Meta Is Reportedly Negotiating to Rent Its AI Chips to Rival Anthropic

Meta is reportedly negotiating to lease roughly $10 billion of its AI computing power to rival Anthropic, a deal that would make Mark Zuckerberg's company a landlord for one of its fiercest competitors. The arrangement, which includes 90-day cancellation options, reflects a broader industry shift where AI labs rent compute to rivals. The talks are early and unconfirmed, but could turn Meta into a vendor competing with AWS, Google Cloud, and Azure.

Meta is reportedly negotiating to lease roughly $10 billion of its AI computing power to Anthropic, a deal that would make Mark Zuckerberg's company a landlord for one of its fiercest rivals. Meta and Anthropic are reportedly negotiating a deal that would have sounded absurd eighteen months ago. Zuckerberg's company would rent out slices of its own AI computing power to one of the labs racing hardest against it. Landlord and rival, all at once. According to a report from SemiAnalysis analyst Jeremie Eliahou Ontiveros, the arrangement could be worth roughly $10 billion. It would include 90-day cancellation options for either side, the same structure SpaceX has already signed with Anthropic and Google. Nothing is signed. Talks are still early, and neither company has confirmed them publicly. The deal, as SemiAnalysis described it, would give Meta private instances of Claude, the kind of arrangement Amazon already runs with Anthropic through its Bedrock service. Meta's own researchers keep building Llama and whatever comes after it inside Superintelligence Labs. It would also be paying to run Anthropic's models on servers Meta built for itself. That's an unusual position for a company that has spent months insisting Llama and its successor models will define its future. The talks sit inside a bigger shift. In January, Zuckerberg announced Meta Compute, a new top-level initiative built to manage how the company engineers, invests in and partners around its data centers. He put Santosh Janardhan, who already runs Meta's global data center and network fleet, in charge alongside Daniel Gross, the former CEO of Safe Superintelligence who now leads long-term capacity planning. Dina Powell McCormick, Meta's new president and vice chairman and a former deputy national security advisor under Donald Trump, works alongside them to line up government and sovereign financing for the buildout. The plan is straightforward. Rent out what Meta isn't using. According to SemiAnalysis, Meta contracted more than 5 gigawatts of cloud and colocation capacity in just the first six months of the year, on top of its own accelerating data center construction. The firm estimates Meta could pull in more than $10 billion a year just by allocating 200 megawatts to outside customers, at the roughly $50 billion per gigawatt annual pricing SpaceX has set as the going rate. An Anthropic deal, the report said, is what kicks off the flywheel. Renting to the competition is now normal A year ago, an AI lab handing its compute to a direct competitor would have looked like a strategic error. Now it's routine. Anthropic already rents the entirety of Colossus 1, xAI's 220,000-GPU supercomputer in Memphis, Tennessee, in a deal reportedly worth $1.25 billion a month, more than $40 billion through 2029. Google pays SpaceX roughly $920 million a month for a separate slice of 110,000 GPUs at the same site, running through June 2029. Anthropic also signed a $19 billion, 20-year lease with TeraWulf for up to 401 megawatts of capacity at a data center campus in Hawesville, Kentucky. Meta reportedly hasn't settled on a model yet. It's weighing whether to sell developers access to AI models running on its own servers, the Bedrock-style approach, or simply sell raw computing capacity the way neocloud operators like CoreWeave do. Either path turns Meta into something it has never been: a vendor competing for the same customers as AWS, Google Cloud and Azure, using infrastructure it paid for out of a capital budget that already runs into the tens of billions annually. Here's the tension nobody at Meta is saying out loud. If you have enough spare capacity to earn $10 billion a year renting to Anthropic, you're either building far more than your own models need, or you've quietly decided owning the infrastructure matters more than owning the frontier model that runs on it. Tom's Hardware reported that news of the talks sent AI infrastructure stocks tumbling, on worry that a compute glut, not a compute shortage, is becoming the more realistic risk. Nothing is signed. SemiAnalysis's numbers describe a deal still being negotiated, and both companies have stayed quiet in public. But the direction is clear enough. The AI labs with the most chips are turning into each other's landlords, and Meta, which spent the past three years insisting it would out-build everyone, is now deciding whether it would rather collect the rent instead. Also read: ASML Gives Nearly Every Worker 20,000 Euros in Stock to Keep Them Through 2030 https://startupfortune.com/asml-gives-nearly-every-worker-20000-euros-in-stock-to-keep-them-through-2030/ • Kimi K3 Shows the Cheap Chinese AI Era Is Coming to an End https://startupfortune.com/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 https://startupfortune.com/upper90-lends-400-million-against-ai-inference-chips-instead-of-nvidia-gpus/