Hyperscaler Vertical Integration Crushes the Neocloud Thesis
The neocloud thesis inverted when Meta announced it would sell excess GPU compute directly to customers. TeraWulf's $19 billion twenty-year lock-in with Anthropic validated the original thesis: hyperscalers buy, model providers grow, independent operators capture margin. But hyperscaler vertical integration just closed that moat. CoreWeave dropped 14% in a single day, Nebius 17%—not because demand disappeared, but because the addressable market for independent operators just halved into hyperscaler surplus sales.
The capex cycle is genuinely constraint-driven, not speculative. Alphabet secured $80 billion from Berkshire Hathaway; Meta committed $50 billion to Louisiana plus $13 billion to Canada; SK Hynix's record $26.5 billion raise is explicitly tied to memory shortages through 2030. But scale is a weapon. CoreWeave's $24.9 billion debt deal with circular Nvidia-customer lending exemplifies the trap: neocloud providers are debt-financed marginal players racing against hyperscalers with captive demand and captive balance sheets. When Anthropic buys from TeraWulf's dedicated capacity and others buy from Meta's excess sales, independent operators lose pricing power and addressable market.
Geography and governance now favor hyperscaler build. Meta's Canadian expansion locked stable power and favorable regulation; site selection has become competitive moat. Inference plays like SambaNova (raised $1 billion at $11 billion valuation) face margin compression as hyperscalers capture workloads. The BIS warned this week that AI investment is a "debt-driven boom" mirroring past tech bubbles—and neocloud carries the highest leverage of all. Clear winners: hyperscaler lock-in deals, memory suppliers (SK Hynix), debt lenders betting captive infrastructure as durable risk. Clear losers: independent compute and alternative-chip providers facing a closing window to differentiate.
Watch whether CoreWeave's $20 billion capital raise can execute without anchor-customer defection to hyperscaler captive offerings. Meta's compute-sales channel is either genuinely incremental revenue or margin cannibalization rebranded. If model-provider unit economics soften, the $80+ billion capex cycle and its debt-financed structure collapse.