Anthropic ended talks to acquire Decart after conducting due diligence, walking away from a deal reported at roughly $6 billion that would have been its largest known acquisition, per Bloomberg. No final agreement had been reached, and people familiar with the talks say the two could still cooperate in other forms. Talks were first reported in mid-August. Both companies declined to comment.
Decart is worth understanding to see what was on the table. Founded in 2023 by Dean Leitersdorf, Orian Leitersdorf, and Moshe Shalev, it raised a $300 million Series B led by Radical Ventures at a valuation near $4 billion, with Nvidia and Amazon among its backers. Its DOS optimization stack targets inference speed and cost across Nvidia GPUs, Google TPUs, and Amazon Trainium, and the company also ships real-time video and world models. The relevant capability is the cross-silicon part, because Anthropic runs on all three of those chip families and has committed enormous sums to capacity it does not own.
That is why the walk-away reads as a decision rather than a stumble. Anthropic has signed roughly $45 billion with Nscale, a 20-year TeraWulf lease, and a series of other commitments this year, and every point of inference efficiency compounds directly against those fixed obligations. A company buying that much capacity has an obvious reason to own the software that makes it go further. Paying $6 billion for a business last marked near $4 billion, when the acquirer is preparing for an IPO and public investors will scrutinize both the price and the goodwill, is a different calculation. Ending talks after diligence rather than before it suggests the price stopped clearing once the numbers were open.
For the market the useful signal is negative evidence, which is rarer than the announcement kind. Frontier labs have spent the year demonstrating they will pay nearly any price for compute and talent, and this is a case of one declining to. Anthropic can also rent the outcome instead of buying the company, which is what a continued commercial relationship would look like, and that option gets more attractive the closer a listing gets. Decart, for its part, keeps its independence and its Nvidia and Amazon relationships intact, which are harder to preserve inside a lab that competes with both.
Bottom Line
A lab that has committed tens of billions to leased compute decided a $6 billion inference-efficiency acquisition was not worth it after seeing the books. Read it as pre-IPO price discipline, and watch for a commercial deal between the two, which achieves most of the same thing without the goodwill.