Bloomberg reported on July 1 that Meta is standing up a cloud business, internally dubbed Meta Compute, that will do two things Meta has spent a decade insisting it did not do. It will rent GPU capacity in the CoreWeave and neocloud style. And it will let developers call hosted Meta models over an API, including the newly closed-weight Muse Spark, in a setup that looks a lot like AWS Bedrock with the serial numbers filed off. Leadership is a Superintelligence Labs slate: infrastructure head Santosh Janardhan, ex-Y Combinator partner Daniel Gross, and president Dina Powell McCormick.
The market reaction was clean. Meta shares climbed about 9 percent on the news. CoreWeave, whose entire business is renting GPUs to hyperscalers and enterprises, closed down about 14 percent. When the customer becomes the competitor, the multiple compresses fast.
The strategic logic is not subtle. Meta has committed roughly $182.9 billion to AI infrastructure and has been noticeably quiet about how much of it is being used by its own consumer AI products. Meta AI and Llama do not report separate revenue. The company built like an OpenAI and monetized like an ad network. Meta Compute is the answer to “who else can we sell this to before the next earnings call.”
The move also arrives a few weeks after SpaceX and xAI announced their own capacity-resale program, with anchor customers reportedly including Anthropic, Google, and Reflection AI. The pattern is now visible enough to name. Everyone with more GPUs than they can immediately use is turning the surplus into a product line, and the pure-play neoclouds are watching their moat get filled in by the same hyperscalers who used to be their tenants.