Meta turns its idle GPUs into a cloud business

Meta spent $125 to $145 billion on AI infrastructure in 2026. Wall Street kept asking where the money comes back. On July 1, Bloomberg reported the answer: Meta is building a cloud business.

The internal project, called Meta Compute, would sell excess GPU capacity and hosted AI model access to outside customers. It is led by infrastructure chief Santosh Janardhan, Daniel Gross from Meta Superintelligence Labs, and company president Dina Powell McCormick. Two business models are on the table. The first mirrors CoreWeave: sell raw compute by the GPU-hour. The second resembles Amazon Bedrock: charge per API call for access to Meta’s Muse Spark models, with Meta running the infrastructure underneath.

The market reacted fast. Meta stock jumped 8 percent, adding roughly $120 billion in paper value for a service with no paying customers yet. The other side of the trade was brutal. CoreWeave fell 13 percent. Nebius dropped 15 percent. Nvidia slipped on the logic that if Meta sells compute instead of buying more of it, one of the world’s largest GPU buyers just signaled weaker future demand.

Here is the part that makes no sense on paper. Meta has $35 billion committed to CoreWeave through December 2032. It signed a $27 billion deal with Nebius just months ago. Both were gap-filler contracts to cover capacity while Meta’s own data centers came online. Those agreements do not unwind because Meta is now building competing infrastructure. But the optics of committing nearly $62 billion to companies you are about to position as rivals, within months of signing the deals, is something.

The infrastructure behind this is unlike anything else in corporate history. The Prometheus campus in New Albany, Ohio is a 1 gigawatt facility powered by nuclear energy, scheduled to come online later this year. The Hyperion campus in Louisiana will eventually reach 5 gigawatts, covering what Zuckerberg described as a significant footprint of Manhattan, with the first 2 gigawatt phase arriving by 2030. Meta has also been erecting tent-based rapid-deployment structures across the US since April. Each one is about 125,000 square feet and cooled by jet engines.

That scale creates the excess the cloud business would sell. When a large training run finishes, tens of thousands of GPUs sit underutilized until the next one starts. A cloud business converts that idle time into revenue rather than a dead-weight capital expense.

Meta is not the first to try this. xAI already leases capacity from its Colossus cluster to Anthropic, Google, and Reflection AI between training runs. SpaceX sells excess satellite bandwidth. The playbook exists.

What Meta does not have is twenty years of enterprise cloud infrastructure. AWS has been building since 2006. Azure and Google Cloud have spent years assembling support teams, compliance frameworks, SLA guarantees, and enterprise sales organizations. Meta is starting from zero on all of that. Building data center capacity and operating a commercially competitive cloud product are not the same thing.

Zuckerberg told investors at the May shareholder call that selling excess compute was “definitely on the table.” The table just got a lot bigger.