Meta is reportedly talking about renting roughly $10 billion of AI computing power to Anthropic, which tells you how strange the AI infrastructure race has become. The company that wants to beat Claude may also help run it.
Meta and Anthropic are reportedly negotiating a deal that would have sounded absurd not long ago. Mark Zuckerberg's company would rent slices of its AI computing capacity to one of the labs racing hardest against it. Rival first. Landlord second. According to SemiAnalysis analyst Jeremie Eliahou Ontiveros, the arrangement could be worth roughly $10 billion and would include 90-day cancellation options for either side, similar to structures the firm says SpaceX has already signed with Anthropic and Google. Nothing is signed. The talks are early, and neither Meta nor Anthropic has confirmed them publicly.
The point isn't only the number. It is what the number says about Meta's bet. SemiAnalysis reported that the deal would give Meta private instances of Claude, in the same broad style that Amazon offers Anthropic models through Bedrock. At the same time, Meta's own researchers are building Llama and the next generation of models inside Superintelligence Labs. You can see the awkwardness. Meta would be paying to build infrastructure for its own AI ambitions while also letting Anthropic use servers Meta built for itself.
That is not a small change in posture for a company that spent years talking as if Llama would define its AI future. Infrastructure is becoming the harder asset to replicate. If you own enough power, chips and data center capacity, you don't need every model to win. You just need to make money from the race.
In January, Zuckerberg announced Meta Compute, a top-level initiative to manage how the company engineers and funds its data center partnerships. Two names run it. The Information reported that Santosh Janardhan, who runs Meta's global data center and network fleet, was put in charge, alongside Daniel Gross, the former Safe Superintelligence CEO. Meta also named Dina Powell McCormick as president and vice chairman that month. She would help manage the physical and financial model behind its next decade of computing, the company said.
The plan is plain enough. Rent out what Meta isn't using.
According to SemiAnalysis, Meta contracted more than 5 gigawatts of cloud and colocation capacity in the first six months of the year, on top of its own data center construction. The firm estimated Meta could bring in more than $10 billion a year by allocating 200 megawatts to outside customers, using roughly $50 billion per gigawatt per year as the pricing benchmark set by SpaceX. That is a brutal little piece of math. A modest slice of spare capacity, if the pricing holds, becomes a new business line.
The landlords are also rivals
A year ago, an AI lab handing compute to a direct competitor would have looked like a strategic mistake. Now it is normal enough to price. SemiAnalysis said Anthropic already rents the whole of Colossus 1, xAI's 220,000-GPU supercomputer in Memphis, Tennessee, in a deal worth about $1.25 billion a month and more than $40 billion through 2029. The same report said Google pays SpaceX roughly $920 million a month for 110,000 GPUs at the same site through June 2029. Anthropic has also signed a $19 billion, 20-year lease with TeraWulf for up to 401 megawatts at a data center campus in Hawesville, Kentucky.
Those figures look wild until you follow the pressure underneath them. Claude needs capacity. Gemini needs capacity. Meta needs capacity. OpenAI needs capacity. Everyone says the bottleneck is chips, power and buildings, so the companies with buildings and power are learning they can charge like landlords in a housing shortage.
Meta reportedly hasn't settled on the exact product. SemiAnalysis said the company is weighing whether to sell developers access to AI models running on its own servers, closer to Amazon Bedrock, or simply sell raw computing capacity the way neocloud operators such as CoreWeave do. Either route pushes Meta into a market where AWS, Google Cloud and Azure already know how to sell enterprise infrastructure. That isn't a side quest. It is a new business with different customers, different contracts, different pricing and different expectations.
Here's the thing. If Meta has enough spare capacity to earn $10 billion a year from Anthropic, then either it is building far more than its own models currently need, or it has decided the infrastructure layer may be more durable than the model leaderboard. Tom's Hardware noted that reports of Meta renting out compute hit AI infrastructure stocks, as investors worried the market could be moving from shortage to glut faster than expected.
That fear is reasonable. A company can overbuild data centers just as surely as it can overbuild offices. The difference is that the AI industry has taught itself to treat every gigawatt as proof of inevitability, when it may only be proof that very rich companies are willing to sign very large contracts.
For Meta, the Anthropic talks are still only talks. SemiAnalysis is clear about that, and the companies haven't put their names on a public announcement. But the direction is hard to miss. The AI labs with the most chips are becoming each other's suppliers, and Meta, after years of saying it would out-build everyone, now has to decide whether collecting rent is more valuable than keeping every rack for itself.
Also read: ASML Gives Nearly Every Worker 20,000 Euros in Stock to Keep Them Through 2030 • Kimi K3 Shows the Cheap Chinese AI Era Is Coming to an End • Upper90 Lends $400 Million Against AI Inference Chips Instead of Nvidia GPUs
