI. The Other Ledger: What El Paso Signed

The instruments El Paso executed in 38 days in late 2023 bind its governments for 35 years — and against the bond file, they read like documents for a different project. Wall Street’s ledger is denominated in billions and secured by Meta Platforms, Inc.; El Paso’s is denominated in 50 jobs and $800 million per phase, and signed by the project SPV, Wurldwide LLC.

Start with the land. On December 29, 2023, the City of El Paso sold the campus site to Wurldwide for about $8.5 million, equivalent to $8,156 per acre. The deed carries two protections the City’s negotiators did win: a “Springing Repurchase Right” — develop within five years or the City may repurchase at the original price, or take $5 million in liquidated damages — and a full reservation of the groundwater estate to the City.

Then come the four incentive agreements, two per government, approved on December 4 and 5 — weeks before the deed:

Each instrument runs at 80%, across up to five phases that “may run concurrently or sequentially,” under 35-year terms. The abatements rest on City of El Paso Reinvestment Zone No. 1, created November 21, 2023, with a five-year, renewable life. The zone expires November 21, 2028, a lapse that cannot undo the signed agreements (Tax Code § 312.203) — it ends only the deal’s ability to take in new land.

What did El Paso get in covenant form? An investment commitment of $800 million per phase — a contractual floor that, across all five phases (i.e., 10 data center buildings), comes to $4.0 billion against a $14.3 billion announced build. And an employment commitment of exactly 50 full-time jobs — cumulative across all phases, due within four years of a completion deadline that is itself seven years from the start of vertical construction. “Full-Time Job” is defined to admit sub-40-hour schedules and remote employees who qualify as “Regional Residents”; local hiring is a non-binding “goal” of 25 of the 50, drawn from within a 50-mile radius; the wage floor is the area median. City staff’s presentation told Council to “expect 200+ FTEs at full ramp-up” — but the economic model it presented ran on the covenanted 50.

The City of El Paso has put numbers to what it signed: roughly $110 million of combined City–County incentives for Phase 1, up to $550 million across all five phases, projected net collections of $83.4 million per phase, and up to $5 million in funding for intersection and roadway improvements at Stan Roberts and US-54.

The ledger runs to Austin too, and the state’s line is the only one that reaches the power bill. On September 17, 2025 — four weeks before groundbreaking — the Texas Comptroller registered “Wurldwide LLC DBA Statue LLC” as a Qualifying Large Data Center Project: a 20-year exemption from state and local sales and use tax on qualifying purchases — servers, cooling and electrical systems, generators, and the electricity itself, through 2045. The statute’s thresholds are 250,000 square feet, $500 million of investment within five years, and 40 jobs in the county — the covenanted 50 clears the state’s bar by ten. Wurldwide is registered as the project’s owner, occupant, and operator.

For now, the property-tax side is still theoretical: the appraisal district carries the two parcels at $5.4 million for 2026, and the 80% abatements have nothing to abate until the buildings land on the books. The sales-tax exemption, by contrast, has been live since September 2025, while the campus is still being built.

The Rules Written for the Next One

On July 20, 2026, City Council adopted El Paso’s first Data Center Policy Framework — the product of a six-month process Council ordered in February 2026: six community meetings and more than 2,300 public comments. The City’s own summary of that input records “repeated requests for the City Council to cancel the 80% tax abatement.”

Council did not cancel anything. What it adopted is aimed at the next project: future hyperscale data centers should face a Special Permit — public notice, Plan Commission review, a public hearing, final Council action. The framework also calls for enhanced standards for water, energy demand, wastewater, noise, lighting, and ecological impact; community benefit agreements; parent-company commitments; protections against infrastructure costs migrating to existing ratepayers; and no local economic incentives at all. Every element is scoped to future projects. The framework is a roadmap, not yet law — each piece returns to Council as ordinance amendments — but its premise is already plain: the Meta deal is one El Paso does not intend to offer twice.

VII. What to Watch

Source: Meta Platforms.

Most of what happens next is already on a docket, a filing calendar, or a lease schedule. The markers, in rough order:

• The McCloud decision — by December 9, 2026. A proposal for decision is expected around early October 2026; the statutory deadline is December 9.

• The closing paper trail — Q3 2026. Meta’s 10-Q said closing was expected in the third quarter; the first post-closing disclosure will be Meta’s Q3 filing, due at the end of October 2026.

• The wires — September 2026 through 2030. El Paso Electric files for the Pine–McCloud Central 345 kV line around September 2026, targeting approval in March 2027 and construction through March 2029; McCloud Central and West substations are scheduled for January 2028 and the 200 MVAR STATCOM for May 2028. These are the physical preconditions of the 1.4 GW option — three substations and 225 MW of on-site capacity.

• December 1, 2028 — the machine turns on. Rent commences date-certain whether or not construction is done; the venture expects first capacity online in 2028.

Subscribe now

Measured AI provides institutional-grade analysis of the physical infrastructure powering AI data centers. For access to our full research library, regulatory intelligence, and weekly briefings, visit MeasuredAI.com.

Consider the two ends of the category. CoreWeave, the largest pure-play neocloud, reported a revenue backlog of $99.4 billion as of March 31, 2026, to be delivered from 49 data centers — essentially all of them leased from third parties, with a first self-build not due online until later this year. Fluidstack, one of the smallest of the six companies profiled here, operates five U.S. sites totaling roughly 1.4 GW of contracted critical IT capacity and owns none of the buildings; at all five of the locations, Google has guaranteed the lease obligations so that the landlords could borrow against them. One company carries a hundred-billion-dollar order book on rented floors. The other has made rented floors into its entire business model, with a hyperscaler’s balance sheet standing behind the rent.

Source: Core Scientific.

The prize is large and compounding. Neocloud category-wide revenue stands at ~$12 billion a quarter as of mid-2026 — a ~$48 billion annualized run-rate — and, on our forecast, reaches $300 billion by 2030, a ~50% compound annual growth rate. Those figures cover the whole category; the six companies profiled here — Crusoe, Nebius, Nscale, CoreWeave, Lambda, and Fluidstack — make up the vast majority of that revenue and are collectively valued at over $150 billion, with a next tier already signing leases behind them.

What organizes this cohort is not chips or software. It is ownership. The six neoclouds arrange themselves along a spectrum that runs from a company that builds power plants down to a company that owns nothing but GPUs and contracts — and underneath every point on that spectrum sits the same load-bearing structure: landlords who own the power, neoclouds who hold the lease, and anchor customers who increasingly co-sign the obligations of the layer below them. The neocloud layer scales precisely because the demand side of the market has agreed to underwrite the supply side.

Subscribe now

I. GPU Demand Outran the Hyperscalers, and a New Layer Formed to Catch It

A neocloud sells dedicated GPU capacity in large contracted blocks. That is a demand shape the hyperscalers’ general-purpose clouds were not architected for: AI labs and enterprises do not want elastic instances by the hour; they want thousands of interconnected accelerators, reserved for years, delivered on a date. The deals are bilateral multi-year contracts that run from tens to hundreds of megawatts, the largest of them covering roughly 100,000 GPUs. The differentiators are speed-to-capacity and a willingness to sign single-tenant, single-site commitments that a general-purpose cloud cannot justify.

Source: Nebius.

The economics of GPU rental pricing have held up better than the glut narrative suggested. SemiAnalysis’s NVIDIA H100 GPU rental index rose roughly 36% between May 2025 and May 2026, from about $1.95 to $2.65 per GPU-hour on one-year rental terms. This is a seller’s market renting scarce capacity, not a commodity spot market.

II. Three Layers Deep: Landlords Own the Power, Neoclouds Hold the Lease, Customers Guarantee the Rent

The Stack, Bottom to Top

Layer 1 is the powered-shell landlord, and in this cohort it is predominantly a converted bitcoin miner: TeraWulf, Cipher Digital, Hut 8, Core Scientific, Applied Digital, Galaxy, IREN, and Ionic Digital. These companies own what is scarce — energized land, interconnection positions, substations — and they learned to monetize megawatts in the crypto era.

Layer 2 is the neocloud: Crusoe, Nebius, Nscale, CoreWeave, Lambda, and Fluidstack. They sign 10–15-year leases on those shells and fill them with GPU assets whose premium-pricing window runs two to five years. That duration mismatch is the structural fact of the business model, and Section V returns to it.

Layer 3 is the anchor customer — Microsoft, Google, Meta, OpenAI, Anthropic — which increasingly does something anchor tenants in ordinary commercial real estate never do: it guarantees the obligations of the layers below it.

Case Study: A Landlord Built for Nscale and Microsoft

One neocloud lease by Nscale shows the whole machine. Ionic Digital was formed in January 2024 out of the Celsius Mining bankruptcy; its Ward County, Texas site mined bitcoin until the rigs were decommissioned in Q4 2025. In mid-October 2025, Ionic signed a 10.5-year triple-net lease with neocloud Nscale’s subsidiary Nscale Ward County LLC, covering an original 234 MW with about $1.95 billion in contracted revenues.

At the same time, Nscale announced it would deploy approximately 104,000 NVIDIA GB300 GPUs in this Ward County, Texas data center, with phased delivery of NVIDIA AI infrastructure services to Microsoft from Q3 2026.

Ionic Digital’s Ward County, Texas site showing 1) Data Hall Cluster, 2) Ionic Substation, and 3) TNMP Transmission Substation. Source: Ionic Digital.

A February 2026 amendment added 89 MW that lifts the lease to 323 MW and roughly $2.6 billion in contracted revenues — contingent on Ionic actually delivering the additional capacity, which it expects in the second half of 2027. Base rent is $65 per kW per month on gross power ($90/kW-mo critical-IT equivalent at an assumed 1.4 PUE), escalating 3% annually after year five. Cash rent starts in August 2026.

In this triple-net lease the tenant, Nscale, bears taxes, insurance, maintenance, and construction. So Ionic’s rent is close to pure yield on its power position — the site’s West Texas electricity runs around $0.05/kWh, and the lease even obligates the tenant to apply for Qualified Operator status under Texas’s data center tax regime: a contract term dressed as regulation.

Above the lease sits the guarantor layer, priced: NVIDIA bought $60 million of Nscale warrants and, as consideration for the warrants, guaranteed the first five years of Nscale’s rent up to $860.3 million — a rent guarantee, not a construction commitment — with Nscale’s parent guaranteeing the rent from year six.

If Nscale becomes insolvent, its anchor customer, Microsoft, may step into the lease within 30 days. And Microsoft holds an option on additional Ward County power starting in late 2027, on a property whose path runs to roughly 700 MW (323 MW contracted plus a tenant right of first refusal on 377 MW more).

Landlord, neocloud, chipmaker, hyperscaler: every layer named and priced in one lease.

Ionic’s Nasdaq direct listing is the first landlord flotation built on a single neocloud lease.

Why the Guarantor Layer Exists

No lender will finance hundreds of millions of dollars of single-purpose construction against the unsecured promise of a private company founded a few years ago. So the credit reaches down the stack from whoever has it: the chipmaker guarantees the rent, the hyperscaler backstops the lease, and the landlord borrows against the guarantee rather than the tenant’s credit. Section IV prices that layer deal by deal.

Subscribe now

III. The Spectrum Runs from Power Plants to Pure Tenancy

The defining variable for this neocloud cohort is physical-asset ownership, and the six companies are arranged below on a descending ladder: the Power-First Builder (Crusoe), the Self-Builder (Nebius), the Campus Acquirer (Nscale), the Leased Giant (CoreWeave), the GPU Veteran (Lambda), and the Pure Tenant-Operator (Fluidstack).

The order is the argument: each rung down, the company owns less of what it operates, and the external credit propping up the structure matters more. The ladder lands on Fluidstack, whose entire model is a preview of Section IV. One timing fact is worth noting: every owned gigawatt-scale campus in this section — Independence, Butler Township, Monarch — delivers first power in 2027 or later, while essentially every floor operating today is rented.

One table carries the breadth of the cohort:

Crusoe — 4.9 GW Contracted, One Stargate Built (The Power-First Builder)

Crusoe is the one neocloud that builds generation. Founded in Denver, Colorado, in 2018 by Chase Lochmiller and Cully Cavness, it began as a flare-gas company — capturing stranded natural gas at oil wells to power modular data centers on site — and that energy-first DNA survived even after Crusoe sold the original flare-gas and bitcoin business to NYDIG in June 2025 to focus entirely on AI.

Source: Crusoe.

Crusoe has contracted 4.9 GW of power, 13.6 million square feet of buildings, and 3,000 acres across its campuses — against a pipeline it puts above 40 GW.

Crusoe has exited its development-partner role on the 1.8 GW Project Jade in Cheyenne, Wyoming. Table also excludes Crusoe Cloud deployments in Houston, Texas; Ashburn, Virginia; and Sparks, Nevada.

Crusoe’s footprint is the most hyperscale-shaped in the neocloud cohort. Crusoe built the Abilene, Texas Stargate campus, live since September 30, 2025, on Oracle Cloud Infrastructure under a 15-year Oracle lease serving OpenAI. Each of the campus’s eight buildings is designed to run up to 50,000 NVIDIA GB200 GPUs in NVL72 systems on a single network fabric; Oracle — not Crusoe — supplies and installs the chips, and began delivering the first racks in June 2025.

Source: OpenAI.

In March 2026, Microsoft signed for 900 MW across a second Abilene campus — two buildings totaling 672 MW of critical IT plus an on-site power plant — taking the site’s projected total to 2.1 GW when Microsoft’s portion energizes from mid-2027. A 1 GW campus at Childress, Texas, followed, and in June 2026 Meta signed for roughly 1.6 GW combined across Childress and a new Warrenton, Missouri site.

Beyond those projects, Crusoe is developing a 1 GW campus for Google at Claude/Goodnight, Texas, and a campus of up to 1.5 GW near Bloomington, Texas, at the Port of Victoria, whose tenant remains undisclosed and whose power is slated to come from Blue Energy.

Crusoe calls itself an “AI factory company,” and its distinguishing asset is a procured generation fleet: 155 gas turbines and gensets — GE Vernova, Boom Supersonic, PROENERGY, Bergen Engines, and Solar Turbines units — totaling about 3.8 GW of nameplate capacity across its portfolio. At Abilene, the >350 MW Longhorn plant was commissioned with the initial campus, and Microsoft’s 900 MW on-site power plant is under development.

Source: GE Vernova.

Although Crusoe is a power-first builder, some of the power behind its campuses is partner-owned or grid-supplied, so it is best understood as a developer that treats generation as part of the product, not a utility.

The financing hallmark is a $1.375 billion Series E closed in October 2025 at a valuation above $10 billion (Valor and Mubadala co-leading), with talks ongoing for roughly $3 billion more at about a $30 billion valuation as of July 2026.

What Abilene shows about the archetype is Crusoe’s actual long-term stance: it doesn’t own the majority of its flagship project. On October 11, 2024, Crusoe sold 92.3% of the campus joint venture to Blue Owl’s real-estate arm, retaining 7.7% and a $13 million development fee. The power-first builder, in other words, monetized the real estate the moment it could — and kept the development role. Childress, energizing next with Meta’s capacity, is where the model’s next test runs.

Nebius — A $19.4 Billion Microsoft Anchor in a Building Someone Else Owns (The Self-Builder)

Nebius is the cohort’s self-builder — with an asterisk that turns out to be the interesting part. The company is the renamed Yandex (Amsterdam-headquartered; Nasdaq: NBIS since its 2024 re-listing), which divested its Russian businesses in May 2024 and redeployed the remainder — and a very large balance sheet — into AI infrastructure.

Excludes Nebius’s planned colocation data center sites in Minnesota and Oklahoma.

Nebius ended Q1 2026 with roughly $9.3 billion in cash after a $6.3 billion quarter of raises — including $4 billion of convertible notes issued in March — and in July 2026 added a $775 million MUFG-led GPU-backed term loan, the company’s first secured debt. Management guides to a $7–9 billion annual recurring revenue run-rate by the end of 2026 against a backlog above $20 billion.

The owned footprint is real: Nebius broke ground May 12, 2026, on a 400-acre campus it owns in Independence, Missouri, sized at up to 1.2 GW with first power expected in October 2027.

Source: Nebius.

The company announced a second owned gigawatt-scale greenfield data center project in Butler Township, Schuylkill County, Pennsylvania, with 1.2 GW of power at completion and initial phase capacity of 260 MW expected to be operational by October 2027.

Source: Nebius.

Nebius is also in the early construction phase, with foundations going in at its two-building AI data center campus on 79 acres in Birmingham, Alabama, sized at up to 300 MW. No other company of the six is self-building at that scale in the U.S.

And yet the flagship is a building Nebius does not own. The Vineland, New Jersey facility, being developed by DataOne at up to 300 MW, follows what Nebius itself describes as a build-to-suit model: a developer owns the land, secures the power, and builds the shell; Nebius leases the result. Vineland is where Microsoft anchored the company: a five-year deal signed September 8, 2025, worth $17.4 billion through 2031, expandable to about $19.4 billion — reportedly more than 100,000 NVIDIA GB300 chips delivered in nine tranches, in a building Nebius says is dedicated solely to Blackwell-architecture GPUs, with liquidated damages if Nebius delivers late.

The first tranche went live in November 2025. So the self-builder’s largest contract rides on delivery milestones inside a leased building — construction risk without ownership, which is precisely the trade the rest of the neocloud cohort makes everywhere.

Source: Nebius.

Nebius struck a deal with Meta in March 2026 spanning multiple locations, worth up to $27 billion — $12 billion of dedicated capacity, one of the first large-scale NVIDIA Vera Rubin platform deployments delivering from early 2027, plus a $15 billion option structured against unsold capacity.

In July 2026, Nebius signed a $1 billion-plus capacity agreement with Reflection AI.

Nebius’s Kansas City origin site (~5 MW leased in a local colocation facility in November 2024, with a path to 40 MW) shows how fast the ladder gets climbed: from five megawatts of rented colocation to two owned gigawatt campuses in under two years.

Nscale — 1.35 GW for Microsoft at Monarch (The Campus Acquirer)

Nscale is the article’s thesis expressed in a single company: owner in one state, tenant in another. The company grew out of bitcoin miner Arkon Energy’s data center business and now sits under a UK corporate structure. Norwegian investment company Aker ASA initially put in $100 million and then a further $350 million of cash across two rounds and, at 22.8%, is the largest shareholder, calling Nscale its second-largest asset.

Nscale’s anchor customer, everywhere, is Microsoft: a company barely two years past launch now supplies Microsoft capacity in three states (West Virginia, Texas, and North Carolina) and even four countries, with contracts in Portugal ($10 billion), Norway ($6.2 billion), and the UK that are outside the scope of this U.S.-focused article.

In West Virginia, Nscale owns. It acquired the Monarch Compute Campus outright in Point Pleasant, Mason County, and Microsoft signed a letter of intent for up to 1.35 GW of AI compute there — specified as NVIDIA’s next-generation Vera Rubin NVL72 systems, delivered in tranches beginning in late 2027; initial site power of roughly 2 GW is targeted for the first half of 2028, on a property Nscale describes as having more than 8 GW of runway. Microsoft rents the whole stack; Nscale owns everything — the full breakdown is in Measured AI’s Microsoft Monarch West Virginia analysis.

Source: Nscale.

In North Carolina and West Texas, Nscale rents. On November 22, 2025, Nscale signed a master services agreement with WhiteFiber for 40 MW of critical IT load in two 20 MW phases in Madison, North Carolina, worth about $865 million over a 10-year term, an implied ~$157/kW-mo of critical load in year one. It is structured as a license rather than a lease — modified-gross, with electricity, PUE, and property taxes passed through, a 3% annual escalator, and a $70.6 million Nscale prepayment.

Source: WhiteFiber.

Nscale’s Ward County, Texas lease is the Section II case study; Nscale is Ionic Digital’s only tenant.

The credit architecture is the point: NVIDIA’s $860.3 million five-year rent guarantee — part of what NVIDIA paid for its $60 million of Nscale warrants — is what made a first-of-its-kind landlord financeable, and Microsoft’s option on additional Ward County power from late 2027 means the customer layer reaches down this stack three times, in three states. The same company that owns a campus Microsoft merely intends to rent is, 1,800 miles away, a rent-guaranteed tenant on a bitcoin mine conversion. That all three structures serve the same customer is the clearest evidence that the spectrum in this section is a set of financing choices, not business models.

NVIDIA, meanwhile, runs through Nscale as both supplier and guarantor: early access to Vera Rubin silicon and roughly 200,000 GB300s planned across the Microsoft footprint (~104,000 in Texas, ~52,000 in Norway, ~23,000 in the UK, ~12,600 in Portugal), alongside the Ward County rent guarantee it gave as consideration for its $60 million warrant purchase. Per Nscale’s recent UK Companies House confirmation statements, NVIDIA also held about 6% of the preferred shares outstanding.

CoreWeave — Nearly $100 Billion Backlog Across 49 Data Centers (The Leased Giant)

CoreWeave is the scale case for pure tenancy. Founded in September 2017 and public since March 2025 — a $1.5 billion IPO at $40 a share, roughly a $23 billion valuation — it reported Q1 2026 revenue of $2.1 billion with more than 1 GW of active power and over 3.5 GW contracted, entirely through long-term leases across 49 data centers (~42 in the U.S.). Its stated ambition is more than 8 GW of active power by 2030, and its 2026 capex guidance is $30–35 billion.

Source: Switch Inc.

Excludes a 180 MW data center expansion in Hammond, Lake County, Indiana, because the development agreement expired on June 30, 2026. CoreWeave also exited its agreement to anchor Poolside’s planned 2 GW Horizon data center campus in Texas’s Permian Basin after Poolside failed to bring the initial GPU cluster online by CoreWeave’s deadline.

CoreWeave’s largest planned data center site is Galaxy’s Helios campus in Dickens County, Texas, which will provide 526 MW of critical IT capacity and 800 MW of gross power across three phases.

Galaxy’s Helios data center campus showing 1) Galaxy’s 345 kV Substation, 2) Chiller Yard, 3) Data Center Building, 4) Electrical Yards, 5) Water Facilities, 6) Pitchfork and Cottonwood 345 kV Substations, and 7) Helios Campus Footprint. Source: Galaxy.

The demand side is extraordinarily concentrated at the top: Microsoft alone was about 67% of CoreWeave’s FY2025 revenue, OpenAI’s commitments stack to roughly $22.4 billion, Meta’s commitment expanded to approximately $21 billion through 2032, and Jane Street signed a $6 billion capacity agreement. The backlog behind all of it: $99.4 billion as of March 31, 2026.

The financing hallmark is the delayed-draw term loan (DDTL) assembly line: $2.3 billion, $7.6 billion, $2.6 billion, and $3.0 billion facilities, then DDTL 4.0 in March 2026 — $8.5 billion for GPU purchases, backstopped by the Meta contract, and CoreWeave’s first investment-grade-rated debt at roughly 5.9%. A DDTL 5.0 followed by May.

NVIDIA sits on both sides of the book — committed to buy $6.3 billion of cloud services through 2032 (a “rent-back” that absorbs unsold capacity) and added a $2 billion equity purchase at $87.20 a share in January 2026. The symbiosis shows in sequencing: CoreWeave was first to market with H100, H200, GB200, and GB300, and in June 2026 took delivery of the world’s first Vera Rubin NVL72 rack from Dell.

Denton, Texas, showcases CoreWeave’s “leased giant” landlord dynamic in one campus. CoreWeave’s hosting relationship with Core Scientific dates to 2019 and escalated through 2024 from 16.5 MW at Austin to hundreds of megawatts across the fleet, adding a 70 MW Denton node in February 2025 (a 262 MW critical IT site in total). CoreWeave tried twice to simply buy its landlord: an unsolicited $5.75-a-share cash offer rejected on June 6, 2024, then a $9 billion all-stock agreement on July 7, 2025, at a 66% premium — $20.40 a share — which Core Scientific’s shareholders voted down on October 30, 2025, terminating the deal. It was never revived.

Source: Core Scientific.

In March 2026, the Denton, Texas site reached roughly 130 MW, carrying more than 50,000 Grace Blackwell GPUs, with about 15,000 more at Marble in Cherokee County, North Carolina — still CoreWeave, now leasing about 590 MW across five Core Scientific sites. That failed takeover explains, in one campus, why CoreWeave is now mid-pivot toward ownership, with its first self-build due online later in 2026 and a $322 million acquisition of the NEST facility in Kenilworth, New Jersey, already closed.

Subscribe now

Lambda — Founded 2012, Under 100 MW Deployed, 3 GW Promised (The GPU Veteran)

Lambda is the cohort’s elder: founded in 2012 by Stephen and Michael Balaban, it spent a decade selling GPU hardware and cloud to researchers before the AI boom found it. That heritage shapes the model — deep GPU operations expertise, a hardware business, and a footprint assembled almost entirely from colocation rather than campuses.

The financing ledger: a Series E above $1.5 billion led by TWG Global in November 2025 (valuation undisclosed; the February 2025 Series D printed a $2.5 billion valuation), a J.P. Morgan-led senior secured credit facility upsized nearly fourfold to $1 billion in May 2026, and NVIDIA on the demand side with two rent-back agreements — $1.3 billion over four years (“Project Comet,” 10,000 GPUs) plus $200 million more for another 8,000 GPUs.

The customer book gained its anchor on November 3, 2025: a multi-billion-dollar Microsoft agreement for tens of thousands of NVIDIA GPUs, including GB300 NVL72 systems. Hudson River Trading followed with a dedicated HGX B200 deployment.

The spotlight belongs to Lambda’s pivot from opportunistic colocation leasing to purpose-built, single-tenant data centers. For years Lambda took existing colocation capacity where it could find it — most recently space at Aligned’s DFW-04 in Plano, Texas (May 2025), Cologix’s COL4 in Columbus (live June 2025), and a 21 MW lease at Prime Data Centers’ LAX01 in Vernon, California (November 2025).

Source: Prime Data Centers.

In August 2025, Lambda flipped the relationship: EdgeConneX is building two dedicated, single-tenant data centers for Lambda — a 23 MW purpose-built data center in Chicago and its twin in Atlanta with more than 7 MW of capacity. The colocation-taker became an anchor tenant worth building for.

The gap for Lambda is between measured and promised: Lambda’s deployed footprint sits under 100 MW, while the company talks about nearly 3 GW by 2030 — a target it has not explained how to reach. A planned Kansas City “AI Factory,” starting at 24 MW and growing past 100 MW “in the future,” is the next increment — more than 10,000 NVIDIA Blackwell Ultra GPUs at launch, dedicated to a single unnamed customer.

Fluidstack — Five U.S. Sites, Zero Owned Buildings (The Pure Tenant-Operator)

Fluidstack is the pure case: the neocloud as tenant-operator, nothing else. Founded in 2017 at Oxford University by Gary Wu and César Maklary, it announced in December 2025 that it was moving its headquarters from London to New York, and it now manages more than 100,000 GPUs for customers including Anthropic, Mistral, Character.AI, Poolside, and Black Forest Labs.

Fluidstack’s U.S. buildout runs through five leased sites, five bankruptcy-remote tenant SPVs, and zero owned buildings. The tenant that anchors $15 billion of landlord debt is nearly invisible in the county record of the sites themselves.

Source: Cipher Digital.

The anchor is Anthropic, whose announced $50 billion U.S. data center investment plan started with Fluidstack sites in Texas and New York. Through Hut 8, the relationship extends to as much as 2.3 GW of capacity for Anthropic on Fluidstack-operated clusters. Notably, no disclosure for the five U.S. sites names an NVIDIA GPU model — the silicon documented on these floors so far is Google’s, with reporting placing TPUs at Lake Mariner and, for Anthropic, at River Bend. Google’s role — backstopping the leases, taking landlord equity, supplying the chips — is Section IV’s subject.

Source: Hut 8.

The financing trail: an $830 million Series A at a $7.5 billion valuation in January 2026, a further $1.5 billion offering underway as of June, and reported talks at an ~$18 billion valuation still ongoing — Jane Street is already on the cap table.

The spotlight site is Lake Mariner, on the shore of Lake Ontario outside Buffalo — a decommissioned coal plant that became a bitcoin mine that is becoming an AI campus, all on one zoning record. Landlord TeraWulf is building Fluidstack three buildings — CB-3 (42 MW), CB-4 (162 MW), CB-5 (162 MW) — totaling 366 MW of critical IT (about 450 MW gross) under two 10-year leases worth roughly $6.7 billion in contracted payments, up to $16 billion with extensions.

Source: TeraWulf.

Google guarantees $3.2 billion of it and took warrants that could reach about 14% of TeraWulf. Note the pattern across the five sites: five different landlord groups (Coatue’s Meridian Arc vehicle, TeraWulf, Hut 8, Cipher Digital, Aligned Data Centers), four states, one tenant. At Abernathy, even the landlord side is layered: Aligned Data Centers owns the ground, and a project JV — TeraWulf 50.1%, Fluidstack 49.9% at formation — ground-leased it and built the data center. In July 2026, Fluidstack bought out TeraWulf’s stake, leaving the pure tenant-operator owning its own landlord at one site.

The next turn was already in motion by June 2026: the Broadcom–Apollo–Blackstone XPU platform’s first gigawatt of Anthropic compute lands “in Fluidstack data centers” starting in mid-2026 — the first site has not yet been named. The platform is designed to exceed 20 GW through 2028, and its initial ~$35 billion of Apollo-led debt buys Google TPUs through an SPV that leases them to Anthropic — Broadcom backstops the senior notes — for deployment across sites in New York, Texas, Louisiana, and Indiana. Read those four states against the site table above: they are the Fluidstack map, state for state.

IV. The Backstop: Google, NVIDIA, and Microsoft Are Underwriting Their Own Suppliers

Google’s Lease Backstop: The Repeatable Template

The mechanism that finances the bottom of this market was invented, at scale, in August 2025: Google guarantees a neocloud’s lease obligations so that the landlord can issue debt against the lease. For Google it is an off-balance-sheet obligation to step in if Fluidstack cannot pay its data center rent. After the first TeraWulf deal, it became repeatable. The anatomy at the three warrant deals: Google backstops the lease, takes warrants or shares in the landlord (up to ~14% of TeraWulf; ~5.4% of Cipher), and secures a path to capacity in the TPU era; the landlord then sells senior secured notes to fund construction.

The path to capacity has a customer attached. In October 2025, Anthropic announced plans to use up to one million Google TPUs — well over a gigawatt coming online in 2026, worth tens of billions of dollars — and in April 2026 it expanded the commitment with Google and Broadcom by multiple further gigawatts from 2027. The NVIDIA-invested neoclouds have little incentive to host rival silicon, which leaves the converted miners and Fluidstack as the floors where those TPUs physically land. Google — itself a roughly 14% shareholder in Anthropic — is guaranteeing the real estate that houses its own chip business.

The math is the revealing part. At Lake Mariner, Barber Lake, and Abernathy, Google’s guarantee equals the landlord’s note principal to the dollar — $3.2 billion, $1.73 billion, $1.3 billion — which is only 0.14×–0.48× the rent those leases nominally owe. Google is not insuring the lease; it is collateralizing the bond. The structure has teeth for six years — during that window Google must assume the lease outright if Fluidstack defaults — and only afterward can it exit by paying a termination fee pegged to the remaining debt.

New Lebanon, financed without landlord warrants, inverted the pattern: Meridian Arc’s $5.7 billion of notes (the largest single raise in the set) carry a Google guaranty that tracks the rent, peaking around $7.8 billion — 1.37× the debt, consistent with the lease’s $18 million-per-MW rent cap. River Bend runs further still: an all-rent guarantee with no disclosed cap, plus a springing Alphabet guarantee if Google’s share of Alphabet’s consolidated revenue ever falls below 51%.

NVIDIA’s Rent-Back and Guarantee Book

NVIDIA’s version of the same instinct — Jensen Huang has said NVIDIA “cultivated” the neoclouds — comes in three instruments.

1. Rent-Back: commitments to buy capacity its customers can’t sell — $6.3 billion from CoreWeave through 2032 (a contract CoreWeave says is for GB300s) and $1.5 billion from Lambda across two agreements.

2. Guarantee: Nscale’s Ward County, Texas structure from Section II, where the price NVIDIA paid for its Nscale warrants was $60 million in cash plus an $860.3 million five-year rent guarantee for Ionic’s benefit — which covers roughly a third of the lease’s ~$2.6 billion contracted revenue, front-loaded onto exactly the years before the site has operating history.

3. Equity Participation: direct investments in CoreWeave ($2 billion of common stock), Nebius (~$2 billion of common stock), and Nscale (~$777 million for ~6% of preferred shares), plus participation in Crusoe’s Series D and Series E rounds and Lambda’s Series D, although NVIDIA’s individual contribution to those financings was not disclosed.

The Landlord Debt Stacked on Top

Against Fluidstack’s five leases alone, five separate issuers have sold over $15 billion of senior secured notes:

Notably, Hut 8 DC is the only investment-grade deal in the set (Fitch BBB−) because the notes fully amortize inside the lease term while the others leave 60–85% balloons.

Microsoft: The Anchor That Needs No Guarantee

Microsoft anchors the cohort from the customer side at a scale that needs no guarantee: up to 1.35 GW intended at Monarch in West Virginia, up to $19.4 billion at Vineland in New Jersey, 900 MW at Abilene in Texas, roughly two-thirds of CoreWeave’s revenue, and a multi-billion-dollar Lambda deal. Denominated in silicon, that book already runs to roughly 200,000 GB300s contracted from Nscale worldwide — ~104,000 of them in Texas — and a reported 100,000-plus from Nebius. Five of the six companies have Microsoft as an anchor customer; the sixth has Google guaranteeing its rent.

What the Middle Layer Actually Pays

The leases with clean disclosure let us derive it (contracted rent ÷ term ÷ critical MW):

The $65 and $100 are year-one contract rates; $/MW-yr figures — and the parenthetical ~$/kW-mo equivalents — are contracted rent ÷ term ÷ critical MW: term averages, escalators included. Ionic’s $65 is on gross power (≈$90/kW-mo critical IT at 1.4 PUE); Core Scientific’s $100 is year one, so its term average runs higher.

The spread down the column is not a bargain ladder — it is scope.

At Ward County the tenant takes a powered shell and pays for its own data center fit-out (triple-net, on ~$0.05/kWh West Texas power).

Core Scientific sits in the middle: it builds the data center, but CoreWeave funds the construction and recovers it through a credit capped at $1.5 million per MW, applied as a 50% offset against each monthly license fee until exhausted — $859 million of credits across the five sites.

At the Fluidstack sites the landlord builds and funds the finished data center, which is exactly why those landlords needed $15 billion of guaranteed debt.

Same product name — “rent” — three different products.

Subscribe now

V. What Breaks First: Concentration, Duration, and the Landlord’s Second Tenant

Concentration Runs Both Ways

Start with concentration. Anthropic sits across the entire Fluidstack chain — the $50 billion plan that starts there, the 2.3 GW via Hut 8, the first XPU gigawatt.

Microsoft’s version of concentration runs the other way: it anchors five of the six, which means the cohort’s largest single risk factor is one customer’s capital-allocation committee.

The Landlord’s Second Tenant

Anthropic, meanwhile, has already shown it can skip the middle layer: its Kentucky campus deal — roughly $19 billion over 20 years — is signed directly with TeraWulf, the same landlord Fluidstack fronts at Lake Mariner, with no neocloud and no Google backstop in between.

Source: TeraWulf.

The squeeze runs from below, too. IREN, one of the converted miners in Layer 1, monetizes its power differently — a vertically integrated AI cloud whose customer list, per a July 2026 contracts announcement, now includes Fluidstack itself. When the landlord class sells compute directly, the neocloud shows up as its customer, not its tenant.

Ten-Year Leases on Two-to-Five-Year Silicon

The duration mismatch is capital structure, not silicon commentary. GPU contracts in this market run two to three years, five at the upper end, while data center developers underwrite seven-to-nine-year paybacks — and the leases the neoclouds sign run 10 to 15 years. A 10.5-year lease like Ward County’s spans roughly three of those GPU contract cycles: the tenant must refill the same building with newly financed silicon three times or more before its rent obligation ends, in whatever price environment prevails. The churn is already observable inside the buildings: Core Scientific describes its CoreWeave halls stepping from H100 through GB200 to GB300 platforms in barely two years.

Subscribe now

Measured AI provides institutional-grade analysis of the physical infrastructure powering AI data centers. For access to our full research library, regulatory intelligence, and weekly briefings, visit MeasuredAI.com.

The machine behind those claims sits on former farmland in Mount Pleasant, Racine County, about 30 miles south of Milwaukee: 315 acres, three buildings, 1.2 million square feet under roof, wired to run as a single cluster of “hundreds of thousands” of NVIDIA GB200 GPUs — part of a campus Nadella says will “scale to 2 GW alone,” on land where Microsoft now holds more than 2,000 acres. A second campus is due for completion in 2028, and 15 more data center buildings — a third and fourth campus — won approval in January 2026.

Microsoft’s stated investment has climbed from $1 billion at the 2023 groundbreaking to more than $7 billion as of September 2025 — and, with a further $13.3 billion expansion in 2026, to a program its utility pegs at more than $20 billion. The campus is set to become the biggest electric load served by We Energies, the regulated utility that powers Wisconsin’s southeastern corner.

What makes Fairwater Wisconsin worth a close read is a split that runs through every layer of it. This is the flagship Microsoft actually owns — the land, the buildings, the machine — its first Fairwater and the template for the family, in contrast to the leased QTS shells of Microsoft’s Fairwater Atlanta and the rented gigawatt of Microsoft’s Monarch development in West Virginia. But everything that feeds and hosts the machine is inherited, rented, or borrowed: a land bank, tax district, and Great Lakes water diversion built for Foxconn’s abandoned LCD plant; a regulated utility’s rate base, now being rebuilt around one customer; another city’s lake water; and the patience of a village that has already hosted one broken mega-promise.

This AI data center primer walks the four Microsoft campuses in Wisconsin layer by layer — the wager, the land, the machine, the power, the water, the neighborhood, and the fiscal bargain — and closes with the markers that will test it.

Subscribe now

I. The Wager: The One Fairwater Microsoft Owns

Wisconsin is the Fairwater Microsoft self-built and owns end to end — the company calls it “our first Fairwater site,” the “largest and most sophisticated AI factory we’ve built yet,” with “multiple identical Fairwater datacenters under construction in other locations across the US.” Wisconsin is the template; the question the campus answers is whether Microsoft can still execute as an owner.

Source: Microsoft.

That question was live because of who the machine serves and what Microsoft lost while building it. The GPU building is fully dedicated to OpenAI, with OpenAI placing its training workloads inside Microsoft’s Fairwater campus. The relationship behind that dedication is enormous — Microsoft holds an approximately 27% investment in OpenAI and $13 billion in funding commitments, $11.8 billion of it funded by March 31, 2026 — but it is no longer exclusive.

Wisconsin was the intended host of the $100 billion “Stargate” supercomputer Microsoft and OpenAI discussed in early 2024, on a roadmap that would have taken the site past 2 GW. The first Stargate contract went instead to Oracle in Abilene, Texas, and by January 2025 the partnership had loosened to a right of first refusal. The loosening was paid for in demand: when the partnership was restructured again in October 2025, OpenAI committed to an incremental $250 billion of Azure purchases. A further amendment on April 27, 2026, loosened it again — Azure remains OpenAI’s primary cloud, and new OpenAI products still ship there first, but Microsoft’s technology license runs non-exclusive through 2032 and the companies’ revenue-sharing now winds down under a cap.

OpenAI’s own Stargate program now includes a separate Wisconsin campus — the Oracle-and-Vantage build in Port Washington, about 50 miles north — that has nothing to do with this site. The label, though, has since bent back toward Mount Pleasant: OpenAI’s January 2026 “Stargate Community” pledge lists “our latest AI campus with Microsoft in Mount Pleasant, Wisconsin” among “all of our Stargate AI campuses” — the site Stargate passed over, now carried under its banner in OpenAI’s own telling.

Microsoft’s second act is the reversal. The stated investment had climbed from $1 billion at the September 2023 groundbreaking to the $3.3 billion President Biden traveled to Mount Pleasant to announce in May 2024. In September 2025, Microsoft committed an additional $4 billion for a second campus of similar size — taking the first two campuses to $7.3 billion — and branded the site “the world’s most powerful AI datacenter.” The third and fourth campuses approved in January 2026 add another $13.3 billion, putting the full program at more than $20 billion.

Then came the schedule beat: online in April 2026, “6 weeks ahead of schedule,” fully operational June 23. Wisconsin became the proof that Microsoft’s self-build arm still works.

The wager needed ground first — and the ground was already there, cleared and entitled by someone else’s failure.

II. The Land: Foxconn’s Ground, Microsoft’s Campus

Fairwater exists where it does because Wisconsin had already assembled, zoned, and plumbed thousands of acres for a factory that never came. In 2017, the Village of Mount Pleasant created Tax Incremental District No. 5 (TID 5) — roughly 3,900 acres with $763.8 million of planned public spending — for Foxconn’s promised $10 billion, 20-million-square-foot LCD plant and up to 13,000 jobs at an average salary of $53,875. The plant “never fully materialized”; Foxconn opened a spherical data center on the site in 2021 and shelved the LCD plans.

Source: Foxconn.

The village was left holding shovel-ready industrial land, special-purpose water mains, and Foxconn-era debt. Microsoft bought into all three.

The first purchase set the pattern. On March 30, 2023, the Village Board approved the sale of about 315 acres at 90th Street and Braun Road to Microsoft for $50 million — $159,000 per acre — against the $1 billion development-spending floor written into the development agreement approved the same day. Racine County co-signed the development agreement that April on an 18–1 vote. Microsoft broke ground at 90th Street that September.

As of mid-2026, Microsoft’s plan in Mount Pleasant, once fully constructed, comes to four campuses, 19 data center buildings, 11.2 million square feet, and a stated path to 2 GW:

Building counts are data center buildings only — Campus 1’s third building, the Central Utility Plant housing its chiller hall, is not counted.

The land bank grew in two waves, and the sequence shows what ran out: first the Village’s ex-Foxconn inventory, then the farms around it.

In November 2023, the Village sold Microsoft two more tracts — 231 acres at Campus 2 for $36.5 million and 400 acres at Campus 3 for $63.2 million, both at $158,000 per acre — while Microsoft separately bought the 400-acre Creuziger farm (now part of Campus 3), leaving it with 1,346 acres inside TID 5, more than two square miles. Then came the private assemblage: 205 acres for $33.9 million in July 2024, the Hribar trust’s 70.9 acres for $12.75 million, roughly 160 more acres for over $43 million that September, and two Braun Road tracts plus an Espy Farms parcel in late 2025, taking Microsoft past 2,000 acres of land.

The expansion that defines the campus’s next decade cleared in one evening. On January 26, 2026, the Village Board unanimously approved site plans for 15 more data center buildings across Campuses 3 and 4:

• Campus 3’s MKE07–15: Nine one-story data center buildings of roughly 579,000 square feet each on the 791-acre Durand Avenue parcel. Two electrical substations will occupy 36 acres each on the north-central and south-central portions of the campus.

Source: Wisconsin Department of Natural Resources.

• Campus 4’s MKE54–59: Six one-story data center buildings of about 568,000 square feet each just north at 2700 International Drive.

Source: Wisconsin Department of Natural Resources.

Building permits show Microsoft’s ramp on Campus 3 is already underway: the foundation permit for the first Campus 3 building was issued May 14, 2026, three more followed within two weeks, and the first vertical-construction application landed in June 2026.

All of this proceeded while Campus 1 received its conditional certificate of occupancy on May 26, 2026. Meanwhile, Campus 2 (shown below) is progressing toward its planned 1.3-million-square-foot, 337.6 MW completion in 2028.

Source: Walsh Construction.

What already stands operational on the first 319 acres is what makes the remaining land significant.

Subscribe now

III. The Machine: Hundreds of Thousands of GPUs Run as One Computer

Microsoft’s claim for Fairwater is not that it is a big campus of servers; it is that the campus is one computer — “the most powerful, tightly coupled AI supercomputer in the world, purpose-built for frontier models,” in the words of cloud chief Scott Guthrie, promising “10X the performance of the world’s fastest supercomputer today.”

Source: Walsh Construction.

Microsoft has disclosed its architecture for Fairwater, and it follows a single design principle: shorten every cable, flatten every network, and let nothing fragment the cluster. It is also the master copy — this is the template the rest of the Fairwater family inherits — so the choices below are ones Microsoft made for itself, as owner and architect, not specifications negotiated with a landlord.

The unit of the machine is the one layer Microsoft buys rather than invents: the rack. Each one packs 72 NVIDIA Blackwell GPUs into a single NVLink domain with 1.8 terabytes per second of GPU-to-GPU bandwidth and 14 terabytes of pooled memory, processing — by Microsoft’s count — 865,000 tokens per second, which it calls the highest throughput of any cloud platform.

Source: Microsoft.

Azure was the first cloud to bring up the GB200 rack systems, and Wisconsin is Microsoft’s flagship deployment; there are more than 150,000 GB200s in Microsoft’s Campus 1 GPU building. Wisconsin’s silicon story is GB200 — the GB300 generation belongs to the Atlanta sister site and the broader fleet.

The GB200 racks stack into a building engineered around them — the first rung of the ladder Microsoft designs, builds, and owns outright, shaped by the physics of cable runs. The GPU building is two stories and roughly 750,000 square feet, sized at 294.4 MW of critical IT load, because stacking racks vertically shortens the copper and fiber paths that set training latency.

Source: Microsoft.

Nadella told investors the “two-story design and liquid cooling allow us to run higher GPU densities” and improve both performance and latency for high-scale training. The two-story signature stops with the flagship generation, though: on the approved site plans, all 15 expansion buildings across Campuses 3 and 4 are drawn one story.

Beside the GPU building sits a dedicated roughly 300,000-square-foot, 43.2 MW CPU-and-storage building — a roughly 1:7 CPU-to-GPU power ratio that exists because reinforcement-learning loops need thousands of CPUs adjacent to the cluster for code compilation, verification, and data processing, with storage systems that run “five football fields in length” — roughly 1,500 feet.

Source: Microsoft.

The scale of the Campus 1 self-build, executed by general contractor Walsh Construction, is easiest to read in what construction consumed: 46.6 miles of deep foundation piles, 26.5 million pounds of structural steel, 120 miles of medium-voltage underground cable, and 72.6 miles of mechanical piping.

Source: Microsoft.

Above the rack, the innovation is Microsoft’s. The GPU building is run as one flat fabric — both InfiniBand and Ethernet at 800 Gbps in a full fat-tree, non-blocking architecture, so that every GPU can communicate with every other GPU without oversubscription. The failure-tolerance layer is a Microsoft-and-OpenAI co-design developed with AMD, Broadcom, Intel, and NVIDIA: the Multipath Reliable Connection (MRC) protocol. MRC’s multi-plane design splits each 800 Gbps network interface across eight parallel planes, letting just two tiers of switches fully connect 131,072 GPUs, whereas a conventional network would need three or four tiers.

Source: OpenAI.

OpenAI says MRC is already deployed across its largest GB200 supercomputers, Microsoft Fairwater among them — OpenAI’s own confirmation of whose models Wisconsin’s machine trains.

The top of the ladder leaves the state. Fairwater Wisconsin is the anchor node of Microsoft’s AI WAN — a dedicated long-haul fabric running at 300 terabits per second with headroom toward 10 petabits, across a global fiber backbone Microsoft says spans more than 500,000 miles. When Fairwater Atlanta came online in late 2025, Microsoft joined the two sites into what it calls “the world’s first AI superfactory.”

As Azure infrastructure GM Alistair Speirs defines it, “it’s running one complex job across millions of pieces of hardware … it’s not just a single site training an AI model, it’s a network of sites supporting that one job.” Wisconsin’s stated ceiling within that system is Nadella’s: the Wisconsin site will “scale to 2 GW alone.”

A 2-GW machine, though, is only as real as its electricity — and power is the layer Microsoft does not own here.

IV. The Power: The Utility Rebuilds Itself Around One Customer

At Fairwater Wisconsin, the grid layer is a significant part of the story. The campuses are entirely grid-powered and their supplier is a regulated monopoly, We Energies — with the campuses set to become the biggest electric load the utility serves.

Source: Microsoft.

What Microsoft’s bet has set in motion is a rebuild of southeastern Wisconsin’s energy system: roughly $2.77 billion of Public Service Commission–approved generation, gas, and transmission projects justified on this load, a new tariff class invented for it, and a utility holding company whose five-year capital plan has doubled since the campus was announced.

The Customer: Named Late, Sized in Fragments

Microsoft’s name first surfaced in the Public Service Commission of Wisconsin regulatory record only in January 2025 rebuttal testimony, under intervenor pressure: Microsoft, Mount Pleasant — and the one hard load figure, 450 MW for Phase 1. The Commission’s final decisions then ground the need for new natural-gas-fired power plants in one customer, Microsoft. Docket testimony projects that just two data center customers — Microsoft (Mount Pleasant) plus Vantage (Port Washington) — will consume 50% of the utility’s peak demand and energy by approximately 2030.

Microsoft’s own filings size the first two campuses’ substations — named Nimbus and Stratus — at no less than 752 MW apiece, and Campus 3’s two planned substations — named Jupiter and Cosmos — at no less than 1.2 GW per substation, against a stated Campus 3 target of 2.6 GW of grid power to the data center buildings. For Campus 4, an electrical substation will occupy a 54.4-acre pad in the southeast corner of the site — the area for which We Energies’ five-year plan already carries about 500 MW.

To its investors, We Energies’ parent company, WEC Energy Group, discloses Microsoft’s announced $20 billion investment but folds its megawatts into aggregates. The megawatts appear only in a five-year demand-growth forecast now at 3.9 GW, of which 2.6 GW sits in the southeastern corridor that contains Fairwater alongside other industrial loads, with Vantage’s Port Washington campus anchoring the northern balance.

The Buildout: $2.77 Billion, Approved on One Customer’s Load

Between May and December 2025, the Commission approved five projects that read as one supply chain, traced backward from the campus’s meter. The wires come first: the ATC Western Feed builds the two substations — Jupiter and Cosmos — that physically interconnect Campus 3, on a schedule American Transmission Company (ATC) said was “driven by the customer’s desire to ensure the new data center is fully operational as soon as possible.”

The wires need generation behind them: the Paris Reciprocating Internal Combustion Engine (RICE) plant is the fast block, slated to land weeks after the first campus reached full operation — a schedule that has since slipped — and the Oak Creek turbines are the bulk block, timed to the second campus. The generation needs fuel: the Rochester Lateral moves the gas to the new plants, and the Oak Creek LNG plant stores two billion cubic feet of it for the days when demand peaks.

Little of it is at the campus: only the Western Feed’s substations sit on Microsoft’s land — the new plants and their fuel rise miles away, at existing utility sites on the Oak Creek lakefront in Milwaukee County and beside the Paris Generating Station in Kenosha County. None of these assets is dedicated to Microsoft — each is a system resource in a utility’s rate base — but every approval rests, directly or one step removed, on the load forecast the Microsoft campus anchors:

The wires are where the chain touches the campus. The first two Microsoft data center campuses draw on an inherited feed: ATC’s $117 million Mount Pleasant Tech Interconnection, approved in 2018 to power Foxconn, whose Mount Pleasant Substation serves them at 138 kV today. That substation could not carry the expansion — ATC found the new load “unprecedented for the applicant’s transmission system.”

American Transmission Company’s Racine County Western Feed — the two new 345/138 kV substations, four 500 MVA transformers apiece, a 350 MVAR STATCOM, and about 9.6 miles of new double-circuit lines through the Villages of Yorkville and Mount Pleasant — was authorized at $625 million “to support new datacenter load addition.”

Source: American Transmission Company (ATC).

MISO classifies the Western Feed as “Other–Load Growth” with no regional cost sharing: Wisconsin transmission customers bear all of it. ATC itself is roughly 60% owned by WEC Energy Group — the wires builder and the utility roll up to the same shareholders.

The Western Feed’s right-of-way politics were sharper: farmland owners objected to easements taken, as one put it, solely for the benefit of a company with “a net worth of over 3 trillion,” and the state agriculture department noted the single largest impacted landowner on the route was Microsoft itself, at 212.9 acres. Construction began in October 2025; the Jupiter Substation is due to take load in January 2027, with the Cosmos Substation following in June 2028 — a date ATC says is “subject to further coordination with Microsoft.”

Behind the wires, the generation is gas. The Paris RICE plant — seven Wärtsilä 18V50SG reciprocating engines of 18.8 MW gross each, about 128 MW net — was approved at a site beside the existing Paris Generating Station in the Town of Paris, Kenosha County, with a July 2026 in-service date. The schedule has not held: all seven engines were reassembled on site by late February 2026, but We Energies now shows construction continuing through summer 2027. It is being built as a single long engine hall on roughly 11 acres, with a row of seven 100-foot exhaust stacks along the roof, one per engine, and a short 138 kV tie into the substation it shares with the co-located Paris solar-battery park.

Source: We Energies.

The Oak Creek Combustion Turbines (CTs) are the big block: five simple-cycle GE 7FA.05 units, roughly 1.1 GW, rising on the Lake Michigan shore at Oak Creek, 15 miles south of Milwaukee, on the site where We Energies retired about 1.085 GW of coal capacity across 2024–2025.

Source: We Energies.

The peakers “power up in minutes,” shed heat through closed-loop air-cooled exchangers, and are stitched into the 1,000-acre lakefront site by roughly 36 miles of underground electrical conduit, more than 55,000 feet of buried process piping, and an expanded ATC 345 kV substation. The utility’s clinching argument was a floor, not a forecast: even if only 450 MW of the anticipated new demand materializes — “approximately 20% of the forecasted new demand” — the turbines remain part of the least-cost portfolio. Construction began in July 2025, targeting August 2028.

Feeding the new fleet takes its own infrastructure, justified almost entirely by generation demand — electric generation accounts for 94% of the gas utility’s projected demand growth. The 33-mile Rochester Lateral — a 24- and 30-inch steel transmission main rated at 650 psig, running from the ANR interstate pipeline system’s Rochester Gate, in the Village of Rochester, across Kenosha, Racine, and Milwaukee counties to the Oak Creek plant site — is the delivery leg, with 90% of its cost allocated to generation.

Source: Public Service Commission of Wisconsin.

The Oak Creek LNG plant is the peak-day buffer behind the pipeline: a 200-foot-diameter full-containment tank on a former coal pile at the same Oak Creek generating site, holding 2 billion cubic feet of gas at −260°F — filled by on-site liquefaction over roughly 200 days, then drainable in about seven at its full 300,000-dekatherm-per-day send-out.

Source: We Energies.

However finely the dockets sliced the chain, We Energies’ own witness stated where it ends: the LNG plant’s costs would be charged to the new gas plants “and then passed through to [the data centers] via the Very Large Customer and Bespoke Resource tariffs.”

The wave now pending drops the system-resource framing altogether. In November 2025, We Energies filed to acquire Red Oak Ridge — a 1.125 GW, $1.77 billion simple-cycle peaker: five GE 7F.05 turbines on roughly 200 acres in Kenosha County’s Town of Paris, the same town that hosts the new RICE block.

Source: Invenergy.

The structure inverts the self-builds above — Invenergy develops and builds, and We Energies takes ownership at completion, putting the plant into its rate base. Microsoft is subscribed to 100% of Red Oak Ridge as a dedicated “Bespoke Resource” — its costs borne by no other customer — and We Energies’ need case is one line: “additional dispatchable generation would not be required but for” the new tariff-class load.

The Tariff: Pricing “Paying Our Own Way”

Until 2026, Microsoft took service under a tariff written for its predecessor — the Foxconn-era EITM rate, capped at 200 MW. The replacement is the Very Large Customer tariff We Energies filed on March 31, 2025, expressly to “shield other ratepayers” from data center costs. Microsoft’s energy VP (at the time) Bobby Hollis endorsed it as ensuring “we are protecting other ratepayers, paying our own way.”

As ordered in May 2026, the Very Large Customer tariff applies mandatorily to loads of 100 MW or more, requires 15-year minimum contracts, charges a fixed $213,118 per billing period plus $305 per MW of maximum demand, sets minimum transmission billing at 100% of forecast load, and demands collateral unless the customer clears a raised credit-waiver bar (A−/A3 ratings plus balance-sheet tests).

The utility’s case for the tariff quantified the stakes: without it, roughly $1.5 billion of data center–driven plant would shift onto other customers, raising residential bills about 5%. The contest was over whether the shield is complete — Commission staff’s own analysis found that a large customer whose transmission needs run beyond forecast could still push $441 million onto other transmission customers over 45 years, and the Citizens Utility Board of Wisconsin argued the standard should be “no harm,” not less harm.

The public read the docket the same way. The tariff drew 2,130 comments, 88.5% opposed — but the coded record shows nearly nine in ten of those opponents were attacking the tariff’s design, demanding data centers pay 100% of the new costs, not opposing data centers as such; only about one in ten opposed the build-out itself.

The counterweight came from labor: union bulk petitions carrying roughly 2,400 signatures backed the companion gas projects. The first stress test arrived within weeks of the order — Microsoft filed a request for clarification, and Oracle (whose roughly 1 GW Port Washington campus tripped the new credit bar) petitioned to reopen; both were pending as of mid-2026, and actual cost recovery lands in the utility’s 2027–2028 rate case.

The Transformation: A Capital Plan Doubled

For WEC Energy Group’s shareholders, the same story reads as transformation. The parent company’s five-year capital plan — a system-wide figure spanning utilities in four states, not just southeastern Wisconsin — stood at $18.2 to $20.1 billion in the 2023 vintage. It now runs to $37.5 billion for 2026–2030.

Source: WEC Energy Group.

WEC now expects electric demand to rise “approximately 45%,” with very-large-customer and dedicated (”bespoke”) assets reaching about 15% of its asset base by 2030.

That fleet has now moved further along: recitals in a July 2026 We Energies filing disclose eight signed Microsoft payment-and-cancellation agreements to date — five simply “necessary to serve” the load, one covering capacity for Microsoft’s first roughly 1 GW of firm load, one for the 1.125 GW Red Oak Ridge gas peaker in Kenosha County, and one for the 150 MW Superior Solar farm — each obliging Microsoft to reimburse the utility’s sunk costs if regulators withhold approval or the load shrinks.

The Backup: Engineered Under the Line

The campus’s own generation is strictly for emergencies — and engineered, permit by permit, to stay under the federal major-source line. In four construction permits issued from November 2023 to June 2026, the DNR has authorized 229 emergency generators and 10 fire-pump engines across eight permitted data center buildings.

The fleet’s theoretical NOx potential runs to 2,779 tons per year; enforceable limits pin its permitted potential at 99.21 tons — just under the 100-ton threshold that would make it a Title V major source, the same under-the-line engineering AWS used at New Carlisle against the PSD threshold.

The emergency generator fleet, which comprises mainly Cummins C3000 D6e models each permitted at up to 3.2 MW, has a standby capacity of about 723 MW based on its permitted unit ratings.

In its 2026 permit, Microsoft made a concession with teeth: it elected an outright prohibition on non-emergency peak-shaving and demand-response operation, locking the diesels to genuine emergencies, and it has committed to running them on renewable biofuel, specifically hydrotreated vegetable oil.

The Solar Match: A Contract, Not a Campus Asset

Microsoft’s renewable answer is a contract, not a campus asset. In October 2024 it signed a power purchase agreement with National Grid Renewables (now known as Geronimo Power) for the 250 MW Portage Solar project — roughly 1,800 acres of farmland in Portage County, central Wisconsin, some 150 miles from the Microsoft campuses.

Source: Geronimo Power.

Construction started in April 2025 and commercial operation is projected for March 2027. As of early 2026, construction had not yet begun on a planned 137.5 MW / 550 MWh battery, which would store midday generation and discharge it into the evening and other high-demand hours.

The PPA is a grid-match: Microsoft’s pledge to offset every fossil kilowatt-hour one-for-one with carbon-free energy added to the grid, plus a $20 million community fund over the contract term. Critics did the math early — Clean Wisconsin and 11 other groups wrote in September 2024 that 250 MW “represents just a fraction of the data center’s energy needs,” warning the campus would anchor “30 more years of fossil fuels”; We Energies’ spokesman answered that gas is “the cheapest, most reliable and lowest carbon approach” when wind and solar fall short. The generation table above is the Commission’s verdict on whose math prevailed.

The solar match now has a sequel, structured the opposite way. We Energies and Madison Gas & Electric are buying Invenergy’s 150 MW Superior Solar project in Michigan’s Upper Peninsula — a roughly $393 million asset the utility will own, with Microsoft subscribed to “80% of the total nameplate capacity of the project or roughly 120 MW“ under the new bespoke machinery, targeting service in Q4 2028. The executed agreement, filed July 1, 2026, named Microsoft — bringing the nameplate solar matched to its Wisconsin load to roughly 370 MW, still a fraction of a campus pointed at 2 GW.

Electricity is the rented layer with the biggest price tag. The other fluid running through the campus has an older, stranger provenance.

Subscribe now

V. The Water: Lake Michigan, on Foxconn’s Terms

Cooling is where Fairwater’s engineering does exactly what Microsoft promises; water is where the campus is most visibly a tenant of someone else’s deal. The buildings themselves barely consume water: more than 90% of the facility runs closed-loop liquid cooling, filled once during construction and recirculated indefinitely — “zero water waste” — with the remaining traditional servers on outside air, drawing water only on the hottest days, less than 5% of the year above 85°F.

Source: Microsoft.

What the loop rejects, hardware handles at scale: Microsoft calls the campus’s chiller plant “the second largest water-cooled chiller plant on the planet,” shedding heat through 172 twenty-foot fans.

Source: Microsoft.

Microsoft’s public quantification is an analogy — annual water use “roughly the amount … a typical restaurant uses annually.” The Village of Mount Pleasant record supplies the hard numbers at both scales: about 2.8 million gallons annually for the first phase — the facility the analogy described — and about 8.4 million gallons drawn from the Racine Water Utility for the full build-out, the “all-encompassing number for all of these sites” Microsoft’s engineers gave the Plan Commission.

The supply chain behind those gallons is the borrowed part. The campus sits in Mount Pleasant, but Mount Pleasant treats no water; Microsoft “will purchase water from the City of Racine,” a different government, under an intergovernmental Retail Water Service Agreement dated May 24, 2004, with sewage returning to Racine’s wastewater treatment plant under a 2002 sewer agreement.

Source: Applied Technologies.

And the marquee water instrument in the area — the Great Lakes diversion — is a Foxconn-era artifact end to end. Racine obtained it from the DNR in April 2018, for up to 7 million gallons per day averaged over a calendar year, to serve the part of Mount Pleasant that lies outside the Great Lakes basin; the application leaned on the state’s Foxconn-specific manufacturing-zone statute, and Foxconn’s own cooling data supported the consumptive-use analysis.

Seven years on, the diversion’s sole customer is still Foxconn, whose annual draw — 13.8 million gallons in 2022, rising to 25.6 million in 2025 — reached only about 1% of the permitted volume even at its peak. Racine’s own diversion reports state plainly that Microsoft’s first phase “will not be located in the diversion area”; only future phases sit on diversion-area land. The most contested water approval in the region’s history, in other words, is not yet Fairwater’s water supply — it is the option Microsoft inherited with the ground.

What Microsoft’s arrival did change is the deal’s politics. In June 2024, Racine’s city attorney served the Village with a Notice of Potential Material Change of Circumstances under the 2004 water agreement — and a parallel notice under the 2002 sewer agreement — arguing that the Village had swapped tenants on the district Racine built its waterworks case around: Foxconn’s promised plant and payroll, replaced by a Microsoft project the notice pegs at $3 billion to $10 billion in value but “approximately 200 jobs associated with each of the 5 phases.” The city calls it the “Job Disparity.”

The fights over water stayed procedural. The fight over sound did not.

VI. The Neighborhood: What the Flagship Sounds Like Next Door

Every friction around Fairwater traces to the same fact: the flagship landed beside subdivisions, in a county that had already absorbed one broken mega-project — and the second time around, the neighbors’ tolerance is thinner. The hum made that concrete.

Residents north of Campus 1 (4800 90th St) began noticing a constant tonal sound in early February 2026, as the first data centers started up. Microsoft’s April 15 statement acknowledged it directly: the sound “comes from cooling fans that are currently operating at high speeds”; the facility “meet[s] the requirements set by the local ordinance”; the company “did not expect the tonal quality of the sound to travel as far as it has.” The village president of neighboring St

urtevant, Mike Rosenbaum, put the local mood in one line at a Microsoft Q&A: “I don’t need to be driving people out of my village because of some fan noise.”

On July 1, 2026, the hum became a federal case. Ostergaard v. Microsoft (E.D. Wis., No. 2:26-cv-01169) is a class-action nuisance and negligence suit by three Sturtevant residents, on behalf of everyone in a defined residential area north of Campus 1 — “over a thousand residential households” — alleging a 24/7 low-frequency hum from generators and cooling equipment, much of it infrasound that standard decibel meters miss. The suit seeks damages above the $5 million class-action threshold and injunctive relief.

The permitting record shows the same curve of rising resistance. Microsoft’s fourth air-permit revision drew a public hearing on April 28, 2026 — requested jointly by the Villages of Mount Pleasant and Sturtevant — with 69 attendees and 21 speakers, all in opposition. Ten miles north, community feedback had already put a stop to a sibling project: Microsoft withdrew its 244-acre Caledonia “Project Nova” campus in October 2025, with its land-use counsel conceding “nobody really wants a data center in their backyard.”

Mount Pleasant itself, though, is the exception, and the reason is timing: data centers were made a permitted use by ordinance in early 2023 through the Foxconn-era planning apparatus — five public hearings across the re-zonings, all passed unanimously — which is why January’s expansion could clear without floor debate, even as the air-permit hearing three months later drew only opponents.

These frictions are one side of the ledger the region keeps. The other side is the bargain.

Subscribe now

VII. The Bargain: What Wisconsin Traded This Time

The Microsoft deal is Foxconn’s inverted. Foxconn came with a state package of roughly $3 billion in refundable credits (later renegotiated down) and promises that never survived contact with reality.

Microsoft came with no state mega-package: it bought its land outright, guaranteed a tax base, and pays its way into a district built — and indebted — for its predecessor. The structure runs through three governments: the Village of Mount Pleasant owns and sells the land, Racine County co-signs the development agreements, and the City of Racine supplies the water.

The enforceable core is a guarantee stack. Microsoft commits to at least $1.4 billion of assessed value on the expansion areas by January 1, 2028, with shortfalls made up in cash at up to $15 million a year through 2047, and to at least four buildings on the expansion land by milestone dates in 2030, 2037, and 2044, backed by Village of Mount Pleasant rights to repurchase undeveloped land at Microsoft’s acquisition price. The clawbacks fall away only once $3 billion of aggregate guaranteed value is on the rolls and vertical construction has begun on Campus 3.

Against those commitments, the Village rebates 42% of the incremental property taxes from buildings on the first campus area, capped at $5 million a year. The one large subsidy is the state’s qualified data center sales-and-use-tax exemption, for which Microsoft was certified on October 1, 2023 — a statewide statute rather than a negotiated award. Beyond that, the state’s direct spending is modest: two $500,000 WEDC innovation grants and $2 million for technical-college AI programs.

The money already moves the other way. The Village’s three land sales brought in $149.8 million; the first sale’s proceeds went to Foxconn at closing under the old development agreement, and Microsoft separately funded an $18.4 million defeasance that retired $19 million of the county’s Series 2019B “Foxconn” land bonds — eliminating $24.4 million in debt service. The district Foxconn was supposed to fill is now filling: TID 5’s incremental value rose from $906 million to $1.7 billion in a year, its projected lifetime surplus from $80.6 million to $136.7 million, and the joint review board projects $2.1 billion of value by January 1, 2028.

On Racine County’s tax roll, Microsoft went from unlisted in 2023 to the county’s largest taxpayer by 2025, at $1.22 billion of taxable assessed value — 4.5% of the county’s total. The January 1, 2026 roll more than doubles it: $2.58 billion across Microsoft’s 30 parcels, $1.87 billion of it on the first campus alone — the buildings themselves now carry nearly four times the value of the land beneath them.

Jobs are the bargain’s soft spot. Construction is the boom: more than 3,000 workers at daily peak, nearly 10,000 over two years, $700 million spent with Racine County construction suppliers and purchases from 29 businesses across 11 Wisconsin counties — part of the $4.7 billion Microsoft estimates it will spend on hyperscale construction in the state between 2024 and 2028. Permanent operations are the modest tail: about 500 full-time roles at the first data center campus — “nearly 550” at go-live — growing to around 800 full-time roles once the second campus is complete, with a Datacenter Academy at Gateway Technical College in Racine training more than 1,000 students over five years.

The bargain Wisconsin struck this time trades away almost nothing in cash, but it also buys almost nothing in payroll: what it gets is tax base, and the record says the tax base is real.

VIII. What to Watch

Source: Microsoft.

Most of what happens next at Fairwater is already on a docket, a permit, or a contract calendar. Five markers:

• The build cadence toward 2 GW. The second data center campus (MKE16–17) is due for completion in 2028, with its fire-protection plan reviews still working through mid-2026. On the 15-building expansion (Campuses 3 and 4), the foundation permit for the first Durand Avenue building was issued May 14, 2026, with vertical construction applied for in June — the pace of those permits is the real test of Nadella’s “scale to 2 GW alone.”

• The generation calendar. Paris RICE — approved for July 2026 — now shows construction running through summer 2027. The ATC Western Feed substations energize in sequence — Jupiter in January 2027, Cosmos in June 2028; the Oak Creek turbines target August 2028, with the LNG plant and Rochester Lateral under construction beneath them. Red Oak Ridge needs a written order by October 1, 2026 to close and break ground, its five turbines phasing in from December 2029 to June 2030.

• The tariff endgame. Microsoft’s request for clarification and Oracle’s petition to reopen the Very Large Customer tariff (both filed June 10, 2026) were as yet undecided, and the tariff’s real incidence gets set in We Energies’ 2027–2028 rate case.

• The noise docket. Ostergaard v. Microsoft runs on a 12-to-14-month clock under the court’s standing order, with Microsoft’s response the first tell. In parallel, watch whether the cooling-fan mitigation Microsoft promised in April 2026 quiets the DNR complaint file.

• The January 1, 2028 test. Microsoft’s guarantee of $1.4 billion in assessed value on the expansion areas comes due, with shortfalls payable at up to $15 million a year — against a joint review board projection of $2.1 billion by the same date.

Subscribe now

Measured AI provides institutional-grade analysis of the physical infrastructure powering AI data centers. For access to our full research library, regulatory intelligence, and weekly briefings, visit MeasuredAI.com.

The developer’s head of policy presented, a property-tax consultant and an ERCOT representative spoke, no one named the tenant, and the court took no action. Five days after that, the tenant’s name surfaced anyway, in a Rule 144A bond offering presentation circulated to institutional investors: Amazon Data Services, Inc., guaranteed by Amazon.com, Inc., had signed a 15-year lease for the entire facility back in March.

That sequence — a county objects in principle, a developer presents without naming its customer, and the bond market receives the full terms within a week — is the current state of the Permian Basin AI buildout.

A downloadable PDF edition of this report is attached at the end of the post for paid subscribers.

Read more

It is an unlikely address. West Virginia has almost no data center industry — none of the hyperscale clusters of neighboring Virginia or Ohio — and that near-blank slate is part of the story: the state spent 2025 rewriting its laws to court exactly this kind of project.

Microsoft owns none of it. Not the land, not the gas plant that will power the campus, not the buildings, and not the GPUs. The owner-operator is Nscale, a “neocloud” — a GPU-rental specialist — that will build Monarch, run it, and sell Microsoft the compute that comes out the other end. Microsoft is the anchor tenant. It is a posture Microsoft has become increasingly comfortable with for its AI data centers, as shown by the 20-year triple-net leases it signed at its Fairwater Atlanta facility, and it is why Monarch is worth a close look.

The stakes are large in both directions. Because the land is undeveloped, Nscale must fund the build, which it markets as “more than $20 billion in planned capital investment” — the cost of the campus and its power plant, not the chips inside.

Source: Nscale.

The servers are a separate, larger layer: at recent NVIDIA NVL72 rack prices, roughly $50 billion for about a gigawatt of compute, a cost Nscale procures and expects to recover through Microsoft’s multi-year compute payments.

The campus ultimately targets 2 GW online by the first half of 2028 and roughly 8 GW by 2031. Against that ambition sits a thin contractual reality: the Microsoft commitment is a non-binding letter of intent, the site is greenfield, and the entire first phase rests on a single customer.

This piece walks the Microsoft wager in order: why a hyperscaler rents an entire gigawatt instead of building it, the neocloud and the deal behind it, the off-grid gas plant that is the real product, the campus and the cluster it will house, the water it will and won’t use, the fiscal bargain West Virginia struck to land it, and the stack of permits that still stands between Monarch and first power.

Subscribe now

I. The Wager: Microsoft Rents the Whole Stack

The Monarch campus will rise at 5533 Ohio River Road just north of Point Pleasant in Mason County, West Virginia.

It is situated on the banks of the Ohio River — at the state’s western edge, where the river forms the line between West Virginia and Ohio.

Most of the AI infrastructure Measured AI has covered is, underneath, a story about ownership. At AWS’s New Carlisle campus, Amazon owns the land, the buildings, the chips, and the network, and points all of it at one customer. At xAI’s Colossus cluster, the company went a step further and built its own grid rather than wait on the utility. Monarch is the third corner of that triangle, and the most counterintuitive: Microsoft owns nothing.

Microsoft has signed a non-binding letter of intent to rent the entire 1.35 GW first phase of an off-grid, gas-powered campus that Nscale will build, own, and operate. It is the hyperscaler bet inverted — not vertical integration, but near-total outsourcing of the stack, with the grid, the gas plant, the buildings, and the GPUs all carried by someone else.

Source: Nscale.

The Inversion — Outsourcing the Stack

Renting is not new for Microsoft; renting everything is the escalation. Microsoft’s head of business development and ventures, Jon Tinter, has described the company’s capacity strategy as a blend of “owned datacenters, leased facilities, and strategic collaborations” — three legs, of which Monarch is the purest expression of the third. The scale of that third leg is already substantial.

Nscale has become Microsoft’s largest single GPU-leasing counterparty, at around $46 billion in commitments:

The Monarch LOI is incremental to that ~$46 billion, not part of it.

Why rent the whole stack rather than build it? The deal structure and what is being built show Microsoft’s rationale:

1. The LOI is non-binding. Microsoft has committed to nothing it cannot step away from, while the capital at risk in the build — the $20 billion-plus Nscale must spend, the gas plant, the permits — sits on Nscale’s balance sheet, partly backstopped by guarantees from NVIDIA. This is an asymmetric structure: Microsoft holds an option on the capacity it reserved, and Nscale and its financiers hold the construction risk.

2. Renting buys geography and time. Monarch places NVIDIA’s newest generation Vera Rubin systems in a jurisdiction engineered for speed-to-power — cheap Appalachian gas, a statute that lets the campus skip the grid — without consuming Microsoft’s own interconnection queue or self-build pipeline.

3. Renting keeps a large liability off Microsoft’s own books. The Monarch power plant’s potential-to-emit runs to 11 million tons of CO2-equivalent per year, and the facility is a major source under federal Prevention of Significant Deterioration rules — a permitting posture that has already drawn organized local opposition. That permitted ceiling alone is roughly 77 times Microsoft’s entire FY24 direct-emissions (Scope 1) footprint. Microsoft pulls data centers it leases and operates into its own Scope 1 and 2; Monarch’s combustion stays out of both, because Nscale owns and operates the generation and sells compute as a service. That leaves the plant-level exposure and the regulatory fight on Nscale’s ledger — but not the carbon itself, which still lands in Microsoft’s Scope 3, the column its carbon-negative pledge also covers.

The Anchor — One Tenant, the Whole Phase

The clearest sign of how committed the two sides are is also the clearest risk. Microsoft’s 1.35 GW equals Nscale’s entire marketed Phase 1 compute — the sellable output of the ~2 GW of generation coming online by 2028 — all of it underwritten by one anchor customer rather than a diversified roster of hyperscalers. Before Nscale acquired the site, Amazon, Meta, and Fluidstack were all reported to have looked at Monarch; none transacted. The compute Microsoft has reserved will be delivered, per the LOI, as “multi-year compute services, alongside a long-term data center lease structure,” in multiple tranches beginning in late 2027.

Microsoft also claims to be the first hyperscaler to “power on” a Vera Rubin NVL72 system, which makes Monarch a showcase as much as a supply line.

Source: NVIDIA.

The tenant explains the demand. The owner explains the build — and that is a stranger story.

II. The Neocloud and the Deal: Nscale Owns Everything

Behind a hyperscaler that owns nothing stands a neocloud that owns everything. Nscale is a vertically integrated owner-operator: it builds and runs the underlying data center, supplies the GPUs and the software orchestration layer on top, and sells the resulting compute capacity. That integration is the pitch — one counterparty for land, power, hardware, and software — and it is why a company few in the industry could have named a year ago is now Microsoft’s largest GPU-leasing partner.

Nscale — One Owner, From Land to Orchestration

Monarch is Nscale’s flagship, and it arrived by acquisition rather than origination. In mid-March 2026, announced at GTC, Nscale agreed to buy American Intelligence & Power Corporation (AIP Corp) — the developer that held the up to 2,250-acre campus, its permits, and its power agreements — and folded AIP and its backers into a new “Nscale Energy & Power” division headquartered in Houston, Texas.

Rendering of AIP Corp’s prior data center design for the Monarch Cloud Campus in West Virginia. Source: AIP Corp.

AIP itself was a new joint venture of two energy and venture firms, Fidelis New Energy and 8090 Industries, with the project backed by LuminArx Capital Management. The campus traces back to “more than three years of development led by Fidelis.” That development began as something else entirely — a hydrogen venture, the “Mountaineer GigaSystem,” that drew West Virginia economic-development support back in July 2023, before a May 2026 filing recorded the project’s shift “from hydrogen production to the immediate establishment of advanced computing data centers.”

Ultimately, a hydrogen developer’s site became a neocloud’s flagship AI data center campus, and because the land was undeveloped, Nscale must now pay to build it — the “more than $20 billion” of planned capital investment it markets.

The Money — Series C and the NVIDIA Backstop

Nscale’s ability to make that commitment rests on a capital base assembled at remarkable speed and on a familiar backer. The company closed a $2 billion Series C in March 2026 at a $14.6 billion valuation — co-led by Aker ASA and 8090, the same 8090 that is half of the AIP venture that originated Monarch. This Series C round was billed as the largest such round in European history, with investments from Astra Capital Management, Citadel, Dell, Jane Street, Lenovo, Linden Advisors, Nokia, NVIDIA, and Point72.

NVIDIA sits at the center of the structure: it is Nscale’s largest preferred shareholder, with a stake of “more than 10%,” and it has separately agreed to guarantee up to $860 million of Nscale’s lease obligations for a 240 MW AI data center in Ward County, Texas. As 8090’s co-founder put it, “AI leadership is not a software problem. It is an energy and infrastructure problem.”

That structure is the neocloud model in miniature, and the model is the reason Monarch matters beyond Microsoft. Nscale is one of a cohort of “neoclouds” — GPU-rental specialists including CoreWeave, Crusoe, Core Scientific, Lambda, and Nebius — in a sector that, per Synergy Research, exited 2025 at roughly a $36 billion annual run-rate — fourth-quarter revenue of $9 billion, up 223% year over year.

The economics are capital-dominated, with GPU lifecycles around four years, and they are de-risked by exactly the kind of long anchor-tenant contract and hyperscaler backstop on display here. The “Microsoft-leases-from-a-neocloud” arrangement is not a one-off; Monarch is the template at its most leveraged, which is what makes its single piece of physical infrastructure — the power plant — the part worth understanding in detail.

III. Power Is the Product: The Behind-the-Meter Gas Microgrid

What Microsoft is really buying at Monarch is not real estate or even chips — it is power, delivered faster than any grid could promise. The campus generates all of its own electricity on-site and, by design, draws nothing from the public grid. Everything else at Monarch is downstream of that one choice.

Islanded by Law — HB 2014 and Speed-to-Power

Monarch is a behind-the-meter, islanded natural-gas microgrid that “generates all of its own power on-site, fully independent of the public electric grid.” That is not just an engineering decision; it is a legal one. West Virginia’s House Bill 2014, the “Power Generation and Consumption Act,” lets the campus operate as “its own islanded utility with special microgrid status,” and the air-permit application states plainly that the plant “will not be connected or capable of selling electricity to the local utility power grid.” There is no conventional interconnection.

As a certified microgrid, the campus is exempt from Public Service Commission jurisdiction over rates, certificates of convenience and necessity, and conditions of service, and from net-metering and interconnection standards; it may export no more than 10% of what it generates, and only to the wholesale market. A future grid tie-in for that capped export is contemplated, but the campus runs solely on the microgrid at launch.

The point of all of this is time. The developers cast islanded-utility status as a way to “significantly reduce the time to power,” and have marketed the site as “America’s first AI-dedicated, unregulated electric utility built to serve hyperscalers faster than the grid.” It is the same logic Measured AI traced in its Grid Bypass analysis — when interconnection queues run 4–7 years, on-site generation is not an ideological choice but a mathematical one — taken to its endpoint. Monarch is not on the grid. By law, it is its own grid.

The Generation Fleet — 864 Engines, 2.16 GW

The plant that delivers this is built from reciprocating engines, not turbines. The WVDEP air permit specifies 864 Caterpillar G3520K natural-gas reciprocating internal combustion engines, each driving a 2.5 MW generator, for 2.16 GW installed.

The prime-power train, one of 864 identical units: pipeline gas is preheated and pressure-regulated, burned in a Caterpillar G3520K genset, and its exhaust scrubbed by a dedicated SCR and oxidation catalyst before release. Source: WVDEP.

They sit on 12 power nodes of 72 engines each — four nodes powering each of the three “H”-shaped data center buildings, or 288 engines per data center building.

Each power node = 72 engines = 180 MW installed (eight engine halls of nine engines).

One of the 12 power nodes: eight engine halls of nine engines each (72 gensets, ~180 MW), arrayed around central electrical and mechanical equipment and routed toward the data center and the adjacent nodes. Source: WVDEP.

Total site peak demand is approximately 1.8 GW — about 0.36 GW below the 2.16 GW installed, the ~17% headroom built into the fleet.

Caterpillar agreed to supply 2 GW of fast-response G3500 series gas gensets for the initial phase, with deliveries scheduled September 2026 through August 2027 via dealer Boyd CAT and vendor financing from Caterpillar Financial.

Source: Caterpillar.

Caterpillar notes the units can “ramp from zero to full load in approximately seven seconds.” That speed matters because AI training does not draw power smoothly.

Battery Storage — Smoothing the AI Load

To this end, the gensets are paired with battery energy storage to manage the rapid load fluctuations associated with AI workloads, with battery power delivery expected to begin in 2026. For this off-grid data center campus, battery storage does two distinct jobs — riding through outages and smoothing the load before it ever reaches the generation.

That second job — load smoothing — is the key advantage of using battery energy storage for AI data centers, and it is the rationale most specific to a data center campus like Monarch. Synchronized AI training does not draw power evenly: hundreds of thousands of GPUs step in and out of computation in lockstep, producing extreme power jitter with 10–20 MW shifts several times per second. A battery energy storage system wired in parallel counter-cycles against that load, discharging on the spikes and charging on the dips so the source sees a smooth draw.

Backup and Emissions — The Major-Source Footprint

Behind the prime fleet sits a separate emergency tier and a large emissions footprint. This backup power equipment includes 48 Caterpillar C175-20 diesel emergency engines (16 per data center building), 12 Generac SD-1000 diesel units, six Clarke diesel fire-water-pump engines, and six natural-gas fuel-gas heaters.

The emergency and fire-pump units are modeled at 100 hours per year, while the prime fleet’s potential emissions assume 8,725 operating hours annually. Each prime engine carries a clean-emission module: selective catalytic reduction, stated to cut nitrogen oxides by 98%, paired with an oxidation catalyst cutting carbon monoxide 98% and formaldehyde 95%.

Each of the 864 reciprocating engines vents through its own stack, but the stacks are grouped tightly in twos and fours between the engine halls, so they are viewed as roughly 240 merged exhaust points rather than 864 — a distributed array of hundreds of small stacks.

Paired engine halls in close-up: the exhaust stacks are co-located in two- and four-stack clusters — the basis for the permit’s ~240 merged-stack sources. Source: WVDEP.

The power node in three dimensions: paired engine halls lined with gensets, rooftop equipment, and the central exhaust stacks rising above the roofline. Source: WVDEP.

Even so, the totals are those of a major source. Facility-wide potential-to-emit is NOx 506.01, CO 1,043.73, SO2 47.79, PM2.5 38.85, and 619.76 tons per year of hazardous air pollutants — the HAP figure formaldehyde-dominated, enough to make the site a major HAP source — with volatile organic compounds stated two ways (382.91 tpy on one basis, carried as 899.00 tpy on the ozone-precursor basis used for PSD) and greenhouse gases at 11 million tons per year of CO2-equivalent.

Notably, the enabling law imposes no obligation to clean any of this up at the source: HB 2014 removed the requirement that microgrid power come from renewables, and the only emissions-offset plan Nscale describes — carbon sequestration — is aspirational, with no tonnage, timeline, or operator attached.

The Gas Supply — The Prosperity Line

A gas plant this large needs a dedicated fuel artery, and Monarch’s is a new lateral with an old trunk behind it. The lateral is Hope Gas’s “Prosperity Line,” announced March 3, 2026 alongside West Virginia Governor Patrick Morrisey. Hope Gas, the regulated West Virginia utility, will invest in and operate it; the permit names Monarch Cloud Campus, LLC as applicant and an AIP affiliate, M2 WV Midstream, LLC, as facility company.

The first phase is a 30-mile line running from Putnam County’s Southern Terminus north into Mason County’s Northern Terminus to the plant just north of Point Pleasant, following the corridor of West Virginia Route 62, the Ohio River, and the CSX railroad.

Source: WVDEP.

The permit authorizes both a 24-inch and a 30-inch line — routine for a system that pairs a larger trunk with a smaller delivery lateral — across a 556.5-acre earth-disturbance corridor. It is a $250 million Hope Gas investment supporting roughly 600 building-trades jobs, built by open-trench construction with dam-and-pump stream crossings, begun in April 2026 and targeted for completion by the end of 2026.

Where the gas ultimately comes from has not been announced publicly. Public statements stop at Appalachian Marcellus-Utica gas sourced from West Virginia. The most likely upstream interconnect is a tap on Columbia Gas Transmission’s Mountaineer XPress (MXP) natural-gas pipeline, the FERC-certificated interstate line whose 164.5-mile, 36-inch mainline carries about 2.66 Bcf/d and runs directly through both Mason and Putnam Counties.

Source: FERC.

Three things point to MXP:

• Geography: MXP is the only high-capacity interstate line WVDEP places in both counties.

• Pressure: Gas arrives at the plant at transmission-class ~1,200–1,400 psig, consistent with an interstate tap rather than a distribution main.

• Capacity: A ~2 GW plant burns on the order of 0.3–0.4 Bcf/d, a small fraction of MXP’s throughput, though the line is roughly 98% subscribed, so Monarch would run on a transportation contract.

At the Monarch plant, gas is stepped down from that transmission pressure to under 150 psig at three pressure-letdown stations, each preheated by the six fuel-gas heaters, before feeding the 864 engines.

The plant is the campus’s heart; the buildings around it are where the plan becomes visible on the ground.

Subscribe now

IV. Site and Build: The Campus in Acres and Phases

On the ground, the wager is about 1,100 acres of West Virginia farmland.

Source: WVDEP.

The buildout is phased: 2 GW online by the first half of 2028, rising to roughly 8 GW by 2031. The building shape is the developer’s AI data center design, not Microsoft’s. The site sits at 5533 Ohio River Road in Mason County, along West Virginia Route 62 about 5.5 miles north of Point Pleasant, with the Ohio River forming the West Virginia–Ohio line just to the west.

The Site — Greenfield Acres, Built in Phases

Company sources market a site of “up to 2,250 acres” — which is also the maximum parcel allowed under West Virginia’s data center statute — while the air-quality permit covers only an “approximate 1,100-acre site.” Either way it is one of the largest available AI data center sites in the country, and it is greenfield: there is nothing there to retrofit.

Construction is underway — the same May 5, 2026 inspection that logged “large dust clouds” came amid dust, silt-fence, and flooding complaints across late April and May 2026.

Source: WVDEP.

The on-site power plant’s pad is the furthest along: “Power Generation Pad South,” an 80.8-acre pad, was filed in January and issued in March 2026.

The “H” — Wings Around a Core, by the Developer

The defining feature of the campus is a building shape. Each of the three data centers is built as four wings and a central core forming an “H.” These are the same wings that the 12 power nodes feed, four per building.

Source: WVDEP.

Each of the three data center buildings comprises ~1.7 million square feet of built area, for a total of ~5 million square feet of built data center area across the campus. Each of the four wings requires nominally 150 MW at peak, or roughly 600 MW per data center building.

The “H” configuration is a purpose-built-for-AI geometry — repeatable high-density wings, each its own power-and-cooling domain, radiating off a shared core that concentrates networking, electrical distribution, and liquid-cooling plant — and it is a layout that can be commissioned one wing at a time.

It is also, pointedly, not Microsoft’s. Microsoft does not use a four-wings-plus-core “H” at any of its own Fairwater AI data centers, which solve the cluster problem differently — three large rectangular buildings in Wisconsin, two-story buildings in Atlanta. The “H” is the developer’s design.

V. Compute and Network: The Vera Rubin Cluster

The compute block is built entirely on NVIDIA Vera Rubin NVL72 systems — merchant hardware, bought off the shelf like the rest of the stack. The cleanest way to read the cluster is to climb it one rung at a time, from the chip to the campus.

The Chip and the Rack — Vera Rubin and the NVL72

Vera Rubin is the rack generation after Blackwell — “Vera” succeeds NVIDIA’s Grace CPU, “Rubin” the Blackwell GPU. The chips assemble into the NVL72 rack, which pairs 36 CPUs and 72 GPUs and is 100% liquid-cooled.

The Vera Rubin superchip pairs two Rubin GPUs with one Vera CPU on a single board; 36 of them make up the 72-GPU NVL72 rack. Source: NVIDIA.

The 100%-liquid-cooled NVL72 compute tray: two Vera Rubin superchips on a cable-free modular assembly, the unit that stacks into the rack. Source: NVIDIA.

What matters for Monarch is not the spec sheet but the power it implies. NVL72 rack power runs 180–220 kW, up from 120–140 kW for the current GB200/GB300 generation — a 50–80% rise that flows straight through to coolant flow, distribution units, and the shift toward 800 VDC (volts direct current) power.

From Rack to Campus — Thousands of Racks, One Fabric

Stacked up, those racks are what the “H” exists to hold. Applying the generic 180–220 kW per rack to the capacity Microsoft reserved implies on the order of 6,000–7,500 NVL72 racks, since that capacity is marketed compute that carries cooling and distribution overhead. The reason to gather them into wings around a core is physical: an AI training run is one coherent job, and scale-up networks are bounded by copper connectivity distances, so keeping every hall within short reach of a shared networking spine is the whole game.

It is the same instinct behind Stargate housing a single cluster across eight connected buildings, and behind Microsoft’s own framing of a “singular AI supercomputer underpinned by a single, flat networking infrastructure.” Monarch’s four wings, fed by four power nodes and cooled as one closed loop, are that idea expressed in a single building.

Subscribe now

VI. Cooling and Water: Closed-Loop, Fill-Once

The water story, like the gas plant, is engineered to defuse the one fight that could stall the campus. Monarch is built to consume almost no municipal water, a political choice as much as an engineering one: in a county where water capacity is already a live issue, the ability to promise that the data centers will not touch the drinking supply is worth as much as any efficiency number.

Cooling Architecture — Closed Loops, Chip to Air

Heat leaves Monarch through closed loops. The data centers run chilled-water plants that reject their heat to air through closed-circuit (dry) coolers rather than evaporative cooling towers, while the on-site gas engines shed heat through dry radiators. Because both reject to air, the water loops recirculate with no continuous evaporative makeup. Inside the halls, the Vera Rubin racks are themselves 100% liquid-cooled, so the building’s thermal design is liquid all the way from the chip to the heat-rejection plant. The cooling loops are closed-loop systems with minimal water makeup — a fill-once character rather than a continuous withdrawal.

The Water Posture — Zero Drinking Water, and the County

The headline claim is specific, and so is its fine print. Nscale states the cooling systems are closed-loop with zero drinking water consumption, promising that “Monarch will not use Mason County’s drinking water supply for cooling operations,” and the joint Microsoft announcement adds a “high-efficiency design that consumes less water with no impact on municipal water supply or residential users.” That zero applies to municipal and drinking water — not to all water: Fidelis’s own launch materials tout “industrial water access to enable large-scale operations,” so the campus does draw process water even as it avoids the public drinking supply.

The site drains toward a tributary of the Ohio River — the nearest surface water sits about 295 feet north of the site’s South pad — and falls inside both the Point Pleasant Water Works wellhead-protection area and the Mason County Public Service District’s Crab Creek source-water-protection area, in a watershed WVDEP has confirmed as impaired though with no Tier 3 streams. Meanwhile, local water infrastructure is already mobilizing around the project: the county has studied doubling the Lakin Water Treatment Plant’s capacity and identified five projects under a $68.3 million capital plan.

VII. Fiscal and the Law: What West Virginia Traded for the Wager

West Virginia didn’t just permit the owner-operator model — it wrote a fiscal regime around it, one that decides who can be charged for a plant the public will never own and bars the tax breaks data centers usually receive. The same law that lets Monarch island its power sets those terms.

That regime was a recruiting tool. West Virginia entered the AI build-out with almost no data center industry — a scattering of small facilities, none of them hyperscale — and set out to build one by statute. HB 2014’s own findings pitch the state as offering “the least restrictive regulatory environment in the Nation,” written against the “national security vulnerability” of so much American compute concentrated in Loudoun County, Virginia. Monarch, at up to 8 GW, is the largest project the effort has landed; Google has since bought land for a campus in neighboring Putnam County, and real estate developer Penzance has announced a roughly $4 billion campus in Berkeley County. The market is real but early, and every entrant carries the same power, permitting, and community-response risk Monarch does.

The Project Economics — Capex, Revenue, and Jobs

The economic case Nscale markets is large and, so far, projected rather than committed: more than $20 billion in planned capital investment, $80-plus million in new annual tax revenue (itemized as $40 million for Mason County schools, $9 million for county services and infrastructure, $23 million for statewide income-tax relief, and $8 million for water and power improvements), and thousands of temporary construction jobs plus hundreds of permanent operational roles in Phase 1.

Source: Nscale.

The Tax Treatment — No Holiday, No PILOT

Unusually, the project does not come with a property-tax holiday: under HB 2014, operators pay ad valorem real and personal property tax at the same millage as any other business, with no blanket exemption. What the law does bar is payments-in-lieu-of-taxes and tax-increment financing for the microgrid’s generation and distribution property — treatment available only where at least 75% of the output serves the data center load — under a special-valuation article that sunsets December 31, 2055.

Where a positive tax increment does arise, the statute splits it: 50% to the state Personal Income Tax Reduction Fund, 30% to the situs county, 10% to all counties per capita, and 5% each to an economic-enhancement fund and a grid-stabilization fund.

Ratepayer and Local Preemption — Insulation Both Ways

The law’s other half is about insulation — of the public from the project’s costs, and of the project from local control. West Virginia law has the state “occupy the whole field,” barring counties and municipalities from limiting a certified microgrid district or high-impact data center and exempting projects from local zoning and building-code enforcement. On the cost side, utility customers outside the district “shall not bear any construction or operational costs” of utility property built solely to serve inside it — the statutory basis for Nscale’s claim that “state law prohibits passing any project costs on to ratepayers.” It is a regime built to move fast and to keep the public ledger clean, which is also why the only friction left is regulatory.

VIII. What to Watch

Source: Nscale.

Monarch is, for now, a marketed gigawatt and a greenfield with dust on it. The next two years will show whether the hyperscaler-owns-nothing model holds. The markers, roughly in order:

• Caterpillar genset deliveries — September 2026 through August 2027. The 2 GW of fast-response gas engines are the long-pole item; the delivery window is the real schedule for first power.

• The Prosperity Line — main line targeted for completion by the end of 2026. No gas, no plant; watch for final WVDEP action on the still-pending WVR312628 permit.

• Microsoft’s first tranche — late 2027. The non-binding LOI converts to reality (or doesn’t) when the first Vera Rubin capacity is actually delivered and taken.

• 2 GW online by the first half of 2028; roughly 8 GW by 2031. The phasing targets that turn a single anchor lease into a flagship campus.

Subscribe now

Measured AI provides institutional-grade analysis of the physical infrastructure powering AI data centers. For access to our full research library, regulatory intelligence, and weekly briefings, visit MeasuredAI.com.

By early 2026 — eighteen months later — the bet had tripled. Colossus is now a three-building cluster straddling the Tennessee–Mississippi line — about 2.6 million square feet, roughly 770,000 GPUs, and about 1 GW of IT compute that Elon Musk projects will reach 2 GW in the coming years.

Source: X, @xAIMemphis.

What makes Colossus unlike the other gigawatt-class campuses Measured AI has covered is how it is powered. Amazon’s New Carlisle and Microsoft’s Fairwater bet everything on the utility grid; xAI did the opposite. Only one of its three data centers draws firm power from the grid. The other two run entirely off-grid, behind the meter, off a 1.2 GW natural gas plant xAI built for itself across the state line in Mississippi.

That self-supplied power is the engine of the cluster’s defining trait — speed. xAI retrofits existing industrial shells rather than building greenfield, stands up its own generation rather than waiting years for a grid connection, and leases rather than buys much of the equipment in between. It built the machine for its own Grok models, then turned it into a merchant utility: Claude developer Anthropic now leases roughly 325,000 GPUs at $1.25 billion a month, Google has contracted about 110,000 more, AI coding editor Cursor trains on it, open-source startup Reflection rents GB300s at Colossus 2, and the U.S. government runs classified workloads on it. The same off-grid choice that bought the speed has also drawn a federal Clean Air Act suit — and a Department of Justice intervention to keep the gas plant running.

This primer walks the bet layer by layer: the strategy and the tenants, the SpaceX-scale money behind it, the brownfield sites and the speed they bought, the compute that behaves as one machine, the self-built power that feeds it, the cooling and water it consumes, the two-state fiscal bargain, and the litigation now testing it.

Subscribe now

I. The Bet: Build Your Own Grid, Rent the Speed

Most frontier AI capacity is tethered to the utility. The developer waits in an interconnection queue, signs an Energy Service Agreement (ESA), and builds around the date the grid says power will arrive. Colossus inverts that model. xAI treats the grid as too slow, generates its own electricity behind the meter, drops its hardware into buildings that already exist, and races competitors to capacity measured in months rather than years. Then it rents the result.

Source: X, @xAIMemphis.

Every layer of the cluster expresses the same wager — the shells, the turbines, the megawatts, the tenants, the deal terms, and the lawsuits all trace back to a single choice to own the schedule by self-supplying the power.

The Inversion — Off-Grid by Design

The cluster’s signature is what it refuses to wait for. Rather than tether its data centers to the utility, xAI generates its own electricity behind the meter — only Colossus 1 draws firm grid power at all — and that choice is what lets it build at a pace with no real peer (Section V traces the per-site split and the plant behind it). When you do not wait on a utility, you build at the speed of the building and the turbines.

The Tenants — Rivals and the Pentagon

xAI built Colossus for itself — it runs its own frontier Grok models on the cluster, including Grok-4.3 (April 2026) and Grok-5, now training at Colossus 2 and 3 — but the machine is no longer xAI’s alone. It has become a merchant compute utility whose customers include xAI’s most direct rivals and the Pentagon. xAI is simultaneously the landlord and an anchor tenant.

The lead customer is Anthropic. In May 2026, xAI and SpaceX signed Cloud Services Agreements giving Anthropic access to roughly 325,000 NVIDIA GPUs across the Colossus data centers to run its Claude models, with Anthropic paying $1.25 billion per month through May 2029; after an initial three-month period either party may terminate on 90 days’ notice. The May 6 announcement first framed it as “all the compute capacity at Colossus 1” — over 220,000 GPUs / more than 300 MW — but that was the capacity live within the first month; 325,000 is the full contracted scope across Colossus 1 and Colossus 2.

A second customer is a competitor too. On June 5, 2026, xAI signed a Cloud Service Agreement with Google for about 110,000 NVIDIA GPUs at $920 million per month from October 2026 through June 2029 to expand capacity for its Gemini models. The deal carries a delivery-milestone gate: if SpaceX fails to deliver the committed GPUs by September 30, 2026, Google may terminate after a one-month grace period, and after December 31, 2026 either party may exit on 90 days’ notice.

The third commercial tenant is Cursor (Anysphere, Inc.), which signed a compute agreement with SpaceX on April 19, 2026 — structured as a compute-for-data/IP collaboration rather than a cash compute lease — and ran reinforcement-learning mid-training of its Composer model on the Colossus 2 cluster. On June 16, 2026, SpaceX announced an all-stock acquisition of Cursor at a $60 billion equity value, expected to close in Q3 2026.

The fourth customer is an open-source challenger. On June 22, 2026, SpaceX signed a computing-power agreement with Reflection AI — a startup last valued at $25 billion, building American open-source models to rival OpenAI, Anthropic, and Google — giving it immediate access to NVIDIA GB300 chips at Colossus 2. Reflection pays $150 million per month from July 1, 2026 through 2029, roughly $6.3 billion over the term, with either party able to exit on 90 days’ notice after the first three months.

The fifth customer is the U.S. government. Per the U.S. Department of Justice’s national-security filing and a Department of War declaration, the Colossus 2 and 3 campuses run classified workloads for the military. A Department of War Chief Digital and AI Office declaration describes a Grok Gov model running in Maven Smart Systems that “enabled U.S. forces to deploy over 2,000 munitions to 2,000 distinct targets within 96 hours during Operation Epic Fury,” consumes roughly 2 billion tokens per day on the Top-Secret network, and is one of only four frontier models cleared for national-security use — and one of three cleared for Secret and Top-Secret networks. That national-security role becomes the fulcrum of the litigation in Section VIII.

Read together, the roster says something about Grok itself. xAI built Colossus to train its own frontier models, yet Grok has trailed its rivals by a wide margin: xAI’s own IPO filing counts about 117 million monthly users of Grok’s AI features as of March 2026, against roughly 1 billion monthly users for ChatGPT. In March 2026 Musk posted that xAI “was not built right first time around, so is being rebuilt from the foundations up.” A data center developer whose own models were absorbing a machine this size would have little of it left to lease; instead xAI rents more than half the cluster — all of Colossus 1 among them — to rivals it set out to beat. The merchant-utility model, in that light, is as much a concession that Grok underperformed as the deliberate business line xAI presents.

The customers explain the urgency; the next layer explains how a company posting a $2.5 billion operating loss in Q1 2026 alone can underwrite a gigawatt of compute at this speed.

II. Ownership and the Money: Underwritten at SpaceX Scale

Colossus exists at this scale because SpaceX absorbed xAI and put a rocket company’s balance sheet behind it. The cluster is financed with bridge-loan and IPO capital, built across a layered stack of single-purpose entities spanning two states, and underwritten by purchase commitments that dwarf current revenue.

SpaceX Owns xAI — the Entity Map

On February 2, 2026, SpaceX completed the “xAI Merger,” making xAI a wholly-owned subsidiary. The consideration was 321.7 million Class A plus 121.7 million Class B shares plus $2.95 billion cash (pre-split), valuing the deal at $125 billion. xAI is now SpaceX’s “AI segment“ — AI compute, Grok, and X.

Source: Space Exploration Technologies Corp.

Beneath the parent, three entities carry the assets:

• CTC Property LLC, a wholly-owned xAI subsidiary organized as a Nevada LLC, operates the Tennessee data centers — Colossus 1 (3231 Paul R. Lowry Road) and Colossus 2 / MACROHARD (5420 Tulane Road) — plus the future Colossus Water Recycling Plant.

• MZX Tech LLC, a wholly-owned xAI subsidiary organized as a Wyoming LLC, holds the Mississippi assets — the Southaven Gas Plant (2875 Stanton Road South) and Colossus 3 / MACROHARDRR (2400 Stateline Road West).

• Stateline Power LLC is the turbine-owning joint venture between Solaris Energy Infrastructure and xAI.

The ground and buildings, unlike the equipment, xAI mostly owns outright:

• xAI, via SpaceX, bought the 785,000-square-foot Colossus 1 building in Memphis — two parcels of land totaling 217 acres — for $185 million (May 23, 2026) from Phoenix Investors, having leased it since March 22, 2024.

• xAI’s CTC bought the Colossus 2 building at 5420 Tulane Road in Memphis and two adjacent Tulane sites (5414 Tulane and 5408 Tulane) for $79.9 million (February 26, 2025).

• In Mississippi, MZX owns both the Colossus 3 building at 2400 Stateline Road West (the former GXO warehouse, acquired December 19, 2025) and the Southaven Gas Plant at 2875 Stanton Road South (the former Duke plant, acquired July 16, 2025).

How the Build Is Financed

The financial backdrop is a company spending far ahead of its revenue. AI-segment capital expenditure ran $12.7 billion in FY2025 plus $7.7 billion in Q1 2026, against FY2025 segment revenue of $3.2 billion and an operating loss of $6.4 billion.

Source: Space Exploration Technologies Corp.

That gap is bridged at SpaceX scale, chiefly by the June 2026 IPO — priced at $135.00/share and closed June 15, 2026 (over-allotment exercised in full) for roughly $86.2 billion gross ($85.7 billion net), its first stated use of proceeds “the expansion of our AI compute infrastructure.” Those proceeds must also repay, within six months, a $20 billion unsecured bridge loan a Goldman Sachs-led syndicate extended in March 2026 to refinance legacy X/xAI debt. Total unconditional purchase obligations stand at $25.6 billion, $21.5 billion of it concentrated in 2027 — several times current segment revenue.

What the spend is racing to meet is contracted demand: the Anthropic and Google leases alone commit a forward run-rate of roughly $26 billion a year at full ramp (about $15 billion and $11 billion respectively), with the Reflection lease adding about $1.8 billion a year more — a recurring annual figure many times the segment’s FY2025 revenue, and one that arrives only as capacity is delivered.

Off the Books — and On

The gas turbines that power the off-grid sites are not owned by xAI; they are rented through a joint venture structured to keep them off SpaceX’s balance sheet. Stateline Power, formed in April 2025, is owned 50.1% by Solaris Energy Infrastructure and 49.9% by xAI. Each side contributed $86 million (Solaris in-kind via Solar Turbines purchase orders; xAI in cash).

The Stateline Power JV supplies up to ~900 MW off-grid under a seven-year commercial/rental contract “underpinned by a 15-year JV,” with revenue beginning in 2026. Despite the near-even split, Solaris consolidates Stateline as the primary beneficiary of a variable-interest entity and operates it, while SpaceX carries the stake at equity method ($80 million by March 31, 2026) — so xAI rents, rather than owns, the permitted Southaven turbines.

The turbine financing is layered on top. A Stonebriar Commercial Finance delayed-draw term loan funds the fleet — a maximum of $550 million or 80% of equipment cost, 9.85% fixed over 6 years with $324 million drawn by April 28, 2026 — and the equipment is rented to CTC and MZX. Separately, SpaceX agreed on April 30, 2026 to buy mobile gas turbines for about $2.0 billion, atop $925 million of forward turbine commitments through 2029.

The same lease-don’t-buy logic reaches the most expensive layer inside the buildings — GPU accelerators: Valor Equity Partners, whose founder Antonio Gracias has sat on SpaceX’s board for over 15 years, raised a $5.4 billion fund (Apollo $3.5 billion, with NVIDIA itself a backer) to “buy and lease” NVIDIA GB200 GPUs to xAI under a “triple net lease structure.”

Source: Space Exploration Technologies Corp.

But where the turbines stay off the balance sheet, the GPUs do not: SpaceX records the Valor deals as sale-leasebacks — financings that keep the chips in its own PP&E against about $9.0 billion of related-party debt by March 31, 2026, the latest tranche (April 24, 2026) booked as a five-year lease worth $6.6 billion in undiscounted payments.

III. Site and Speed: Brownfield Shells, Stood Up in Months

The money buys speed only because of how xAI builds. Every site in the cluster is a retrofit of an existing industrial building, not a greenfield construction project — and that single choice is the mechanism behind build times no competitor matches.

The Retrofit Model

xAI’s playbook is a brownfield retrofit strategy leveraging existing industrial sites with direct-to-chip liquid cooling. Each Colossus building began as something else: Colossus 1 an Electrolux home-appliance factory, Colossus 2 a logistics/distribution warehouse, Colossus 3 a GXO Logistics warehouse/distribution center, and the Southaven Gas Plant a former Duke Energy natural gas power station.

Speed rests on three legs — the retrofit itself, on-site gas-plus-Megapack power, and equipment leasing (servers from Dell and Supermicro; turbines via Solaris and VoltaGrid). To move at permit-speed, xAI used Darana Hybrid, a midsize Ohio-based general contractor, as a primary permit-submitting contractor for electrical, mechanical, plumbing, and related work on the Memphis buildings.

The construction implication is that each new GPU generation (NVIDIA H100, H200, GB200, and GB300) lands in a building that already stands, rather than triggering a fresh greenfield build.

The payoff is cost as well as speed: by xAI’s own accounting, building Colossus 2 ran about $2.7 million per megawatt, against an industry benchmark of roughly $12.3 million per megawatt — a more than four-fold cost advantage from dropping hardware into a standing shell rather than building from scratch.

The Cluster at a Glance

Together the three data centers span Memphis (Shelby County, Tennessee) and Southaven (DeSoto County, Mississippi).

These three Colossus data center buildings occupy roughly 207 acres (excluding adjacent land) and about 2.6 million square feet of building area, and together hold roughly 770,000 GPUs. As of March 31, 2026, the cluster’s combined IT compute power was about 1 GW, with Musk projecting roughly 2 GW once the Colossus cluster is fully scaled. The table below is the cluster’s data spine; the sections after it unpack each layer.

Colossus 1 — the Flagship

Colossus 1 is where xAI proved the model — the founding site, occupying the former Electrolux home-appliance factory at 3231 Paul R. Lowry Road in the Frank C. Pidgeon Industrial Park of southwest Memphis. The decommissioned plant was converted into a single data center building of 785,000 square feet on an 81-acre site, operated by CTC Property LLC, the utility customer of record.

Source: Space Exploration Technologies Corp.

Operational since July 2024, its first cluster of 100,000 H100 GPUs (about 130 MW) was built in 122 days — the build Jensen Huang called “superhuman” — and xAI then doubled it to 200,000 GPUs in a second phase of about 92 days.

Source: Supermicro.

Today the site holds over 220,000 GPUs — a mix of NVIDIA H100, H200, and GB200 accelerators — delivering more than 300 MW of IT compute, and it is the only site in the cluster that draws firm power from the grid (Section V).

Colossus 2 — MACROHARD

Colossus 2 is the cluster’s largest building and its fastest to stand up — branded MACROHARD in giant rooftop lettering, it sits at 5420 Tulane Road in the Whitehaven district of Memphis. The site spans 78.6 acres and a single converted warehouse of 1,007,838 square feet.

Source: Space Exploration Technologies Corp.

xAI kicked off the project on February 26, 2025, acquiring the warehouse and two adjacent sites, and brought the building online in January 2026 — its first cluster of about 110,000 GB200 GPUs (~210 MW) in 91 days.

Source: X, @elonmusk.

This was followed by a second cluster of about 110,000 GB300 GPUs (220 MW) in 64 days. Colossus 2 now holds approximately 350,000 GPUs (NVIDIA GB200 and GB300) and delivers about 450 MW of IT compute.

Source: Space Exploration Technologies Corp.

Colossus 2’s supporting build sprawls across neighboring parcels: a 312,365-square-foot pre-engineered metal building (with a 131,040-square-foot expansion permitted) plus Megapack pads at 5414 Tulane; three central utility chiller plants and electrical buildings on the 5420 Tulane parcel itself; and a 230 kV–500 kV gas-insulated substation in a 16,000-square-foot building on a 27.3-acre parcel at 5408 Tulane, distributing behind-the-meter power.

Colossus 3 — MACROHARDRR

Colossus 3 carried the playbook across the state line — brought online in Q1 2026 as MACROHARDRR, it sits at 2400 Stateline Road West in Southaven, DeSoto County, Mississippi, about 0.9 miles from the Southaven Gas Plant. The site is a former GXO Logistics warehouse built in 2001, retrofitted to host data center operations on roughly 47 acres in a single building of 810,225 square feet.

Source: Space Exploration Technologies Corp.

It holds approximately 200,000 GPUs and delivers about 250 MW of IT compute. Like Colossus 2, it runs off-grid and serves the multi-tenant roster set out in Section I, with xAI training its own Grok-5 there.

The shells and the speed are only half the bet. What goes inside them — and binds the buildings into one machine — is the compute.

Subscribe now

IV. Compute and Network: Hundreds of Thousands of GPUs as One Machine

The reason xAI concentrates all of this hardware in a few enormous buildings is architectural: the cluster is engineered to behave as a single coherent machine, not a collection of server halls. Every accelerator deployed to date is merchant NVIDIA silicon, fused by an Ethernet fabric that is itself the cluster’s most distinctive technical choice — with custom silicon held in reserve as the forward bet. The layers below build from the smallest unit upward: the chip and the rack, then the fabric that links racks, then the cluster that behaves as one computer.

The Chip, the Server, and the Rack

The smallest coherent unit is the rack, and its design changed with the GPU generation. In the H100 generation, a rack holds 64 GPUs as 8 HGX H100 NVL8 servers — but there is no rack-wide coherent scale-up domain; the NVLink domain spans only 8 GPUs per server.

Source: Supermicro.

The GB200/GB300 generation collapses that limit: NVL72 racks create 72-GPU coherent NVLink domains at roughly 130 kW per rack. These appear in limited numbers at Colossus 1 (about 30,000 GB200) and as Colossus 2’s building block, and they raise the coherent scale-up domain from 8 to 72 GPUs in a single generational step.

Source: Space Exploration Technologies Corp.

The Fabric — Spectrum-X, Not InfiniBand

Above the rack, at Colossus 1, the scale-out fabric is NVIDIA Spectrum-X Ethernet, not InfiniBand — a defining networking decision, used in xAI’s 200,000-GPU H100/H200 deployment.

Source: Supermicro.

Each GPU connects through one 400GbE NVIDIA BlueField-3 SuperNIC into Spectrum-X SN5600 switches (64-port, 800 Gbps, 128×400G links), a design for which NVIDIA claims 95% data throughput. The fabric has a ceiling: the reference architecture (v1.2) supports up to 100,000 GPUs, and “up to 512,000 GPUs with a non-reference custom design“ — still short of the 1-million-GPU cluster Musk has said he is targeting.

Colossus Is One Computer

xAI frames itself as unique among AI labs in keeping all of its GPUs in one location, and its Anthropic-access statement describes Colossus 1 as “one of the world’s largest and fastest-deployed AI supercomputers … [featuring] over 220,000 NVIDIA GPUs.” Across the Colossus cluster, the GPUs are tied together by thousands of miles of high-bandwidth fiber so that training jobs can run across hundreds of thousands of accelerators without the communication bottlenecks of geographically dispersed systems.

Source: Google Earth.

Proximity is key. Colossus 1 sits just over 8 miles from Colossus 2 and Colossus 3, which themselves stand only a few hundred feet apart — directly across the road from one another, with the Tennessee–Mississippi border running between them.

Source: Google Earth.

The Forward Bet — Vera Rubin and Custom Silicon

The forward roadmap keeps the brownfield logic intact: merchant NVIDIA now, with Vera Rubin (VR200) flagged as the next generation, and custom silicon later.

Whatever the silicon, the construction model holds: each generation drops into an existing industrial shell. And those shells have to be fed — which brings the bet to the layer xAI chose not to outsource.

V. Power and the Grid: The Layer xAI Builds Itself

Power is the layer every other hyperscaler outsources to the utility, and the layer xAI chose to build itself. The split is clean, and it is the cluster’s defining electrical fact: only Colossus 1 draws firm power from the grid; Colossus 2 and Colossus 3 run entirely off-grid, behind the meter. What follows traces the supply from the thin grid tie that feeds the flagship, through the curtailment terms that came with that tie, to the self-built plant that powers everything else, and the battery layer that lets none of it go dark.

The Thin Grid Tie — Colossus 1 Only

Colossus 1’s grid power runs through a federal-wholesale / municipal-retail chain. The Tennessee Valley Authority (TVA) is the wholesale generation-and-transmission provider; Memphis Light, Gas and Water (MLGW) is the local distribution utility that contracts with the customer. MLGW is TVA’s largest customer, serving about 440,000 electric customers and accounting for roughly 9% of TVA’s 2025 operating revenue, and the contracting entity is CTC Property LLC, its retail customer of record. Because TVA Board approval is required to make available firm power exceeding 100 MW, the grid tie advanced in board-sized increments: 150 MW approved November 7, 2024 and energized May 1, 2025 through a newly constructed substation (#63), then up to 150 MW more approved February 11, 2026, bringing the site’s ceiling (once energized) to as much as ~300 MW.

Who pays for the connection follows one principle: xAI fronts the cost, and MLGW ends up with the asset. xAI builds the substations and covers the transmission upgrades they require — a system-impact study flagged $1.7 million of improvements to a 161 kV transmission line that had to be completed before service could begin. It then recovers the outlay through monthly rebates against its utility bill until the build cost — about $35 million for the first substation, plus another $20 million for a second 200 MVA substation — is recouped, after which MLGW takes ownership. So if xAI ever leaves, “a brand new substation … at their cost“ stays with the utility.

On-Site Gas and Solar — Colossus 1

Colossus 1 also generates on-site. It is permitted for 15 Solar SMT-130 turbines, each rated 16.5 MW, for about 247 MW, each unit fitted with an oxidation catalyst, DLE/SoLoNOx, and selective catalytic reduction to a NOx limit of 2 ppmv, with the fleet capped at 87.14 tons of NOx per year. The combined generation total of nearly 400 MW — comprised of ~247 MW of on-site gas turbines plus the 150 MW utility feed — supports the site’s more than 300 MW of IT compute described earlier.

xAI is layering on renewable generation as well: an 88-acre ground-mounted solar array — the “xAI Photovoltaic Generation Station” — is going in on the adjacent 136.4-acre parcel at 3223 Paul R. Lowry Road. It was issued a $25.5 million construction permit in May 2026 and feeds power through medium-voltage switchgear into the data center.

Source: Space Exploration Technologies Corp.

Mandatory Curtailment — the Price of the Grid Tie

The grid tie came with a condition xAI’s other off-grid sites avoid entirely: xAI must agree to be cut. Both TVA approvals were “contingent on the new customer’s agreement to certain demand response terms,” and MLGW’s CEO was blunt — “I do not support 150 megawatts unless they are enrolled in an aggressive demand response program.” By February 2026, xAI’s current load was enrolled in a TVA Demand Response program, distinct from TVA’s standard voluntary program. The terms are deep but bounded: cuts of between 50% and 80% of xAI’s demand when the grid is stressed, “maybe four to five times a year,” in events “only between 2-4 hours,” against TVA notice windows of 5 minutes, 30 minutes, 60 minutes, 4 hours, and 12 hours — and an annual interruption total of 24 to 96 hours. Non-compliance is penalized financially, with the ultimate sanction that the utility can just not provide the power.

The point of the off-grid architecture is that none of this takes the Colossus data centers offline. When the grid is curtailed, Colossus 1 rides through on behind-the-meter generation and Megapacks — over 240 batteries let the Electrolux facility go completely offline during emergency situations or times of peak demand, and xAI’s Memphis executive Brent Mayo has said the company would go 100% off TVA’s grid and rely on Megapacks when the grid is stressed. That interruptibility was the enabler: demand-response participation was key to accessing grid power faster than typical timelines, letting xAI draw 150 MW less than a year after requesting it, precisely when TVA was at risk of not meeting “planning reserve margin” targets over the next five years.

Critics flag the catch — curtailing the grid draw can simply mean running the on-site gas turbines instead, so it need not cut local emissions.

The Self-Built Grid — the 1.2 GW Southaven Gas Plant

The defining electrical infrastructure sits across the state line. Colossus 2 and Colossus 3 are powered behind the meter by the Southaven Gas Plant at 2875 Stanton Road South in Southaven, Mississippi — a former Duke Energy facility, now a self-built simple-cycle combustion-turbine plant operated by MZX Tech LLC, with approximately 1.2 GW of continuous generating capacity.

Source: Google Earth.

The plant will not export power to the grid — it is the cluster’s private generation source, not a grid draw. A gas turbine is, in essence, a jet engine bolted to the ground — it burns natural gas to spin a shaft that drives an electrical generator, rather than producing thrust to push a plane. The 41 turbines themselves are rented by xAI, not owned, through the Stateline JV described in Section II.

The roster above is the plant’s permitted, permanent configuration. The turbines physically running at Southaven today are a larger, temporary fleet of portable units bridging the load until that permanent plant is built — and it is that portable fleet, not the permitted one, that the Clean Air Act suit in Section VIII contests.

The arrangement is set down in the plant’s own tax agreement, which defines its purpose as generating electricity for xAI’s own use, consumed directly by data centers that sat, at signing, outside Mississippi — the Tennessee buildings — and delivered across the line by “temporary transmission lines and permanent below ground transmission pipelines.” It is the off-grid, cross-state architecture written into a primary legal document, down to the buried lines that carry the power over the border.

The off-grid posture is documented on the Tennessee side by what xAI did not request. MLGW supplies the Colossus 2 site only around 0.5 MW of general service and has stated flatly that “MLGW is not supplying power to their supercomputer, Colossus 2.“ An initial 260 MW request for the Colossus 2 site was not studied, and a “260 MW to 1.1 GW“ range was discussed, but no final request was ever filed or studied. Colossus 2 runs off the Southaven Gas Plant, and Colossus 3 shares the same supply.

Batteries — the World’s Largest Megapack Deployment

The layer that lets the whole arrangement ride through disturbance — and absorb the violence of the AI load itself — is storage, not generation. Across the cluster, Tesla Megapacks valued at around $1 billion provide battery backup, peak shaving, and load smoothing.

Source: Tesla Inc.

SpaceX calls the Colossus cluster the “world’s largest Megapack deployment” and the “first gigawatt-scale Megapack battery installation.” Colossus 1 grew from 168 Megapacks (installed May 22, 2025) to over 240.

Source: Tesla Inc.

Colossus 2 holds more than $375 million worth of Tesla Megapacks. They connect into the building’s medium-voltage distribution — wired through the switchgear and step-down transformers that take in the Southaven Gas Plant’s behind-the-meter power.

Source: X, @xAIMemphis.

Each grid-forming Megapack 2 XL pairs roughly 2 MW of power with more than 3.9 MWh of storage. Across the fleet, that storage does two distinct jobs — riding through outages and demand-response events, and smoothing the load before it ever reaches the generation.

That second job — load smoothing — is the key advantage of using Megapacks for AI data centers, and it is the rationale most specific to a cluster like Colossus. Synchronized AI training does not draw power evenly: hundreds of thousands of GPUs step in and out of computation in lockstep, producing what Elon Musk has called “extreme power jitter” — “like having an orchestra [that] can go loud to quiet very quickly, at the sub-second level,” with “10–20 MW shifts several times per second.” A Megapack wired in parallel counter-cycles against that load, discharging on the spikes and charging on the dips so the source sees a smooth draw.

Source: Tesla Inc.

Megapack’s grid-forming control is credited with stripping out 70%+ of the variability while holding enough charge to do it around the clock. For a data center cluster that generates its own electricity, that is closer to a requirement than a refinement. Field data from Tesla on the Megapack shows the same jitter destabilizing generator power, frequency, and even mechanical vibration when the batteries are switched off — so the Megapacks are part of what lets the Colossus cluster’s private gas-turbine fleet serve a load that can swing by tens of megawatts in a second.

Subscribe now

VI. Cooling and Water: The Mitigation That Isn’t Running

If power is where the bet succeeds, water is where it strains. The Colossus cluster’s cooling is retrofitted into buildings that were never designed for it, and it draws today on the Memphis Sand Aquifer — the region’s drinking-water source — because the flagship mitigation engineered to offset that draw, the Colossus Water Recycling Plant, is permitted, partly built, and on indefinite pause.

Cooling Architecture

Colossus 1’s cooling is retrofitted into the Electrolux shell and built around liquid cooling at the rack level. Each rack combines two methods: direct-to-chip cooling, which circulates coolant through cold plates placed directly on the GPUs, and rear-door heat exchangers — liquid-cooled coils in the back of the rack that capture the heat the servers exhaust.

Source: Space Exploration Technologies Corp.

Together direct-to-chip cooling and the rear-door heat exchangers handle roughly 30–40 kW per rack, rising past 50 kW when the rear doors’ fans are engaged. A Supermicro coolant distribution unit (CDU) at the base of each rack isolates this rack loop from the building’s water and drives it through a redundant pair of pumps.

Because the rear doors absorb the exhaust heat, each rack is “room-neutral” — it adds no net heat to the data hall, sparing the building the room-scale air handlers a conventional data center relies on. That rack-level liquid path is also what makes the brownfield strategy work: at roughly 130 kW apiece, the densest GB200/GB300 NVL72 racks throw off more heat than any air-cooled room could remove, so only direct-to-chip cooling can pack GPUs at that density into a shell never designed for the load.

At the facility level, the captured heat is rejected outdoors through a hybrid of wet and dry equipment. Colossus 1 pairs open-loop evaporative cooling towers and water-cooled chillers with air-cooled chillers that shed heat straight to the air; the chillers need only drop the loop temperature a few degrees before the water recirculates.

Source: Space Exploration Technologies Corp.

Colossus 2 carries the same wet-and-dry logic into a newer design — its central utility plants house a hybrid of dry coolers and adiabatic cooling units, which run dry for most of the year and switch to evaporative assist only on the hottest days.

Aquifer Draw Today

Today, the Colossus cluster’s cooling still runs on potable water from the Memphis Sand Aquifer — the same source that supplies the region’s drinking water — and it does so by deliberate design rather than by drilling private wells.

Source: Protect Our Aquifer.

When xAI arrived in 2024, MLGW steered it away from sinking its own wells, which would have pulled unmetered groundwater, and toward buying metered municipal water instead. xAI agreed, tapping an existing water main at Colossus 1 that already had spare capacity, and it pays the standard commercial rate with no incentive. Its roughly 1 million-gallon-a-day draw adds only marginally to MLGW’s Davis Wellfield, which already pumps well over 20 MGD (and has a 30 MGD capacity).

Source: Protect Our Aquifer.

Colossus reaches the aquifer through the municipal system, not through wells of its own. Most of that water is consumed as makeup water for evaporative cooling, replacing what boils off Colossus 1’s cooling towers. xAI’s stated design demand is 1.1–1.3 MGD, but its actual draw is harder to pin down. The clearest public read comes from the advocacy group Protect Our Aquifer, which in June 2026 said MLGW’s own meter data showed the Colossus facilities using over 800,000 gallons a day — approaching 1 million, and rising sharply over the prior year even as the Colossus Water Recycling Plant meant to offset that draw sat unfinished.

Colossus Water Recycling Plant

The asset meant to take the cluster off the aquifer is the Colossus Water Recycling Plant, a water-reclamation facility that xAI’s CTC Property LLC is building at 3644 Riverport Road in Memphis.

Source: Protect Our Aquifer.

Permitted by the state of Tennessee as a non-potable reuse system, it is designed to take up to 13.5 MGD of treated effluent from the City of Memphis’s T.E. Maxson Wastewater Treatment Facility and treat it into industrial-grade reclaimed water — cutting the cluster’s aquifer withdrawals by up to 13 MGD. It is a zero-discharge plant — and a substantial one: the engineering plans show a full membrane-bioreactor (MBR) treatment complex with biological anoxic and aerobic stages, closer to a municipal wastewater plant than the simple filtration tie-in the word “recycling” might suggest.

Source: City of Memphis.

Once running, the plant would feed the permanent water-cooled chillers and cooling towers that replace today’s interim setup, supplying reclaimed water in place of aquifer draw — enough, in principle, for Colossus 2 to run as a net-zero-water data center. xAI puts the aquifer offset at roughly 13 MGD, or about 4.75 billion gallons a year (13 MGD × 365).

xAI would build, own, and operate the plant at its own expense. The company puts its own investment at well north of $80 million, but MLGW’s President and CEO, Doug McGowen, told the Memphis City Council in April 2026 that the plant now looks likely to cost closer to $200 million. The reclaimed water would not serve Colossus alone. The state permit names the TVA Allen Combined Cycle Plant as a second end user alongside the data center, with a slot held open for a third, as-yet-unnamed user.

For now, none of this is happening. xAI broke ground on October 27, 2025, but by April 2026 the work had stopped — an indefinite pause, with crews redirected to other sites.

Source: X, @xAIMemphis.

As Musk put it: “We need to focus on finishing Colossus 2 and ensuring it is extremely stable, then we will build the water recycling plant.” Until the plant runs, the interim water-cooled and rental air-cooled chillers carry the cooling load, and the makeup water keeps coming from the aquifer.

VII. Fiscal: Two States, Two Opposite Bargains

The cluster sits in two states that struck opposite bargains to host it. Tennessee gave xAI essentially nothing and collects full freight; Mississippi assembled a sales-tax exemption and fee-in-lieu deals around a pledge it calls its largest economic-development project ever. The contrast is the fiscal expression of a single cluster split across a state line.

Tennessee — No PILOT, Full Freight

Tennessee’s posture is unusual for a multi-billion-dollar build: no payment-in-lieu-of-taxes (PILOT) and no abatement. The Shelby County Assessor’s office put it plainly — “they did not get a PILOT … so they pay full freight” — and Shelby County Assessor Javier Bailey confirmed “xAI did not get a PILOT nor any other kind of tax rebate … no abatement whatsoever,” where a standard PILOT would have meant “a 75% decrease.” The Greater Memphis Chamber’s Dr. Bobby White explained why: “PILOTs don’t work for startups … they don’t have a lot of property” to abate.

Tennessee did, however, attach a binding community-benefit mechanism. Memphis Ordinance No. 5953 passed 11-0 on August 19, 2025, routing 25% of xAI’s AI-property tax revenue into a South Memphis Revitalization Action Plan fund within a five-mile radius, capped at $100 million, plus 1% for education and marketing. xAI-funded items to date include $750,000 to restore two Boys & Girls Club afterschool programs, and MLGW separately spent about $22 million on Orange Mound electric-reliability upgrades.

Mississippi — Sales-Tax Exemption and Fee-in-Lieu

Mississippi did the opposite. The Mississippi Development Authority (MDA) “approved xAI for its Data Center Incentive, which provides a sales and use tax exemption for all computing and equipment software” for certified data centers, and the City of Southaven and DeSoto County “are also supporting xAI’s projects through fee-in-lieu agreements,” with “MZX Tech – MDA-approved FILA Fully Executed” entered into the county record on March 2, 2026. Governor Tate Reeves and the MDA framed the stakes accordingly, announcing “investment exceeding $20 billion … the largest economic development project in Mississippi’s history” on January 8, 2026.

DeSoto County and Southaven granted MZX Tech two separate 30-year fee-in-lieu agreements — one covering the data center, with a projected capital investment north of $10 billion and at least 20 jobs, and one covering the power plant, effective November 21, 2025, scaling from over 400 MW to more than 1.2 GW, with over $900 million invested and at least 10 jobs. On the power plant, the fee is fixed at one-third of the taxes it would otherwise owe, and while the term runs 30 years, no single asset is abated for more than 10 years. Where Tennessee abated nothing, Mississippi forgives two-thirds of the gas plant’s tax bill — the price it paid to be the state that took the turbines.

Property-Tax and Utility Revenue

The investment claims split sharply by state — Tennessee’s TVA board memo cites Colossus 1 as “$5 billion … of capital with over 200 high-tech jobs,” while Mississippi’s pledge exceeds $20 billion. Tennessee’s property tax revenue is real and growing: the Shelby County Assessor estimated about $13.5 million of new county revenue for FY2025 (personal property only, since the building was leased prior to its purchase by xAI in May 2026). City building permit valuations corroborate the order of magnitude xAI is spending — the single largest is a $1.8 billion permit for the Colossus 2 main building alteration.

The cleanest local upside is utility revenue, and it is unsubsidized. MLGW expects xAI’s electric bill to run $7 million per month, exceeding cost-to-serve, and the added sales contribute ~$500,000 of additional PILOT payments annually to the City of Memphis (a municipal-utility PILOT, not an xAI tax break). xAI is on the customary commercial rate / published industrial rates with “no … incentives being sought for taxes,” and funds its own service infrastructure (the transmission-line upgrade at xAI expense, Section V).

On jobs, the official counts are only claims — Tennessee figures range from “over 200” (TVA) to “~300+” (MLGW) to “hundreds of permanent jobs” (xAI), while Mississippi claims “hundreds of permanent jobs,” and “in excess of 100 immediately” for Colossus 3.

Subscribe now

VIII. Regulatory and Litigation: The Bill for the Off-Grid Bet

The off-grid choice that bought the speed also bought the legal exposure. The Southaven Gas Plant is the crux: an as-built turbine fleet that bears little resemblance to the one Mississippi permitted, the basis for a Clean Air Act citizen suit — and, in turn, for an extraordinary federal intervention to keep the plant running on national-security grounds.

Permitted vs. As-Built — the Turbine Gap

On paper, Southaven is a 41-turbine fleet of about 1.2 GW under a Mississippi Department of Environmental Quality (MDEQ) Prevention of Significant Deterioration (PSD) permit issued March 11, 2026:

• 17 Solar PGM-130 (16.5 MW each)

• 16 Solar Titan 350 (35 MW each)

• 8 ProEnergy 6000PE (50 MW each)

• 10 PLUM T6500 pressure-reduction systems

Every permitted turbine carries SCR plus an oxidation catalyst, with NOx BACT at 2 ppmv under NSPS Subpart KKKKa, and the plant is classified a PSD major source; its PSD application’s potential-to-emit table (Trinity Consultants) lists NOx 423.39 tpy, VOC 417.40, CO 364.16, SO2 156.53, and CO2e 6,410,729 tpy.

But that permit covers the permanent plant, still under construction. The turbines actually running at Southaven are a separate, temporary bridge: in a July 29, 2025 letter, MDEQ found these portable, trailer-mounted units exempt from permitting altogether as “mobile sources” — so long as each stays affixed to a flatbed trailer and on-site under twelve months — and xAI began deploying them the following month to supply the off-grid sites until the permanent plant is finished. Whether that exemption holds is what the suit actually turns on.

The bridge fleet is mechanically different, too: it grew over time from 18 → 27 → 33 → 46 → 57 units of different makes and models — and kept growing, the defendants’ own June 2026 declaration reporting 59 on-site, with a 60th expected by the end of the month. As of the May 15 count, “none of the 57 turbines … are included in that permit.” That fleet of 57 — as the plaintiffs’ expert (Kim Laufenberg, of the Environmental Integrity Project) classified the Trinity Consultants inventory against aerial imagery — breaks down as:

• 14 Solar SMT-130 — Solar Turbines (Caterpillar) Titan 130, ~16.5 MW; the only units fitted with SCR

• 27 GE Vernova TM2500 — aeroderivative, mixed Gen 4/6/7/8

• 1 M35 — GE LM2500+G4 engine, ~35 MW (mobile-package brand not named in the record)

• 3 Jereh JT35 — GE LM2500+G4 engine

• 12 Mitsubishi FT8 — Mitsubishi Power package, Pratt & Whitney FT8 engine

Per Laufenberg, none of these 57 units are among the 41 the MDEQ permit authorizes. The defense — and Trinity’s own list — classify the lone M35 and the 3 JT35 as TM2500s instead.

Because emissions are measured in tonnage, not turbine count, the NOx figures span an enormous range strictly by basis: 87.14 tons of NOx per year is the Colossus 1 permit limit; 423.39 tpy the Southaven permanent-plant PTE (from the PSD application); and a plaintiff expert (Laufenberg) puts the unpermitted Southaven fleet’s potential-to-emit at 1,732.71 tpy (27 turbines, March 3, 2026), 2,507.96 tpy (33 turbines, April 14, 2026), and 5,300.54 tpy (57 turbines, May 15, 2026).

At roughly 5,300 tpy, the plaintiff calls Southaven the largest NOx emitter across Mississippi, Tennessee, and Arkansas and in the top 0.1% of about 52,000 U.S. facilities — against Memphis International Airport at 1,077 tpy and the comparable 1.1 GW TVA Allen plant at 230 tpy (a plaintiff estimate, not adjudicated).

Colossus 1 set the precedent. Its own temporary fleet peaked at about 35 turbines (~421 MW) by the Southern Environmental Law Center’s (SELC) March 31, 2025 aerial count — largely unpermitted, “the largest industrial emitter of NOx in Memphis,” having begun as “up to 18 gas turbines in June and July of 2024.” SELC demanded a cease order and $25,000/day penalties, and the temporary units were demobilized from about May 2025 as grid power and the permitted 15 came online. The same permitted-versus-as-built pattern then reappeared, larger, in Mississippi.

NAACP v. X.AI and the DOJ Intervention

The dispute became NAACP v. X.AI Corp. & MZX Tech, a Clean Air Act citizen suit — and then something far less ordinary. On June 15, 2026, the United States moved to intervene as plaintiff and to dismiss the case with prejudice, arguing that the NAACP’s suit “threatens American national, economic, and energy security.” The government’s filing leans on the cluster’s defense role: a Department of War CDAO declaration ties the “Grok Gov” model in Maven Smart Systems to the same Operation Epic Fury mission and Top-Secret workload described in Section I, now invoked as grounds to keep the plant powered.

The defendants — represented by Butler Snow, Vinson & Elkins, and Boyden Gray — confirmed on the record what the cluster runs and for whom: the Colossus 2 and Colossus 3 facilities operate “Grok, and other frontier AI models, including Anthropic’s Claude and Google’s Gemini,” hold “approximately 550,000 GPUs,” and serve the tenant contracts set out in Section I. Mississippi’s governor weighed in alongside them: Tate Reeves’ May 29 letter framed the project as the state’s largest-ever investment (Section VII), called the turbines exempt “mobile sources,” and cited “vital contracts with the War Department.”

The result is a federal posture without much precedent in data center permitting — the executive branch moving to dismiss an environmental suit so that a private, off-grid power plant can keep a gigawatt of AI compute online.

IX. What to Watch

Source: Space Exploration Technologies Corp.

Most of what comes next at Colossus turns on a handful of public markers — a federal docket, a TVA decision, a permit, and a scale-up.

• The NAACP / DOJ docket. The decisive near-term event is whether the court grants the United States’ June 15, 2026 motion to intervene and dismiss with prejudice. The outcome determines whether the as-built Southaven fleet keeps running as-is, or whether the Clean Air Act claim over the allegedly unpermitted plant — 57 turbines at the May 15 count, since grown to 59 — proceeds.

• The Colossus cluster’s scale-up toward ~2 GW. Cluster IT compute stood at about 1 GW as of March 31, 2026; Musk’s ~2 GW projection rides on Colossus 3 scaling up. Its progress is the clearest gauge of whether the off-grid pair reaches its stated ceiling.

• Whether the recycling plant restarts. The Colossus Water Recycling Plant is permitted and partly built but on “indefinite pause.” A restart — and approval of the Reclaimed Wastewater Management Plan “required prior to operation” — is what would finally take the cluster’s cooling off the Memphis aquifer; the aquifer draw remains the live number.

• Google’s September 30, 2026 delivery gate. If xAI fails to deliver access to the committed ~110,000 GPUs by September 30, 2026, then after a one-month grace period (end of October 2026) Google may either walk away or accept the GPUs delivered at a pro-rata-reduced fee — the first hard test of whether xAI’s build speed can keep its merchant commitments on schedule.

• Grok-5. Grok-5’s training run at Colossus 2 and 3 is the in-house workload against which the rented capacity competes for the same machine. Having slipped past Elon Musk’s Q1 and Q2 2026 targets, Grok-5 is now expected to launch in the second half of 2026.

Subscribe now

Measured AI provides institutional-grade analysis of the physical infrastructure powering AI data centers. For access to our full research library, regulatory intelligence, and weekly briefings, visit MeasuredAI.com.

Source: Amazon.

The campus reads as much like a utility project as an AI infrastructure deployment. Eighteen data center buildings are operating today, with fourteen more under construction. Finished, its 32 buildings will hold about 6.5 million square feet and draw 2.25 GW from the grid — about the electricity used by 1.5 million average homes, and enough to make the campus the single largest customer in Indiana Michigan Power’s 100-plus-year history. Behind them: a planned fleet of 878 backup diesel generators, an $11 billion price tag that ranked as the largest capital investment in Indiana’s history when announced, and a water debate that has filled county hearing rooms.

What makes New Carlisle worth a close read is that everything here is Amazon’s own — the land, the buildings, the chips, the network, and the deal terms around them — an unusually complete picture of how a frontier AI campus actually gets built. This overview walks that bet layer by layer: why the campus exists and who it serves, then the land and its three phases, the silicon and the network that make 32 buildings behave like one machine, the power, the water, the investment — and the markers that will test it.

Subscribe now

I. The Bet: One Customer, One Chip, One Machine

Most frontier AI capacity is assembled from other people’s parts — a developer’s shell, merchant NVIDIA silicon, a shared tenant roster. New Carlisle is the opposite choice taken to its maximum: Amazon owns the land, the buildings, the chips, and the network that fuses them, and has pointed all of it at exactly one customer. Each layer of the campus expresses the same bet in a different way — acres, silicon, megawatts, water, deal terms.

Flagship Campus for Project Rainier

New Carlisle is the flagship and largest site of Project Rainier, AWS’s large-scale AI compute cluster built with Anthropic.

Source: Amazon.

Project Rainier is deliberately distributed across multiple U.S. states — New Carlisle, St. Joseph County, Indiana, is its primary campus, with a second, smaller campus near Canton, Madison County, Mississippi, which has more than 350 MW of IT capacity today and is set to ramp past 1 GW, also on Trainium2.

The Indiana and Mississippi campuses reflect a single strategy: AWS is building multiple large AI campuses around its own accelerator platform, Trainium. For training the current generation of frontier AI models, AWS treats the entire New Carlisle campus as a single unit of compute — one vast, tightly networked machine.

Anchor Tenant — Anthropic

The New Carlisle data center campus is devoted to a single customer, Anthropic, which uses it to train and serve its Claude models. Amazon and Anthropic co-engineered both the site and the silicon inside it — Anthropic’s input on what to accelerate in the chip shaped the hardware. Project Rainier was sized to deliver more than five times the compute — the exaflops — that Anthropic used to train its previous generation of Claude models.

The Amazon–Anthropic alliance runs deep. Amazon invests in Anthropic: up to $25 billion more (a $5 billion equity stake plus a $20 billion milestone-gated financing facility), on top of the $8 billion already invested. Anthropic buys from AWS, committing to spend more than $100 billion on compute over 10 years and reserve up to 5 GW of Trainium, with roughly 1 GW of Trainium2/Trainium3 online by the end of 2026.

The first layer of the bet is land.

II. Site and Land: The Bet in Acres and Phases

Ownership starts with the ground. The lots, the buildings on them, and the construction program are all Amazon’s own — and the phasing below is the bet: scaling in place, phase added onto phase, as Anthropic’s training demand grows.

The Campus at a Glance

Amazon’s New Carlisle data center campus spans two subdivisions inside the Indiana Enterprise Center (IEC) that sit across State Road 2 from each other, in Olive Township, St. Joseph County, northern Indiana. New Carlisle is situated approximately 80 miles east of Chicago and about 15 miles west of South Bend, Indiana.

Source: St. Joseph County Office of Economic Development.

The data center campus is developed and operated by Amazon Data Services, Inc. This hyperscale self-build is organized into three phases:

• Phase 1 is the North Campus, known as SBN100

• Phase 2 is the South Campus, known as SBN201

• Phase 2 Expansion, known as SBN210, is also part of the South Campus

Source: Google Earth.

Fully built, the three projects will comprise 32 data center buildings on roughly 882 acres formally mapped for development, with about 6.5 million square feet of building area and a grid draw of 2.25 GW:

* The three building lots total ~805 acres (232 + 179 + 394); the remaining ~77 acres of the 882 mapped are dedicated road right-of-way carved out of the IEC-3 subdivision.

North Campus — Phase 1 (SBN100)

Phase 1 anchors the campus on the single IEC-2 lot north of State Road 2: nine data center buildings, each about 35 feet tall and 216,000 square feet in size — roughly the footprint of a Walmart — together delivering an estimated 590 MW of IT capacity, about 65 MW apiece. All nine buildings are operational.

Source: Google Earth.

After applying in April 2024 and breaking ground in September 2024, Phase 1 received its Indiana Department of Environmental Management (IDEM) air permit in January 2025, drew first power that same month, was certified “construction complete” in December 2025, and had its operating permit validated in February 2026.

South Campus — Phase 2 (SBN201)

Just over a mile south, across State Road 2, Phase 2 sits on Lot 1 of the IEC-3 subdivision and mirrors the North Campus in form — nine buildings again, each slightly taller at about 39 feet.

Source: Google Earth.

Phase 2 applied in September 2024 and received its IDEM air permit in May 2025, then mirrored Phase 1 — “construction complete” in December 2025 and operating permit validated in February 2026.

South Campus Expansion — SBN210 (Under Construction)

The campus’s next stage, SBN210, extends Phase 2 onto Lot 2 of IEC-3, directly adjoining SBN201 and forming one continuous South Campus. Its buildings vary in size (eight single-story, six two-story) and footprint (elongated linear, large-format rectangular, paired blocks) rather than repeating the earlier phases’ uniform design. AWS’ wetlands filings have put the expansion’s purpose in capacity terms: a stated minimum operational power capacity of 767 MW, with the preferred 14-building design rated at 775 MW.

Source: Indiana Department of Environmental Management (IDEM).

Build-out of SBN210 is underway. Amazon began initial site prep in Q1 2026, including grading for laydown areas, a substation, and some building pads. The South Campus Expansion filed its IDEM air-permit application in May 2026 and its wetlands/dredge-and-fill permit application in June 2026.

Satellite imagery from June 2026 shows Amazon has begun constructing multiple data center buildings on the SBN210 site. AWS has previously projected full New Carlisle data center campus build-out no earlier than late 2027.

Developer and Construction

Behind the build-out is Amazon itself, building and operating the campus directly rather than contracting the project to a third-party data center developer. To hold an aggressive schedule, AWS’s team brought on four general contractors to construct the campus — Walbridge, Clayco, Ryan Companies, and Angelo Iafrate.