Meta is building a cloud business to rent out its surplus AI compute

Meta spent the past two years buying land, power, and chips to train AI. Now it wants to rent some of that hardware out.

According to a Bloomberg report published July 1, Meta is developing a cloud infrastructure business that would sell outside customers access to its AI compute and, potentially, its models. The effort runs through an internal group called Meta Compute, led by infrastructure chief Santosh Janardhan, Meta Superintelligence Labs leader Daniel Gross, and company president Dina Powell McCormick. Meta had not commented publicly when the report ran, though Mark Zuckerberg had already left the door open: in May he said a cloud business was "definitely on the table."

What Meta would actually sell

Two things, if the reporting holds.

The first is raw compute: bare access to GPUs and the data-center plumbing around them, billed by the hour or the reservation. That is roughly CoreWeave's business, and it is the easy thing to sell because it asks nothing of the customer except a workload. Meta already owns the hardware. The product is the capacity it is not currently using.

The second is model access, closer to what Amazon does with Bedrock: hosted endpoints to run models on Meta's infrastructure, including Muse Spark, the closed-weight model Meta released recently. That carries higher margins and keeps customers around longer, but it puts Meta in direct competition with the same clouds it would be renting bare iron to.

Either version rests on one fact about data centers: idle ones depreciate whether or not anything runs on them. Selling the slack is a way to make expensive, half-used hardware earn its keep.

The number behind it

As of the end of the first quarter, Meta had committed to about $182.9 billion in AI infrastructure spending, per its own SEC filing. That covers sites including a campus in Louisiana and one in Ohio that Zuckerberg has said would roughly match the footprint of Manhattan.

Set against that figure, the cloud plan reads as much like an answer to a revenue problem as an ambition. Meta does not break out revenue from Meta AI or its Llama models, and its executives mostly describe internal uses. In plain terms, its AI spending does not yet have a standalone revenue line to point at. Renting compute puts a number next to the capex.

Meta is not first

xAI got there a few weeks earlier. In early May, SpaceX bought out the entire capacity of its Colossus 1 data center and began leasing it, first to Anthropic, then to Google and Reflection AI. The pattern behind both moves is a bet on ownership: the companies that pour the concrete and buy the chips may capture more durable value than the ones shipping the best models.

That bet holds under two conditions. Demand for compute has to keep climbing, and the hardware has to retain enough value to earn back its cost. Both are contested. AI accelerators lose value quickly, and plenty of people have argued the infrastructure race is running well ahead of the end-user revenue that would justify it. Leasing surplus capacity looks shrewd if the demand is real and looks like a clearance sale if it is not.

What to watch

This is a reported plan, not a shipping product. A few things would show it is real: a named service with public pricing, a first external customer that is not already a Meta partner, and whether Meta sells only raw compute or actually opens Muse Spark to outside developers. Until then it is a strategy leak with an org chart attached. Worth noting anyway, because it is the second time in two months that a company which built data centers for itself has decided to become a landlord.