The Bottleneck Isn't Algorithms Anymore. It's Power and Buildings, So Anthropic Is Financing Data Centers Like Toll Roads.

On 10 August, Anthropic, Macquarie Asset Management, and GIC, Singapore's sovereign wealth fund, announced a new company called Theseus Infrastructure. Its job is unglamorous and enormous: develop, own, operate and lease data centers purpose-built for Anthropic under long-term agreements. The interesting thing is not that Anthropic needs more compute. It is how it has chosen to pay for it.

What was announced

Per Macquarie's own release, funds managed by Macquarie Asset Management and GIC will jointly own the platform and provide the majority of the equity for each project, with Anthropic as the anchor tenant. Initial focus is the United States, and the build-out will require significant capital and, in the companies' words, create thousands of construction jobs and permanent operational roles. One detail stands out: Anthropic committed to covering electricity price increases that might otherwise land on consumers near the sites, a tell about how politically sensitive the power draw of AI has become.

Why the structure is the story

Read past the press release and this is a financing move as much as a compute one. Building a frontier data center is a multi-billion-dollar, multi-year, capital-intensive act that looks far more like a toll road or a power plant than a software product. So Anthropic is doing what capital-intensive industries have always done: it is not putting the buildings on its own balance sheet. Infrastructure investors, Macquarie and GIC, who between them run digital infrastructure globally and invest across more than forty countries, own the physical layer and lease it back. Anthropic keeps its capital for models and talent, and turns a giant capex problem into a long-term operating cost.

What it signals

Two things. First, the binding constraint on AI has moved. It is no longer mainly about algorithms or even chips. It is about power, land, and the money and years to build, which is why the smartest labs now spend as much energy on project finance and grid access as on research. Second, expect AI supply to keep scaling regardless of any single lab's cash position, because the capital to build the physical layer is now flowing through infrastructure funds and sovereign wealth, pools far larger and more patient than venture money. It is the same instinct we saw when Alphabet backed its own departing researchers as a cloud partner: compute is the scarce input, and everyone is arranging their affairs around securing it.

The other direction

It is worth holding this next to the opposite trend. While Anthropic industrialises centralised compute at utility scale, Meta just shipped a capable agent that runs on a single consumer GPU. Both are true at once. The frontier is getting more capital-intensive and more centralised, and the useful floor is getting cheaper and more local. If you are building on AI rather than building AI, the second trend is the one you can bank on: the capability you need keeps getting cheaper to run, precisely because giants like Anthropic are financing the expensive end.