Hyperscaler Consolidation Kills Neocloud Economics

Nebius lost $120B in market cap in 24 hours. CoreWeave and IREN followed. This is price discovery, not panic. When Meta announced Meta Compute, the market suddenly priced what was always true: the neocloud category exists only because hyperscalers weren't willing to sell excess capacity at margin compression. That constraint dissolved. These companies were never sustainable independent operators—they were temporary fill-ins for a capacity shortage that hyperscalers are now solving themselves. Margin compression is immediate and structural.

Power is the new infrastructure moat. Brookfield's $25B Bloom Energy fuel-cell partnership and National Grid Ventures' $1.75B Texas gas plant for Microsoft reveal the binding constraint: not GPUs, but generation capacity. OpenAI's 10GW Ohio lease and Meta's multi-gigawatt campus commits signal that hyperscalers are now acquiring power and land as core capex. SoftBank's $500B Stargate consortium is effectively a land-plus-power-plus-compute platform. Whoever controls generation capacity (fuel cells, stranded gas, nuclear, renewables) controls the AI infrastructure layer for the next decade. Independent operators without captive power are structurally exposed.

China's Meituan trained a 1.6T model on 50,000 domestically sourced chips, proving export controls didn't freeze competitive model training—they raised cost and timelines. Domestic silicon scales. Nvidia still owns volume and margin, but the US no longer monopolizes large-model training. The supply geography is resetting.

Structural winners: integrated hyperscalers (Meta, OpenAI, SoftBank, Oracle) and power infrastructure (Brookfield). Crusoe Energy's $30B valuation is the rare alternative—power-efficient hardware instead of capacity sales. Together AI's $800M Series C at $8.3B validates the open-model niche, but this is secondary consolidation at lower growth rates. Tier-two operators are now subordinate.

Watch three signals: (1) Meta Compute adoption rates—if cheaper capacity drives growth, cloud margins compress industry-wide and unit economics get ugly. (2) Whether power-cost pressure and US grid stress force workloads geographically toward fuel-cell or nuclear zones. (3) Whether Meituan and ByteDance's chip scaling reduces Chinese model training cost below US hyperscaler marginal cost, triggering supply-chain realignment toward domestic silicon independent of Nvidia.