Anthropic is buying its way out of a compute shortage

Anthropic, GIC and Macquarie are teaming up to build data centres. Separately, Anthropic has struck a US$9 billion cloud deal with Riot Platforms. Two announcements, one day, and the surge is only just starting.

Both point at the same underlying problem: the company needs far more computing capacity than it currently controls.

What was announced

The first piece of news is a strategic partnership between Anthropic, Macquarie Asset Management, and Singapore sovereign wealth fund GIC to build AI data centres.

The new entity, Theseus Infrastructure, will initially focus on sites in the United States. Macquarie and GIC will fund the bulk of the equity, while Anthropic will cover any consumer electricity price hikes. That last detail is worth sitting with, given how politically charged power prices have become around large data centre builds.

The second is the US$9 billion cloud deal with Bitcoin mining company Riot Platforms. Riot had earlier announced a 20-year deal to supply 191MW of computing to an unnamed customer, reportedly Anthropic.

So why are we seeing data centre deals of this size, and why now? Two reasons.

Anthropic's compute gap

Unlike its flashier competitor OpenAI, which has inked multiple large data centre deals, Anthropic has relatively few data centres of its own.

As I wrote previously, its AI models run mostly on AWS as part of a US$8 billion Amazon investment, as well as on Google Cloud. Renting at that scale works until it doesn't, and the company is working hard to build up its own footprint.

Last year, AWS opened a data centre in Indiana that exists solely to train and serve Claude. More US data centres are expected to come online this year.

Signs of strain

There is also anecdotal evidence that Anthropic is struggling to meet surging demand for AI capacity. Claude has suffered a string of outages, which is rarely a sign of comfortable headroom.

I've frequently encountered rate limits or server errors in my own use of Claude Code. And compared to OpenAI, Anthropic rarely offers usage "resets" for its top-tier plans. I subscribe to the premium plans of both Claude and ChatGPT, so the contrast is hard to miss.

All this points to a battle in the AI clouds, as increasingly capable models prove genuinely useful in areas such as coding and powering agents. Demand is outrunning the concrete, the power, and the racks.

It's also why I think space-based data centres might arrive earlier than you expect. When the constraint is power and land on the ground, how long before someone stops looking down for the answer?