Power Is Now the Primary Infrastructure Moat—Not Compute

Anthropic's $19 billion Terawulf commitment and Brookfield's $25 billion fuel-cell deal reveal the real constraint: grid electricity and dedicated power infrastructure have become the primary lever, not GPU supply.

This week's megadeals and valuations reframe the infrastructure hierarchy. Anthropic locked in 20 years of capacity at Terawulf—not a spot market transaction but a structural power-first arrangement. Simultaneously, Brookfield and Bloom Energy announced $25 billion for dedicated fuel-cell power targeting hyperscale AI clusters globally. Meta, despite building its own grid-constrained Alberta campus ($13B CAD for 1GW capacity), can now monetize "excess" compute—a SpaceX-style margin capture. The message is identical across all signals: power geography and supply contracts now determine who scales, not GPU availability.

The neocloud model is being validated and pressured simultaneously. CoreWeave's Texas campus came online at 133MW via Galaxy Energy; Crusoe tripled to $30 billion valuation on power-efficient demand; SambaNova closed Series F at $11 billion with JPMorgan as an inference partner. These are not signals of a collapsing sector—they're validation that neoclouds solve a structural problem. But the market's visceral reaction to Meta Compute (Nebius -$120B market cap, CoreWeave -48%, IREN down) exposes the real fear: hyperscalers' ability to undercut commodity capacity at margins. Meta didn't build a cloud; it proved that excess capacity at any giant can become a margin weapon. That distinction matters.

Geopolitically, Meituan's 1.6 trillion-parameter model on 50,000 domestic chips is a watershed. China can now field competitive frontier-scale models independently of US supply chains. This doesn't immediately threaten current US dominance, but it reshapes the compute supply geography and raises the stakes for export controls. Meanwhile, hyperscalers' geographic expansion—Meta to Alberta, implied grid-constrained plays elsewhere—signals they're building dual systems: their own dedicated capacity for core workloads, neocloud partnerships for overflow and specialized workloads (inference, power-efficient, geographically diverse).

The week's concrete math points to a winner: power-anchored neoclouds with long-term model-provider contracts (Terawulf-Anthropic, implied CoreWeave relationships). The exposed play is commodity GPU clouds and publicly traded neoclouds trading on growth without moat differentiation. Crusoe's $30 billion valuation isn't a bubble—it's a signal that power efficiency is now a defensible moat. Grid-constrained siting, fuel-cell deals, and 20-year capacity leases are how you compete in 2026. GPU prices have become a commodity input; electricity and physical site scarcity are the structural limits.

Next watch: Which frontier labs lock in long-term power contracts (Anthropic did; OpenAI and xAI likely next). Brookfield and Bloom's $25 billion is a dry run for a much larger shift—expect hyperscalers to join or countersignals with their own power partnerships. CoreWeave and Crusoe will escape the public market's fear if they can disclose their anchor contracts. Meta's Alberta capacity will signal whether hyperscalers genuinely treat neoclouds as partners or redundancy. Finally: does China's model-training independence accelerate US export controls or lead to price wars in commodity inference (where China can compete at lower margin)?