Here is a sentence that would have read as a typo a year ago: Meta, the company that builds its own Llama models, is in talks to rent computing power to Anthropic, the company that builds Claude. According to reporting first surfaced by The New York Times, the two are discussing a lease of up to $10 billion of AI compute over two years, roughly $417 million a month, with either side free to walk away early. Nothing is signed, and the talks are described as preliminary. But the detail that matters most is who asked. Anthropic went to Meta, back in June, proposing the arrangement. This was not Meta shopping spare racks to a desperate buyer. It was a frontier lab knocking on a rival's door because it had run out of other doors.
The reason is a shortage that has become the defining constraint of the industry. Anthropic's share of web traffic nearly doubled between March and June, and every new user means more inference, more training, and more chips burned each day. The company has been capping usage of its most capable Fable models simply because there are not enough accelerators to serve demand. When a lab throttles its flagship product to ration hardware, that is not a planning slip. It is a structural bottleneck, and it explains why Anthropic has spent the year stacking compute commitments from anyone with racks to spare: a reported $45 billion deal with Elon Musk's SpaceX for its Colossus supercomputers, a multi-billion arrangement with Google for TPU capacity, and smaller pacts with Fluidstack and TeraWulf. Meta is simply the most surprising name on the list.
For Meta, the surprise is the point. Mark Zuckerberg told shareholders in May that entering cloud computing was "definitely on the table," and that companies were approaching him "almost every week" to buy access to spare capacity. The talks reportedly sit inside a new unit, Meta Compute, staffed with a serious bench: infrastructure chief Santosh Janardhan, Superintelligence Labs leader Daniel Gross, and president Dina Powell McCormick, with an 18-year Amazon Web Services veteran, Dave Brown, hired to help build the business. Meta has already struck smaller compute deals with the neoclouds CoreWeave and Nebius. Landing Anthropic would be a different order of thing, less for the money than the symbolism: one frontier lab renting the ground its competitor stands on.
The economics make the strange logical. Meta expects to spend as much as $145 billion on capital projects this year, more than double last year's outlay, and even that is apparently more concrete and silicon than its own roadmap can immediately absorb. Selling the surplus turns a cost centre into a revenue line, and helps justify to shareholders the same buildout that saw the company cut 8,000 jobs in May to fund it. Across Big Tech, AI spending is projected to reach $725 billion this year, and still the labs are short. Ten billion dollars, in that light, is the small deal. It is a third the size of the SpaceX pact.
What lingers is the shape of the market this reveals. Compute has become more valuable than the models it runs, the one resource no lab can conjure fast enough. In an industry where yesterday's rival is today's landlord, the winners may not be the labs with the best system but the ones that own the ground underneath everyone else's. Whoever collects the rent does well whichever model wins.