Power, Not Chips, Is Now the AI Infrastructure Constraint

NVIDIA's dominance is ending because power, not chips, is now the binding constraint for AI infrastructure.

NVIDIA is in crisis because the constraint shifted. B200 Blackwell is stalled by HBM3E memory shortage; stock fell 17%; OpenAI + Broadcom announced a custom inference chip; Anthropic is co-developing on Samsung 2nm. Micron's revenue quadrupled to $41 billion, proving memory is the real bottleneck. Inference—where NVIDIA has pricing power—is being captured by custom silicon. Training, where NVIDIA still dominates, faces supply scarcity. This is not hyperscalers hedging; this is wholesale exit.

But chips are no longer the constraint. Brookfield + Bloom Energy announced $25 billion in fuel-cell partnerships for AI data centers; National Grid Ventures funded $1.75 billion in gas generation for Microsoft; Crusoe Energy hit $30 billion valuation on power-aware infrastructure. Power is now binding. Hyperscalers have 10+ GW of planned capacity and nowhere to plug it in. NVIDIA owns no power assets and has no path to owning them. Value is migrating from semiconductor vendors to power majors and grid operators.

Hyperscalers are going vertical to eliminate middlemen. Meta launched a cloud business to monetize excess capacity. SoftBank, OpenAI, and Oracle announced $500 billion in Stargate campuses—the largest AI infrastructure commitment ever. NVIDIA partnered with Firmus Technologies on a $20 billion Indonesian data center. Compute is moving to power, not power to compute. That inversion is permanent.

China is scaling independent of US sanctions. Meituan trained a 1.6 trillion parameter LLM on 50,000 domestic chips, proving export controls fragment geography but not capability. The US and China now operate separate compute economies. NVIDIA has zero leverage in either region.

The neocloud market is bifurcating. Together AI validated $8.3 billion open-source inference funding. But CoreWeave, Nebius, and IREN tumbled when Meta announced a cloud business. Distributed GPU clouds without power assets, integrated customers, or capital are structurally dead. Only vertically-integrated players survive: Together, Crusoe at $30 billion, hyperscalers. Selling compute commodities is no longer a business.

NVIDIA is a legacy vendor in a power-driven market. Watch Brookfield's fuel-cell deployment speed, Stargate's cost per kilowatt-hour, and whether custom silicon inference reaches cost parity with H100. Those metrics determine who owns the next decade.