Power, Not Compute, Is Now the AI Infrastructure Moat

Hyperscalers and power providers have won the GPU cloud wars by locking up electricity supplies before capacity constraints; neoclouds without power partnerships face extinction as Meta commoditizes excess compute.

The last 18 months proved GPU scarcity wasn't structural—it was organizational. Anthropic's $19 billion TeraWulf deal, Brookfield-Bloom's $25 billion fuel-cell partnership, and Meta's Canadian gigawatt facility all signal the same truth: the binding constraint has shifted. Hyperscalers aren't competing for silicon anymore; they're bidding against each other for kilowatt-hours. SoftBank's surprise SB Neo entry into neoclouds—explicitly positioned against power-starved incumbents—confirms what the market just repriced: electricity is the bottleneck, not GPUs.

CoreWeave, Nebius, and IREN fell 14–17% in a single day because a pure-play neocloud selling compute from the grid is not differentiated from Meta selling excess H100 clusters at scale. Meta can afford margin compression; CoreWeave cannot. Crusoe's $30 billion valuation against the 50% haircuts on peers proves the thesis: neoclouds with a power angle (Crusoe's stranded fuel infrastructure) or bespoke inference-only bets (SambaNova's $11 billion with JPMorgan adoption) survive. Generic GPU clouds face slow extinction.

The $25 billion Bloom-Brookfield partnership and Anthropic-TeraWulf's $19 billion are the largest infrastructure deals in AI history—and both are power-first. Fuel cells are now table stakes for hyperscale deployment. This isn't a luxury; it's structural requirement. Grid power is exhausted in Virginia, Texas, and California. Dedicated power—fuel cells, dedicated renewable, on-site generation—is the new moat. Companies controlling electricity (TeraWulf, Bloom, Brookfield, nuclear bidders) will capture outsized returns.

Meituan's 1.6-trillion-parameter LLM trained on 50,000 domestically sourced Chinese chips proves the bifurcation is real and irreversible. China is validating semiconductor alternatives at competitive scale. This reshapes global pricing: if China fields competitive models independently, US hyperscalers cannot sustain premium pricing globally. The $30 billion annual capex flooding AI datacenters will split between US frontier labs bound by power scarcity and Chinese players bound by export controls but fully resourced.

Neoclouds dependent on grid power in congested regions face slow extinction. Watch SoftBank consolidate weakened players. Watch hyperscalers stop competing on compute and start competing on power procurement—who signs the next $20 billion fuel-cell deal, who gets nuclear plants online first, who captures AI-friendly jurisdictions (Canada, Ireland, Nordic). The winner will own the electrical engineering, not the networking.