After the Power Crunch, AI Infrastructure Hits a Silicon Wall

extract

After the Power Crunch, AI Infrastructure Hits a Silicon Wall A CNAS report argues that chip manufacturing and memory shortages are emerging as major barriers to the expansion of hyperscale AI, alongside growing power demands. For two years, the AI industry warned of an impending electricity shortage. Now, it faces an even greater threat: running out of chips. A new report from the Center for a New American Security (CNAS) argues that semiconductor manufacturing capacity – spanning advanced logic, HBM, and packaging – cannot keep pace with AI demand, threatening to slow the next wave of hyperscale expansion. “The world’s leading AI companies cannot get enough chips,” the report states, describing AI chip production as a “binding constraint on the pace of the AI compute buildout.” Microsoft, Amazon, Alphabet, Meta, and Oracle could collectively spend $700 billion or more on capital expenditures in 2026, with most of that spending tied to AI systems, data centers, and supporting infrastructure. That spending surge is colliding with a semiconductor supply chain that cannot expand at the same pace, the report said. The shift marks a reversal from much of 2024 and early 2025, when operators often described having GPUs they could not deploy because power infrastructure lagged behind hardware procurement. In 2024, Satya Nadella described situations in which Microsoft held AI chips that it lacked sufficient power capacity to energize. “You may actually have a bunch of chips sitting in inventory that I can’t plug in,” Nadella said during a 2024 discussion on AI deployment challenges. The mismatch is already visible across the market. Earlier this month, Data Center Knowledge reported that billions of dollars in AI compute remained underutilized or stranded even as hyperscalers accelerated infrastructure spending. Now the pressure may be moving further upstream into the semiconductor supply chain itself. Power and Silicon Move at Different Speeds Stephen Sopko, semiconductor and deep tech analyst at HyperFrame Research, said the industry now faces two separate physical problems unfolding on radically different timelines. “Silicon is the binding short-term constraint. Power is the binding long-term constraint,” Sopko told Data Center Knowledge. Power shortages stem from generation projects, transmission upgrades, transformer lead times, gas turbine lead times, and interconnection queues. Those systems often take years or decades to expand. Chip production is moving faster, but not fast enough to keep up with hyperscale AI demand. “CNAS and recent earnings commentary from TSMC, Micron, SK hynix, Nvidia, and Broadcom converge on the same message,” Sopko said. The report argues AI compute demand is now “outpacing many chip manufacturers’ forecasts.” That pressure extends far beyond GPU fabrication. Modern AI clusters depend on a tightly coupled manufacturing stack that includes advanced packaging, high-bandwidth memory (HBM), networking silicon, and leading-edge foundry capacity. Several parts of that chain already face severe shortages. Sopko described HBM and DRAM as the tightest pressure points in the current market, noting AI now consumes most global DRAM production capacity. Research firm Epoch AI estimates that total AI chip memory bandwidth has grown roughly 4.1x annually, underscoring how quickly AI demand is accelerating. HBM production also consumes significantly more wafer capacity than conventional memory, tightening supply as hyperscalers scale larger AI clusters. Meanwhile, advanced logic manufacturing remains heavily strained. Industry reports indicate TSMC’s advanced 3 nm (nanometer) capacity remains heavily utilized amid surging AI demand. The CNAS report also highlights advanced packaging as a major pressure point as hyperscalers race to deploy Nvidia Blackwell systems and next-generation accelerators. OpenAI CEO Sam Altman summarized the situation bluntly earlier this year: “Right now, again, it’s chips.” New capacity doesn’t arrive at software speed – tooling and qualification take years. (Image: Alamy) AI Buildouts Hit Physical Limits At the same time, AI expansion plans continue growing. OpenAI, Oracle, SoftBank, and partners tied to the Stargate initiative have outlined plans for massive deployments expected to consume unprecedented volumes of GPUs and supporting hardware. Hyperscalers are also investing more heavily in custom silicon platforms, such as Google’s TPUs and Amazon’s Trainium and Inferentia chips, to reduce dependence on Nvidia’s supply chain. Those efforts may ease some procurement pressure, but they do little to resolve broader manufacturing, packaging, and memory shortages across the AI hardware stack. But semiconductor manufacturing does not scale at software speed. New fabs take years to build. Clean rooms, specialized tooling, and advanced packaging lines further slow expansion. “The 12 to 24 month story isn’t whether silicon constrains the buildout,” Sopko said. “It will.” Those supply pressures are already reshaping deployment timelines. “These are deployment timelines bending around silicon, not power,” Sopko said, pointing to mounting strain across memory, wafer, and accelerator supply chains. Research from SemiAnalysis estimates memory could account for roughly 30% of hyperscaler AI spending in 2026, up from about 8% in 2023 and 2024, as HBM shortages ripple through the supply chain. None of this means power shortages disappeared. Utilities across major data center markets continue warning about AI-driven load growth. Interconnection queues remain clogged. Transmission upgrades, permitting delays, and generator procurement are still slowing deployments. Instead, the industry appears to be entering a phase in which multiple bottlenecks tighten simultaneously. In some markets, power availability still governs deployment schedules. In others, chip supply may dictate how quickly operators can bring new clusters online. That creates a far more volatile planning environment for developers trying to scale AI capacity at hyperscale speed. If GPU deliveries slip, operators risk building shell capacity faster than they can populate it with hardware. Capital continues burning even if racks remain partially empty. The result is an AI buildout increasingly shaped by interconnected physical systems stretching from substations and transmission corridors to memory fabs and advanced packaging lines.