Anthropic May Double Down on Nscale

Anthropic and AI infrastructure provider Nscale are reportedly discussing another $45 billion compute deal, according to Bloomberg, a striking follow-on to a previous agreement between the same two companies for the same amount.

That does not make the deal real yet. TechCrunch, citing Bloomberg, reported that people familiar with the matter said the companies were close to announcing an agreement but that negotiations were still under way and terms could change. Nscale and Anthropic did not immediately provide further detail to TechCrunch.

What changed is simple enough: one enormous capacity commitment may be turning into two. In plain English, Anthropic appears to be trying to lock up far more computing power, not merely talk about ambition in the abstract. In AI, the slogans are cheap and the graphics processors are not.

The useful figure here is $45 billion, because it is large even by the standards of a market that has become oddly casual about writing checks with many zeroes. If a second contract is signed at that level, the combined headline value of the two Nscale deals would reach $90 billion. Headline value, though, is not the same thing as cash paid immediately, recognized revenue, or profit. Compute agreements can stretch over years and depend on buildouts, delivery schedules and usage commitments. Investors tend to discover those distinctions after they applaud the press release.

The timing also matters for both companies’ capital plans. TechCrunch said Anthropic and Nscale are each pursuing initial public offerings. Bloomberg also reported that Nscale has met investor resistance over a proposed $30 billion valuation. A fresh mega-deal could help Nscale argue that demand is deep and durable. It could also sharpen the obvious question: whether the AI industry is booking future infrastructure at a pace that future customers will actually justify.

That tension is not new, but the denominator keeps getting larger. Another $45 billion would not prove the spending is wise. It would show, again, that one of the sector’s biggest model companies is still behaving as if scarce compute is the bottleneck worth overpaying to avoid.