Anthropic is in talks for a second $45 billion compute deal with infrastructure provider Nscale, according to people familiar with the discussions. The combined commitment would reach $90 billion, following a $45 billion agreement announced roughly a month ago.

The scale of capital deployment reflects Anthropic's bet that securing guaranteed GPU capacity is essential to competing for frontier AI models. Training Claude requires hundreds of millions in compute spend per cycle. Pre-purchasing capacity locks in supply and sidesteps the spot-market volatility that has already forced other labs to pause or scale back training runs.

But the strategy exposes a hard math problem: Anthropic has no clear path to offset $90 billion in infrastructure spend through API revenue or licensing deals. The company's largest customer contracts, while significant, generate tens of millions annually—not billions.

Nscale stands to become a quasi-utility for AI training, extracting steady revenue from long-term commitments. Yet the infrastructure provider faces its own valuation pressure. It is seeking to go public with a $30 billion valuation, but investors are questioning whether the company's financial projections justify the ask, according to reporting from Monday.

The tension cuts both ways. Anthropic needs to demonstrate that scale translates to competitive advantage—that a larger training budget produces models that command pricing power or exclusive enterprise deals. Nscale needs to prove its capacity and margins justify valuations that assume years of high-margin utilization.

For Anthropic, the deals also reflect capital market timing. The company is expected to pursue a public offering this year. A multi-year compute contract de-risks investor concerns about supply chain fragility and positions the company as the well-funded player in the AI race.

The broader story is economic: as AI models grow larger, only a handful of well-capitalized players can afford entry. Anthropic is betting that capital intensity itself becomes a moat. Smaller competitors cannot match $90 billion in guaranteed compute, making the race not about innovation but about who can burn cash the fastest and still survive to profitability.