Meta's Entry Collapses Neocloud Economics; Power Becomes the Bottleneck

Anthropic locks in $19 billion of capacity to escape hyperscaler control, but Meta's identical-week entry into compute clouds proves neoclouds can't compete—power, not silicon, is the actual bottleneck.

Anthropic's commitment to a $19 billion, 20-year capacity lease with TeraWulf in Kentucky reads as a declaration of independence from hyperscaler infrastructure monopolies. Frontier model laboratories are no longer comfortable betting on AWS, Azure, and GCP availability; they're locking in dedicated capacity at fixed rates to guarantee foundry access. The precedent matters: a model lab willing to stake $19 billion over two decades signals that compute access has become critical infrastructure, not a utility commodity. TeraWulf's pivot from Bitcoin mining to Anthropic partnership validates the entire neocloud thesis—dedicated providers can win by serving dedicated customers.

Then Meta announced it would commercialize excess GPU capacity and rent it to customers. Nebius lost $120 billion in market capitalization in a single trading day. CoreWeave dropped 17%. The neocloud premise collapsed in hours. Hyperscalers built so much capacity to serve their own AI ambitions that their marginal cost on spare GPUs undercuts every neocloud provider's fully-loaded economics. CoreWeave raised $20 billion on the assumption the market would value independent compute. That assumption expired the moment Meta showed it wouldn't.

Anthropic made the smarter bet anyway, but not for the reason it appears. Anthropic didn't lock in compute abundance—it locked in power. TeraWulf's site selection in Kentucky, Bloom Energy and Brookfield's $25 billion power infrastructure partnership, Meta's commitment to Alberta's 1-gigawatt grid access—all point to the same constraint. Electricity, not silicon, is now the binding resource. Compute will remain abundant as long as power supply is adequate; power is the real limit. Frontier labs understood this. Neoclouds, built on generic spot-market economics, did not.

Memory supply is locking in around the same pressure point. SK Hynix's record $26.5 billion IPO included a CEO warning: HBM shortage will persist until 2030. Memory scarcity forces AI companies to sign long-term contracts with memory suppliers, and long-term contracts require long-term infrastructure commitments. Anthropic's 20-year TeraWulf agreement hedges compute, power, and memory supply simultaneously. CoreWeave's funding assumed commodity pricing could hedge everything. The gap between these strategies defines the next three years.

The market is consolidating around a clear hierarchy. Dedicated partnerships win (Anthropic-TeraWulf). Hyperscalers with diversified revenue win (Meta Compute, Google's $80 billion Berkshire commitment). Power providers and memory makers win. Generic neocloud startups lose. CoreWeave and Nebius could pivot to become regional power-infrastructure partners, but their funding already priced in independence—a bet that just lost 14-17% of its value in hours.

Watch whether other frontier labs follow Anthropic's precedent with long-term capacity partnerships anchored in power, not compute. Watch whether Meta and other hyperscalers use spare-capacity pricing to accelerate neocloud consolidation. And watch whether memory makers can simultaneously raise HBM prices while locking in long-term customer contracts. The $10+ trillion infrastructure buildout is real. The neocloud business model is not.