Anthropic is in early talks to lease computing power from Meta in a deal worth roughly $10 billion, the New York Times reported Friday. The discussions are described as very preliminary: Anthropic proposed the arrangement in June, would pay Meta in monthly installments over about two years, and both companies would keep the option to exit early. Sources cautioned a final deal may never materialize.
The market graded it instantly anyway. Meta shares, lower earlier in Friday’s session, climbed off their lows after the report crossed. That reaction tells you exactly what investors want from Meta right now: any evidence that the most aggressive AI buildout in tech can generate direct revenue — instead of waiting years for ads and agents to justify the spend.
For Anthropic, the logic is simpler: it is buying capacity anywhere capacity exists. This month alone it signed a $19 billion data-center lease with TeraWulf, and weeks earlier it struck a deal to use SpaceX’s Colossus 1 capacity. A Meta lease would be one more shelf in a compute pantry that can never be full enough.
The anchor tenant “Meta Compute” needed
Two weeks ago, Bloomberg reported Meta was standing up a cloud unit to sell its excess AI capacity. That story was an org chart. This one is a customer. Landing a frontier lab as a ~$10 billion anchor tenant would turn Meta’s cloud ambition from concept into business overnight — with Hyperion, its 5-gigawatt, $50 billion Louisiana campus, still to come.
The catch: Meta has never operated as a cloud provider. It has no AWS-style tooling, billing, or enterprise support layer, and building one is years of unglamorous work. Which is why this deal shape is telling — renting raw capacity to a single, sophisticated tenant is the easiest possible version of a cloud business. You don’t need a console and a sales force to serve one customer who knows exactly what it’s doing.
Our take: Rivals renting to rivals is the defining weirdness of this cycle. Meta sells models through its own paid API, and it would now power a lab it competes with — because compute is so scarce that the org chart stops mattering. Every hyperscaler wants its capex line reclassified from “terrifying cost” to “revenue-producing asset,” and leases like this are how it happens. If even the very-preliminary version survives, expect every big-tech earnings call to get a “would you lease spare capacity?” question within a quarter.
What to watch
- Whether “very preliminary” survives contact. The NYT’s sources flagged real odds this never closes; the two-year term with early exits reads like both sides hedging.
- Meta’s Q2 earnings call. Management commentary on leasing capacity — or a named cloud business line — would confirm the Bloomberg-reported unit is real and funded.
- Anthropic’s compute stack. TeraWulf, SpaceX, and now possibly Meta — watch how the lab layers these against its existing cloud relationships.