Meta spent $75B in 3 months on AI infrastructure (CoreWeave, Oracle, Blue Owl) (allenarch.dev)

Between Sept. 30 and Oct. 27, 2025 Meta unveiled roughly $75.5 billion in infrastructure commitments — pushing its Q3 capex to 37% of revenue — via four headline deals: CoreWeave ($14.2B), Oracle (~$20B), Scale AI (49% stake, $14.3B) and a $27B joint venture with Blue Owl called Hyperion. Hyperion’s financing is notable: Blue Owl holds 80% vs Meta 20%, Morgan Stanley arranged $27B+ of debt, PIMCO anchored 144A bonds maturing 2049, Meta took a $3B cash distribution but is on the hook through a 16‑year residual value guarantee. The builds are raw compute and power at scale — Hyperion is a 2GW, 2,250-acre campus (operational by 2030) able to host hundreds of thousands of GPUs and run GPT‑4–class training repeatedly; CoreWeave supplies Nvidia GB300/Blackwell racks (72 GPUs/rack, ~2.5× H100 performance) under long-term contracts; ENGIE deals add 1.3GW of solar for cheaper, steadier power. For AI/ML, the story is both strategic and structural. Meta is treating infrastructure as a competitive moat and is accelerating scale by shifting from outright ownership to JV/debt-financed “infrastructure-as-a-service,” keeping assets off the balance sheet and speeding deployment—but creating 16‑year fixed obligations and higher leverage if AI revenue lags. Technically, the commitments unlock massive training throughput, lower latency and more efficient FP4 inference, but the economics are unsettled: current AI revenue hasn’t caught up to the spending, raising questions about timing of returns, circular vendor financing, and systemic macro exposure even as compute demand and product use cases continue to grow.