Meta Is Quietly Building a Business to Sell Its Leftover AI Computing Power

Meta is developing a cloud infrastructure business called Meta Compute to rent out unused AI computing power, potentially competing with AWS, CoreWeave, and Lambda. The initiative, still in early stages, could turn idle GPUs into revenue and sent Meta's stock up 8% on the news. Meta's massive AI spending—up to $145 billion in 2026—makes a resale business a strategic hedge against overbuilding.

Meta spent as much as $145 billion building AI infrastructure this year. Now it wants to rent out what it doesn't use. Bloomberg reported on Wednesday that Meta is developing a cloud infrastructure business through an internal initiative called Meta Compute, aimed at turning unused computing power into revenue. The report, sourced to people familiar with the matter, sent Meta's stock up 8 percent that day. It's a striking pivot for a company that has spent two decades insisting its data centers exist only to serve its own two billion users. According to Bloomberg, Meta is weighing two distinct paths. One mirrors Amazon Web Services' Bedrock: selling developers access to AI models running on Meta's own chips, billed by usage rather than by the hour of hardware. The other looks more like CoreWeave or Lambda, the neocloud firms that lease raw GPU capacity with none of the managed layer AWS or Azure wrap around it. Meta hasn't picked a lane. Bloomberg's sources describe the plans as still forming, not finalized. None of this arrived out of nowhere. At Meta's annual shareholder meeting on May 27, Mark Zuckerberg told investors a cloud business was "definitely on the table," adding that outside companies routinely ask "if we have compute that they could buy from us at some premium to what we've bought it at." At the time it read as a hedge, a contingency in case Meta overbuilt. Five weeks later, Bloomberg's report suggests that hedge has become an active plan. You don't stand up a resale business unless you expect leftover supply. Meta's own numbers explain why leftovers look likely. The company's AI-related capital spending for 2026 is projected at $125 billion to $145 billion, funding buildouts like the Hyperion campus in Louisiana and the Prometheus site in Ohio. Bloomberg separately reported on June 24 that Meta and Microsoft are leading an $850 billion boom in data center leasing across the industry. Meta has also signed a $14.2 billion compute deal with CoreWeave, buying capacity from the exact kind of neocloud it may soon compete against. That's the tension sitting underneath this story. Meta is simultaneously a customer of neocloud providers and, potentially, their rival. If it starts selling raw GPU access at a premium to what it paid, the way Zuckerberg described, it competes directly with CoreWeave for the same buyers CoreWeave built its business serving. Frankly, the more telling number here is the 8 percent stock jump, not the initiative itself. Investors have spent much of 2026 nervous about whether hyperscaler AI spending will ever pay for itself, a worry Oracle's own data-center debt load has done nothing to ease. A resale business changes that math. Idle GPUs are a cost. Rented GPUs are a revenue line. Wall Street rewarded Meta instantly for signaling it has a plan B if its own AI products don't absorb everything it's building. That's also the read for AWS, Azure and Google Cloud. All three have spent the last two years selling scarcity, telling customers that GPU capacity is tight and reservations book out months ahead. A Meta that shows up selling the same chips at a premium undercuts that scarcity story, even if Meta never approaches their scale. Neoclouds like CoreWeave and Lambda carry thinner margins and less capital than a company that just told shareholders it might spend $145 billion in a single year. If Meta enters as a supplier, it enters with a balance sheet none of them can match. What nobody has confirmed is pricing, target customers or a launch date. Bloomberg's sources describe an initiative still taking shape inside Meta Compute, not a product with a release window. Zuckerberg has said Meta hasn't rented out capacity so far because it still believes it has a use for what it's building. Whether that changes depends on how much of Meta's own roadmap, from Llama training runs to ad-ranking models to the AI agents it recently began selling to businesses on WhatsApp and Instagram, actually absorbs the compute Meta keeps stacking up. Either way, the direction is set. Meta spent years telling Wall Street its data centers existed to serve Facebook, Instagram and WhatsApp users. Now it's telling them the same buildings might serve its competitors' customers too, for a fee. Also read: Twelve Labs raises $100 million as Amazon bets its Trainium chips on video AI https://startupfortune.com/twelve-labs-raises-100-million-as-amazon-bets-its-trainium-chips-on-video-ai/ • Oracle Just Handed Wall Street Its Best Case Against the AI Spending Boom https://startupfortune.com/oracle-just-handed-wall-street-its-best-case-against-the-ai-spending-boom/ • Anthropic Gets Its Most Powerful AI Models Back After an 18-Day Fight With Washington https://startupfortune.com/anthropic-gets-its-most-powerful-ai-models-back-after-an-18-day-fight-with-washington/