Bloomberg first reported it, Reuters confirmed it, and by midday Thursday the figure had become the talk of the AI infrastructure world: Anthropic has agreed to lease computing power worth about $45 billion from Nscale, a British AI infrastructure company that barely existed two years ago. The six-year agreement, expected to go live near the end of 2027, will draw on a data center Nscale is building in West Virginia, powered by Nvidia’s next-generation Vera Rubin systems.
The deal is one of the largest compute agreements in the industry’s short history, and it says as much about the shape of the AI boom as any chip shipment number. Anthropic, the maker of the Claude models, has spent the past eight months signing a series of multi-billion-dollar capacity agreements as it races to secure the computing power it believes it needs to keep its models ahead of rivals. The company has said little publicly about its strategy, but the pattern is clear: secure capacity early, pay for it with borrowed money and investor capital, and worry about the economics later.
Nscale’s role in the arrangement shows how the industry’s supply chain is fragmenting. The company was founded in 2024, has worked with Microsoft on prior projects, and is now building one of the largest AI data centers in the United States in West Virginia. It is not a chipmaker, not a cloud provider, and not a model company. It is a builder and operator of the physical plant that the AI economy runs on, and its ability to land a $45 billion contract two years after founding is a measure of how fast that economy is being assembled.
The West Virginia site is central to the arrangement. The state has become a magnet for data-center development, lured by available land, cheap power, and a workforce hungry for construction jobs. Nscale has said it will build the facility in phases, and the Anthropic contract gives it the demand certainty needed to finance the first stage. The deal’s structure, leasing compute rather than buying chips outright, lets Anthropic avoid the balance-sheet burden of owning hardware while guaranteeing Nscale’s revenue.
The financing behind the deal is as important as the hardware. A $45 billion lease requires capital that neither company carries, so the agreement is being underwritten by lenders and investors who have concluded that AI compute is the most reliable asset class in technology. Nvidia has been encouraging exactly this kind of arrangement, creating financing platforms with Wall Street firms that can mobilize more than $500 billion for AI infrastructure, and its Vera Rubin systems will be the product inside the Nscale facility. The chipmaker effectively sits on both sides of the transaction, selling the hardware and helping to finance the building that houses it.
For Anthropic, the deal is part of a broader preparation for a public listing. The company has been raising money at escalating valuations, and investors have pressed it to show that it can secure the compute it needs to compete with OpenAI and Google. Contracts like the Nscale agreement are the company’s answer: long-term capacity, locked in at today’s prices, before the next round of scarcity pushes costs higher. The trade-off is that the company is now carrying obligations that will weigh on its financial statements for years, obligations that a public market will scrutinize closely.
The deal also sharpens the debate about concentration in AI infrastructure. A handful of model companies now control most of the world’s frontier AI capacity, and the largest orders flow to a small group of chipmakers and data-center builders. Nscale’s rise suggests that the bottleneck has moved from technology to capital and power: anyone who can secure a site, a grid connection, and financing can participate. That is why infrastructure firms are being valued like technology companies, and why venture investors are pouring money into builders that have never run a data center before.
The deal also highlights how much of the AI economy now runs on financial engineering. Anthropic will not own the Nscale facility, and Nscale will not own the chips outright; the arrangement sits somewhere between a lease, a forward purchase, and a structured credit, with lenders taking risk on both the technology and the electricity. That structure is becoming the industry standard, and it is the reason companies with no profits can still secure tens of billions of dollars in capacity. The system works as long as the underlying asset appreciates or at least keeps earning. If compute prices fall faster than expected, the first losses will land on the lenders, and the next funding round will be harder for everyone.
What remains unclear is whether the demand justifies the commitments. Anthropic’s $45 billion lease, Amazon’s two-million-GPU order, and Nvidia’s own growth targets all rest on the assumption that AI workloads will keep expanding for a decade. The companies involved are betting that models will continue to get better, that enterprises will keep adopting them, and that the energy to run them will be found. If any of those assumptions fails, the leases, the guarantees, and the financing platforms will unwind together. For now, the industry is behaving as if that risk does not exist, and the checks keep getting larger.
