Anthropic rents all of Colossus, and immediately spends it

Anthropic announced this week that it is taking the entire compute capacity at Colossus 1, the xAI-built data centre in Memphis, somewhere north of 220,000 NVIDIA GPUs, with another 300 MW of capacity coming online within the month. The deal is on top of, not instead of, Anthropic's existing arrangements with Amazon, Google, Microsoft, and FluidStack. And the announcement landed alongside a second, less-discussed move: Anthropic doubled rate limits across Pro, Max, Team, and Enterprise tiers for Claude Code and the API, and removed peak-hour throttling.

The two announcements are the same announcement. They have to be read together, because separately they don't make sense.

What the deal actually is. Anthropic is not buying Colossus. It is taking the entire output of a facility built and owned by a competitor's affiliated infrastructure arm. The headline number, 220,000+ GPUs, is the installed base; the 300 MW addition is the part that matters for the next twelve months of capacity planning. Memphis is interesting because it was xAI's flagship build, the one Elon Musk repeatedly cited as evidence that xAI could close the compute gap with frontier labs through sheer infrastructure speed. Anthropic has now rented all of it.

The structure of the agreement has not been disclosed in the kind of detail that would let me say whether this is a tenancy, a take-or-pay, a capacity reservation with optionality, or something more bespoke. The press materials describe it as Anthropic using "all compute capacity" at the site, which is the kind of phrase that can mean several things in a contract. I'd watch for the next 10-Q from Amazon, which discloses Anthropic-related commitments in segment commentary, to see whether this shows up as a third-party arrangement or whether AWS is somewhere in the structure. There are reasons to suspect the latter.¹

Why the rate-limit move matters. Frontier labs do not double rate limits because they want to. They double rate limits because they have capacity that needs absorbing or because competitive pressure on usage caps has crossed a threshold. The Anthropic move is almost certainly the former: 300 MW comes online, the marginal cost of an idle GPU is the full cost of the GPU, and the cleanest way to convert capacity into revenue is to let existing paying users consume more.

This is the inference-economics frame doing its job. Training is lumpy and capitalised; inference is continuous and has to be filled. When a lab signs for hundreds of megawatts of new capacity, the question is not "can they train a bigger model", they can, but "what fills the inference side of the cluster between training runs". Doubling rate limits is the answer. Removing peak-hour throttling is the answer. The capacity has to go somewhere.

What this is a case of. It is the second time in eighteen months a frontier lab has effectively absorbed a competitor's infrastructure build. OpenAI's Stargate arrangements pulled in capacity that had been pitched, at various points, as serving a broader market. Anthropic taking Colossus is a sharper version of the same move: the facility was built by xAI's affiliate, branded as xAI's competitive answer, and is now running Claude inference. The capital that built Memphis is being repaid by Anthropic's revenue.

This matters for how to read xAI's position. The Colossus build was load-bearing in the xAI narrative, proof that Musk's infrastructure could outpace incumbent labs. Renting it to Anthropic is a perfectly rational commercial decision (the GPUs earn more leased to a paying frontier lab than running xAI's own workloads at whatever utilisation xAI is achieving), and it is also a quiet admission about where the demand actually is. xAI is not, on this evidence, inference-constrained in the way Anthropic is.

The capital that built Memphis is being repaid by Anthropic's revenue.

The performativity question. The AI-performativity frame asks whether spending at this scale is chasing revenue or chasing positioning. Anthropic's compute stack is now: Amazon (Trainium and Nvidia), Google (TPU), Microsoft (announced earlier this year), FluidStack, and now Colossus. Five infrastructure relationships, at least four of them with multi-year commitments, for a company whose disclosed annualised revenue at the most recent reporting was an order of magnitude smaller than the implied compute spend.

The frame predicts that at some point the spend has to either (a) be backed by revenue that justifies it, (b) be backed by capital markets willing to fund the gap indefinitely, or (c) compress. Anthropic's last round closed at terms that suggest capital markets remain in mode (b). The rate-limit doubling is the company trying to push toward (a) by letting paying customers consume more of what they're already paying for, which is a margin question, not just a revenue question. If usage doubles and revenue per user is flat, the inference-cost line gets uglier, not prettier.

This is where the frame holds and where it doesn't. Inference economics predicts that capacity additions force pricing or rate-limit moves. That has happened. It also predicts that margin pressure shows up in the unit economics over the following two to four quarters. We don't yet have the disclosure to test that. Anthropic is private. The numbers that would settle whether Colossus is accretive or dilutive to gross margin are not in the public record and will not be until either a funding round prospectus or, eventually, an S-1.

What to watch.

The next Anthropic pricing move. If rate limits doubled and pricing held, the company is betting that elasticity works in its favour, more usage at the same price produces more retention and more enterprise pull-through. If pricing rises within two quarters, the bet didn't work and Colossus is a margin problem.

xAI's next infrastructure announcement. If Memphis is leased out, the next Musk-orbit data centre announcement tells you whether xAI has pivoted to being an infrastructure provider for other labs (a perfectly good business) or is doubling down on training its own models with capacity it actually controls.

Amazon's segment commentary. AWS reports include enough texture on Anthropic-related commitments that a careful read of the next two quarters will indicate whether the Colossus arrangement sits inside or outside the AWS umbrella.

The Claude Code usage curves, to the extent Anthropic discloses them. Doubling rate limits is a capacity-utilisation move. Whether it produces the usage that justifies the capacity is the test.

Anthropic has rented an entire data centre from a competitor and immediately given its existing customers permission to use twice as much of it. That is not a contradiction. It is the shape of the inference economy when the capacity arrives faster than the demand.

Footnotes

¶