Why Power, Not Chips, Is Now Reshaping the AI Capex Arms Race

The AI infrastructure wars have entered a new phase: power is the scarcity. This week's data points are unambiguous. Nebius raised GPU prices 17–21% and announced a third hike in three months, signaling genuine supply tightness rather than marketing noise (item 7). CoreWeave locked in $250 billion in new commitments at $40 million per megawatt short-term pricing, while the US government committed $5.25 billion to electrical-grid upgrades across 31 projects in 26 states (items 13, 14). The signal is clear: the bottleneck has shifted from manufacturing GPUs to delivering the megawatts to run them.

Anthropic's IPO filing crystallizes this inversion. The company disclosed $518 billion in infrastructure commitments, the bulk non-cancelable, anchoring a reported $2 trillion valuation and up to $100 billion in IPO proceeds. Broadcom's simultaneous $42 billion financing facility to Anthropic is not a chip deal—it binds Anthropic to Broadcom silicon for the duration of a power-constrained buildout (items 1, 3). The capital structure is inverted: AI labs now command leverage over suppliers. Anthropic is not buying compute; it is selling 20-year power commitments and recovering that risk through equity. That rewriting of who funds whom reshapes every downstream deal.

Power projects are now the deal flow. Vistra secured $4 billion from the U.S. Department of Energy for nuclear projects anchored to a 2,600 megawatt, 20-year purchase agreement with Meta (item 22, C4). GE Vernova received the first U.S. construction permit for a small modular reactor with a $176 billion reported backlog, 22% revenue growth driven by AI demand (C23). Nuclear utilities—yesterday's slow-growth incumbents—are now the go-to counterparties. Oracle's Project Jupiter, Broadcom's data-center division, and Applied Optoelectronics all face months of delays not because chips are unavailable but because grid operators cannot clear the capacity (item 18, C6). The constraint has moved upstream from fabs to substations.

Export-control workarounds are settling into structural form. Tencent leased 100,000 NVIDIA H100/H200 accelerators from Oracle's Southeast Asia data centers for $7 billion over five years, a transaction that circumvents chip-export bans through physical control retained in the US while providing functional compute access to a Chinese anchor tenant (items 19, 23, C3). The lease structure—not outright sale—exploits a gap in the regulatory framework. But the scale (100,000 units) and the fact that Oracle executed it underscore that offshore cloud-compute arbitrage is now mainstream.

Pricing power has visibly shifted to those controlling power. Nebius's 17–21% increase, announced alongside a $120 million acquisition of Israeli inference-optimization startup Inferize, reveals a vendor combining pricing discipline with software to extract margin from compute scarcity (items 7, 16). But the real beneficiary is unmistakable: anyone controlling long-term, zero-carbon power now commands the capital structure. Vistra, Broadcom's financing arm, and nuclear operators have inverted the supply chain. Chip shortage is tomorrow's problem only if power gets solved first.

The question now is not whether AI buildout can meet demand but whether power grids, nuclear regulators, and 20-year PPAs can coordinate at the required scale. Anthropic's $518 billion in commitments assumes they will. If they do not, the first constraint breakers will be the most power-exposed: hyperscalers with locked-in long-term demand and no second site. Watch Vistra, GE Vernova, and grid operators for the next 90 days—that is where the real capital gates now sit.