Anthropic and Nscale Reportedly in Talks on Second $45 Billion AI Compute Deal

Anthropic is in talks with UK-based AI cloud provider Nscale on a second AI compute capacity deal worth roughly $45 billion. The first agreement was struck only two months ago. The pace shows that labs now compete less on model architecture and more on access to electricity, chips and data centers.

What happened?

According to Bloomberg, Nscale is negotiating to sell Anthropic roughly $45 billion in additional compute capacity. The talks come two months after the two companies reached a deal of the same size. The report says the agreement could be signed within days, though the terms may change. Neither company commented.

Why does it matter?

A few years ago, an AI lab's biggest expense was salaries, followed by individual training runs. Today the picture is different: the main line item is multi-year capacity contracts committed in advance. Finding the GPUs needed to train a model on the spot market is getting harder. Locking in capacity years ahead has become a precondition for serving customers, on both the training and the inference side.

Anthropic is clearly not alone in this strategy. OpenAI's large commitments with providers such as Oracle, Microsoft and CoreWeave, along with Meta's and xAI's own data center investments, follow the same logic. Anthropic, for its part, has historically chosen to spread across Amazon, Google and additional providers rather than be locked into a single cloud. A second deal of the same size suggests that this diversification is continuing and that demand is growing faster than expected. That is an inference; the companies have not said so.

The real question is how a relatively young provider like Nscale will carry an obligation of this scale. A contract this size requires the provider to buy the chips, land and energy it needs. The capital required would far exceed what the provider has on its own balance sheet. The common approach in the industry is to raise debt using the customer contract as collateral: a long-term contract with a creditworthy buyer amounts to a cash-flow guarantee for banks and private credit funds. It is reasonable to guess that Nscale is likely following this path, but the details of its financing structure are not public.

Who is affected, and how?

Deals like this shift where risk sits. For Anthropic, the upside is secure access to capacity and less dependence on the pricing power of a single hyperscaler. In return, it takes on a long-term, inflexible payment obligation. If revenue growth falls short of expectations, those fixed costs could squeeze the company.

For Nscale, the deal is a springboard, but it also carries concentration risk. A provider that depends on a single customer for most of its revenue becomes highly sensitive to that customer's decisions. CoreWeave's similar reliance on a handful of large customers has been a risk item that investors have watched for years.

Dependence on Nvidia is a quiet but decisive factor in this picture. Most new-generation cloud providers are built on Nvidia GPUs. Anthropic, meanwhile, also uses alternatives such as Google's TPUs and Amazon's Trainium chips. It is not yet clear which hardware the Nscale deal covers; that detail will be an important clue to Nvidia's bargaining power.

What's next?

There are three things to watch. First, whether the deal is officially announced, and in which region and with which energy source the capacity will be built. Second, Nscale's funding rounds and debt structure, which is the most fragile link in the system. Third, whether rival labs respond with similar moves.

Here is my assessment: two $45 billion commitments in two months confirm that labs have put compute capacity at the center of their product strategy. But this model is sustainable only as long as demand keeps growing. If capacity is built quickly while revenue lags expectations, the financing side, the weakest link in the chain, will be the first place to shake.

The Bloomberg report and similar developments in the industry are available via The Next Web.