Meta Eyes GPU Sales, Pivots to AI Compute Infrastructure

量子位 · wechat · 2026-07-05

According to Bloomberg and SemiAnalysis, Meta is considering launching MetaCompute to open its massive AI compute to external clients. It is reportedly in final talks with Anthropic to secure private instances of Claude and build a model service platform similar to Bedrock. In the first six months of this year, Meta secured over 5GW of data center capacity, with nearly 10GW in related deals since early 2024, and raised its 2026 capital expenditure guidance to between 1250亿 and 1450亿 USD. The economics of compute leasing are highly lucrative (around 500亿 USD revenue per GW annually), prompting Meta's stock to surge nearly 9% and causing a sell-off in AI cloud providers like CoreWeave and Nebius. This shift comes as Meta's in-house model progress falls short of expectations, Gemini usage is restricted, and internal AI agent initiatives advance slowly.

The compute will be allocated across four areas: continuing to train in-house models (MuseSpark and the next-generation Watermelon), scaling the complexity of ad recommendation systems, leasing to external clients at a premium, and hosting third-party models like Claude. If the compute business materializes, Meta will directly compete with AWS, Azure, Google Cloud, and AI cloud providers like CoreWeave and Nebius.

Original post →

More from Companies & People

Companies & People channel →