Meta plans cloud business to rent excess AI compute, targeting AWS and neoclouds
Meta is building a cloud infrastructure business to monetize excess AI computing capacity it built ahead of demand, Bloomberg reported Wednesday. The initiative, internally dubbed Meta Compute and led by infrastructure chief Santosh Janardhan, Daniel Gross of Meta Superintelligence Labs, and president Dina Powell McCormick, will compete directly with Amazon Web Services, Microsoft Azure, and Google Cloud—putting Meta in the same position as SpaceX/xAI, which is already leasing data-center capacity to Anthropic and Google.
The company is weighing two service models: selling access to hosted AI models including its own Muse Spark foundation model (similar to AWS Bedrock), or selling raw GPU compute like neocloud providers CoreWeave and Nebius. Meta committed $125–$145 billion in capex for 2026, up from prior guidance, and has signed contracts worth $48 billion for GPU rental from CoreWeave and Nebius—giving it leverage to supply as well as demand.
The announcement moved markets immediately: Meta stock rose 8.6–9% on the news, while CoreWeave dropped 10.8% and Nebius fell 12.4%, signaling investor concern that Meta could displace neocloud intermediaries. CEO Mark Zuckerberg confirmed in May that cloud services were "definitely on the table" if the company overbuild capacity—a hedge against the risk of captive-compute waste.
For architects, this marks a structural shift: hyperscalers are no longer just buyers of GPU capacity; they are becoming suppliers. This signals both easing GPU spot-price pressure over time and fragmentation of the compute market along hyperscaler-loyalty lines. How Meta prices and allocates capacity—and whether it prioritizes raw GPU or hosted models—will reshape inference economics for the next 18 months.