Arvind Balakumar

From "General-Purpose GPUs" to "Custom Silicon": A Hardware Inflection Point Indian Startups Shouldn’t Miss Broadcom recently announced it has received over $10 billion in orders for custom AI racks and accelerators. That’s not just a procurement milestone. This marks a decisive shift in AI infrastructure—from broad-spectrum GPU reliance to purpose-built silicon engineered for high-efficiency, scalable performance. This surge reflects growing demand from frontier AI developers like OpenAI and others building large-scale foundation models—who increasingly require specialized infrastructure that goes beyond general-purpose GPUs to meet the scale, efficiency, and latency needs of next-generation training and inference. General-purpose GPUs, while versatile, are increasingly being outpaced by custom silicon when it comes to inference-heavy workloads. ASICs and other specialized architectures offer better power efficiency, reduced latency, and tighter integration for predictable, high-volume operations. And Broadcom isn’t alone. I see, Google continues to evolve its Tensor Processing Units (TPUs). Amazon has deployed Inferentia and Trainium across its cloud stack. Microsoft is building its own Azure Maia and Cobalt chips. There are startups that are reimagining compute from the ground up—focusing on memory bandwidth and wafer-scale innovation. But my main point is, this shift opens a rare window for Indian startups. With the right partnerships and design focus, there’s a clear opportunity to: - Build domain-specific silicon for EV, Automotive, AgriTech, Space, wearables, audio, and other edge AI domains - Leverage fabless models and global foundries to bypass capital-heavy barriers - Create IP-led differentiation in smart devices. There’s no need to replicate legacy architectures. The real opportunity lies in out-specializing them—especially in segments they won’t prioritize. This isn’t just a hardware story. It’s a blueprint for deep-tech sovereignty. Let’s not just observe the AI infrastructure race. Let’s build lanes of our own. Thoughts? #AIHardware #CustomSilicon #IndianStartups #SemiconductorIndia #DeepTech #StrategicAutonomy #TechLeadership #FutureProofIndia

40

2 Comments

Looking at $MU Q1 2026 earnings - CEO Sanjay Mehrotra on why AI = memory demand: “The value of memory in delivering AI capabilities, functionality, and performance is critical, and more HBM is required across AI platforms.” Key takeaways: - AI architectures are becoming more memory intensive, not less - HBM supply will be tight in 2026, alongside non-HBM DRAM - Micron now sees HBM as a $100B TAM by 2028 (pulled forward by 2 years) - GenAI’s shift from training → inferencing (and video) is also driving enterprise SSD growth AI isn’t just about compute, memory is the bottleneck.

3

AI tools have started making waves in chip design, and it's exciting to see how they can enhance productivity. While large language models like Claude are proving beneficial in RTL coding and verification, there's a growing array of new tools emerging from both established EDA vendors and numerous startups. With the current talent crunch in the chip development space, the introduction of these innovative tools could be key to improving team efficiency without compromising quality. If you are utilizing any of these AI tools and have experienced significant productivity improvements, please share your insights. Looking forward to a lively discussion on this topic. Let's get it started.

106

18 Comments

𝐆𝐨𝐨𝐠𝐥𝐞 𝐭𝐨 𝐈𝐧𝐯𝐞𝐬𝐭 $𝟏𝟓 𝐁𝐧 𝐢𝐧 𝐀𝐈 𝐃𝐚𝐭𝐚 𝐂𝐞𝐧𝐭𝐫𝐞 𝐢𝐧 𝐕𝐢𝐳𝐚𝐠 Google Cloud CEO Thomas Kurian announced that the tech giant will invest $15 billion to build the world’s largest AI and data centre in Visakhapatnam, Andhra Pradesh. This hub will be Google’s biggest AI facility outside the US and will feature Tensor Processing Units (TPUs) for cutting-edge AI computing. “This gigawatt-scale AI hub represents a major step in supporting India’s digital infrastructure goals, and we’re grateful for the support of the Government of India and the state of Andhra Pradesh,” said Kurian. Google Cloud CEO Kurian made this announcement during Google’s event on “India’s AI-Powered Future and Advancing the Vision of Viksit Bharat 2047,” in the presence of Ashwini Vaishnaw, Minister of Electronics & IT; Nirmala Sitharaman, Finance Minister; and N. Chandrababu Naidu, Chief Minister of Andhra Pradesh. Google Cloud Thomas Kurian Ministry of Electronics and Information Technology Via: Outlook Business | Deepsekhar Choudhury (Story link in the comment)

21

1 Comment

Cerebras Systems inference is now available on Amazon Web Services (AWS) Marketplace—access the world’s fastest AI inference with your existing AWS payer account, start in minutes with no long-term commitments and consolidated billing. We are always looking for ways to to make it easier for customers access the industry's fastest inference speeds.

222

10 Comments

What if the real bottleneck in AI isn’t compute at all… but memory? On CNBC Television, Micron CEO Sanjay Mehrotra described something a lot of AI conversations skip: memory isn’t “just a part in the box” anymore. In an AI world, it behaves more like a strategic asset. Here’s the tension: AI is spreading everywhere at once—from data centers to the edge. Models keep evolving. Workloads move from training into inference, then outward into devices. And every step of that expansion quietly demands the same thing: more memory, and faster memory. That’s the twist. AI accelerators can look infinite on paper, but they still have to feed the model. If memory bandwidth and capacity don’t keep up, “more compute” turns into “more waiting.” Mehrotra also pointed to the uncomfortable reality on the supply side: even when companies decide to add capacity, the timeline doesn’t match the hype cycle. New greenfield fabs and manufacturing ramps take years, which means the infrastructure response lags demand. So the question for anyone building or investing in AI isn’t only “Which model wins?” It’s: who controls the memory roadmap that makes those models usable at scale? Do you think the next AI constraint shows up first in chips, memory, power, or data access? #AI #ArtificialIntelligence #GenAI #AIInfrastructure #Semiconductors #Memory #DRAM #NAND #DataCenters #EdgeAI

4

1 Comment

Micron Technology warns that the global memory shortage is not easing but intensifying and is now expected to persist well beyond 2026. According to Micron COO Manish Bhatia, the scale of the shortage is unprecedented, driven primarily by explosive demand for HBM used in AI accelerators. AI workloads are absorbing so much capacity that traditional markets such as PCs and smartphones are structurally undersupplied. Electronics manufacturers are already attempting to secure memory supply contracts extending past 2026. Across the industry, capacity is being fully booked: SK Hynix has sold out 2026 production, while Micron’s AI-focused output is fully contracted for the current year. Samsung and other producers are also shifting priorities toward AI and strategic customers like Nvidia. To meet demand, Micron is accelerating expansion through fab acquisitions, U.S. megaprojects, and sharply higher capital spending. This level of investment reflects a long-term structural bet rather than a typical cyclical upswing. However, memory remains a deeply cyclical market, and current pricing and margins already assume an “AI forever” scenario. When oversupply eventually emerges, the correction is likely to be fast and painful, potentially worse than in 2022.

2

1 Comment