Anthropic Walks Away From Its $6 Billion Decart Deal

Anthropic has stepped back from what would have been the largest acquisition in its history. According to Bloomberg, the company spent months exploring a purchase of Decart, an Israeli startup whose technology makes AI chips run more efficiently, in a deal valued at roughly $6 billion. After completing its due diligence, Anthropic said no.

The logic behind the talks was easy to follow. Decart's software squeezes more useful work out of the same silicon, which lowers the bill for training and serving large models. Compute is by far Anthropic's biggest cost, so a tool that trims it goes straight to the part of the business that hurts most. Owning that technology outright, rather than renting improvements from a supplier, would have given the company a lever over its own economics.

What makes the collapse sting is the position it leaves Decart in. Reporting from Calcalist and Ynet says the startup's founders had turned down a larger, mostly cash offer from Nvidia in order to pursue Anthropic's stock-heavy proposal. Betting on equity in a fast-rising company is a reasonable wager, right up until the buyer changes its mind. For now Decart is left with neither deal, though both sides have signalled they may still cooperate in some looser form.

Why the timing matters

Anthropic is not short of money or ambition at the moment. The company is preparing a record public listing and recently committed to a $35 billion data centre build in Texas. Against spending on that scale, a $6 billion acquisition to shave running costs looks less like a stretch and more like housekeeping. That it fell through anyway suggests the numbers, or the fit, did not survive a close look.

Neither company has spelled out what changed. Due diligence exists precisely to surface the things a term sheet cannot, and plenty of deals die quietly once the buyer sees inside. It may be that Decart's gains looked smaller in Anthropic's own stack than in a demo, or that the price stopped making sense once the technical picture was clear. Without a statement from either side, the honest answer is that we do not yet know.

What the episode does show is how central raw efficiency has become to the economics of frontier AI. When a lab worth a fortune is willing to spend billions simply to run its models more cheaply, the cost of inference has stopped being a footnote. It is now one of the main events.