Meta plans to spend as much as $145 billion on AI infrastructure this year. Now it is looking at a way to rent out the computing power it does not use.
Bloomberg reported on Wednesday that Meta is developing a cloud infrastructure business through an internal effort called Meta Compute, aimed at turning spare AI capacity into revenue. Citing people familiar with the work, Bloomberg said Meta's shares rose 8 percent after the report. For a company that has long treated its data centers as internal plumbing for Facebook, Instagram and WhatsApp, that is a real change in posture.
The plan is still being shaped, not launched. According to Bloomberg, Meta is considering two routes. One would sell developers access to AI models running on Meta's own chips, closer to the way Amazon Web Services sells model access through Bedrock. The other would look more like CoreWeave or Lambda, where customers rent raw GPU capacity and build more of the stack themselves. Meta has not picked one. That matters less than the fact that it is preparing for both.
You do not build a resale business unless you think some of your supply may sit idle. Meta's own spending explains why that is possible. In April, the company raised its 2026 capital expenditure forecast to $125 billion to $145 billion, up from $115 billion to $135 billion, a move MarketWatch said sent the stock down more than 6 percent after hours. Investors were not confused. They saw the bill.
At Meta's annual shareholder meeting on May 27, Mark Zuckerberg gave them the answer he wanted them to hear. As TechRadar reported from the meeting, Zuckerberg said a cloud business was \"definitely on the table\" and added that outside companies ask almost every week whether Meta can sell API access or compute \"at some premium\" to what it paid. At the time, that sounded like a useful line for nervous shareholders. Five weeks later, it reads more like a product strategy being pulled into the open.
The buildout behind this is huge and very physical. Meta's Hyperion campus in Richland Parish, Louisiana, is expected to draw gigawatts of power, and Bloomberg reported in March that Meta would fund additional natural gas plants through Entergy to support the site. Bloomberg also reported on June 24 that the largest cloud companies now have more than $850 billion in future data center lease commitments, with Meta's future lease obligations rising to $182.9 billion as of March 31. These are not abstract AI bets. They are land, power contracts, chips and leases that have to earn their keep.
Here is the awkward part. Meta is not only a potential seller of AI capacity. It is also a buyer. The Wall Street Journal reported in April that CoreWeave expanded its long-term AI cloud infrastructure agreement with Meta to about $21 billion through December 2032, building on an earlier $14.2 billion deal. If Meta starts selling raw GPU access, it will be competing with the same kind of provider it has been paying to help close its own capacity gap.
Frankly, that is why the stock reaction tells you more than the branding. Wall Street has spent much of 2026 asking whether hyperscaler AI spending can ever justify itself. Oracle's data center obligations, Microsoft's capacity race and Google's compute constraints have all fed the same worry: the industry is building faster than anyone can prove demand will arrive on schedule. Idle GPUs are a cost. Rented GPUs are a revenue line. Investors liked the second sentence much more than the first.
The timing also cuts against the idea that AI capacity is anywhere close to abundant. The Financial Times reported this week that Google capped Meta's use of Gemini after Meta sought more capacity than Google could provide, delaying some internal AI projects and pushing staff to use tokens more carefully. Think about that for a second. Meta may soon rent out compute while still buying and rationing it elsewhere. That is not contradiction. It is what an infrastructure race looks like when demand comes in uneven waves.
AWS, Azure and Google Cloud should still not panic. Meta does not have their enterprise sales machines, partner networks or decades of customer trust in cloud operations. But you should not dismiss the pressure either. Neoclouds like CoreWeave and Lambda were built for a market where GPU scarcity gave specialists room to charge. A Meta cloud product, even a narrow one, would enter with a balance sheet and chip pipeline most specialists cannot match.
What nobody has confirmed is pricing, launch timing or which customers Meta would target first. Bloomberg described Meta Compute as an initiative still taking shape, not a finished product. Zuckerberg has also said Meta has not rented out capacity so far because it believes it can use what it is building. The real test is whether Llama training, ad-ranking systems and business AI tools on WhatsApp and Instagram can absorb enough of the infrastructure before shareholders demand a clearer return.
Either way, Meta has changed the story it is telling. For years, its data centers existed to feed its own apps. Now the same buildings may serve someone else's AI workload too, if the price is right.
Also read: 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 • Anthropic Gets Its Most Powerful AI Models Back After an 18-Day Fight With Washington
