Power, Not Chips, Is Now the Binding Constraint

The $600B AI infrastructure capex cycle has hit a physical wall—energy scarcity is reshaping capital flows, forcing hyperscalers to own power generation and frontier labs to lock in multi-year dedicated capacity.

The data is unambiguous: 1,800 pending projects requesting 474 gigawatts now queue for approval in Texas alone—five times historical peak demand and enough to freeze 20% of the US data center pipeline. Texas Governor Abbott's grid connection pause is not a political stunt; it's a rational response to supply rationing. Duke Energy reports 7.6 GW of new electric service agreements with data center firms (2.7 GW added since Q4 2025), yet even that ramps slower than demand. Morgan Stanley's $2.3 trillion hyperscaler capex backlog ($1.4 trillion AI-specific) cannot execute if the grids that host it cannot supply electrons. Energy has displaced compute as the binding constraint, and capital is reallocating accordingly.

Hyperscalers are integrating vertically: Nvidia committed $3 billion in Lancium, a Texas power developer, to secure long-term energy supply rather than rely on utilities. SpaceX committed 10 GW of compute capacity by 2027 backed by power contracts. Anthropic signed a $10 billion infrastructure deal with Volta Infra and a $19 billion lease with TeraWulf, securing dedicated capacity with implicit power hedges. These are not traditional software-company moves—they are infrastructure operators protecting supply chains. The BitTorrent-miner-to-neocloud transition (TeraWulf, CoreWeave) is validated; idled mining rigs in power-rich regions are now premium real estate.

Geographic arbitrage is no longer a luxury—it is survival. The Texas pause forces projects to migrate east (Duke Energy's Carolinas capacity), south (Kentucky's $100 billion DOE-backed AI infrastructure project), and internationally (CoreWeave in Indonesia, Anthropic in Norway). This diffuses capital across regions with different power profiles and regulatory regimes, raising deployment timelines by 12-18 months and making regional power costs the new leasing multiplier. Frontier labs securing long-term deals are insulating themselves; infrastructure latecomers will face constrained supply and cost creep.

Chip supply remains critical—SK Hynix's $38 billion HBM investment locks Korea's tier-1 role—but it no longer gates buildout. The supply chain has inverted: power availability now gates chip placement; chip cost buys you processor share, but power scarcity buys you nothing if the grid cannot support it. Nvidia's pivot to power generation (Lancium) and SpaceX's $18.4 billion quarterly capex spike signal that mega-labs have internalized this. The hyperscaler capex cycle is no longer elastic on margin pressure; it is now constrained by regional power infrastructure and long-term energy contracting.

Watch three signals: whether Abbott's freeze propagates to other Republican-led energy-rich states (Arizona, Oklahoma), creating sidelined capacity; whether government-backed projects (Kentucky's $100B) move from R&D to operational supply; and whether frontier AI labs outbid hyperscalers for long-term power contracts—Anthropic's $19B TeraWulf deal suggests they are already winning niche markets. The winner will be whoever controls power generation in high-demand regions by year-end 2026, not whoever has the newest GPU die. Capital is repricing accordingly.