Meta May Rent Its Spare GPUs to Anthropic

Meta is in talks to rent out compute capacity from its own data centers, and Anthropic could be the first major customer. The Decoder reports the deal could be worth up to $10 billion over two years, citing the New York Times. Anthropic pitched the arrangement back in June. Meta is still reviewing it, and according to The Decoder, either side could still walk.

Nothing is signed. But the fact that this conversation is happening at all tells you something important about where the AI buildout stands.

Why This Matters

  1. Meta becomes a compute seller, not just a compute buyer. Zuckerberg already floated this to investors: if Meta’s own AI demand didn’t absorb all the capacity it was building, he’d sell the excess. This would be that plan going live.
  2. A rival is renting from a rival. Anthropic and Meta are competitors in the model race. Compute scarcity is strong enough that competitive lines are getting blurry.
  3. The $145 billion question gets an answer. Meta plans to spend up to $145 billion this year, mostly on AI, per The Decoder. Investors have been asking how that gets paid back. Renting capacity to a company with money and urgency is one answer.

The Status Quo It Breaks

Until now, the compute market had a clean structure. Hyperscalers with cloud businesses (Microsoft, Amazon, Google) sold capacity. Everyone else bought it. Meta sat firmly in the buyer column, building enormous infrastructure purely for its own products and models.

This would move Meta into the seller column without it ever launching a cloud business. That’s a meaningful shift. It means any company with a large enough data center footprint is potentially a compute vendor, whether or not that was ever the plan.

What stands out here is the timing. Meta is building capacity faster than its own AI products can consume it, while Anthropic is consuming capacity faster than it can build. Two opposite problems that happen to solve each other.

Why Anthropic Needs It

Demand for Claude Code has surged, according to The Decoder. Coding agents are compute-hungry in a way chat products are not. They run long, they run repeatedly, and they burn tokens on every iteration.

Anthropic has already been stacking supply. The Decoder notes a $45 billion deal previously signed with SpaceXAI. Long term, though, the company wants its own data centers and has hired former Google executives to lead that effort.

So the Meta deal, if it happens, reads as a bridge. Rent now, build later.

Media Watch: What Else Meta Is Doing

Compute rental isn’t the only new revenue line Meta is testing:

  • Paid AI subscriptions under the “Meta One” banner
  • Investment in space-based solar power to feed its data centers

That second one sounds exotic until you remember that power, not chips, is becoming the real constraint on AI infrastructure. Meta is hedging on energy the same way it’s hedging on compute demand.

Implications for Practitioners

If you build on Claude or run agent workloads, the practical read is straightforward:

  • Capacity pressure is real. Anthropic shopping for $10 billion in rented compute is a signal about how tight the supply picture is for coding agents specifically.
  • Rate limits and pricing follow supply. More capacity secured usually means more headroom downstream. Watch for it.
  • Multi-provider infrastructure is becoming normal. Frontier labs are stitching together capacity from whoever has it, including competitors. Vendor loyalty in compute is over.

What Comes Next

Watch for three things. Whether Meta confirms it publicly, since that would formalize a compute-rental business. Whether other large infrastructure owners follow the same move. And whether Anthropic’s own data center project stays on track once cheaper rented capacity is available.

The deal may still collapse. But the logic behind it won’t. When one company overbuilds and another underbuilds, someone eventually writes the check.

Full details are at the original report.

Scroll to Top