Bloomberg reported late Monday that Anthropic completed due diligence on the Israeli startup Decart and then decided against buying it. The number attached to the talks was roughly $6 billion, which would have made it Anthropic’s largest acquisition by a wide margin. Representatives for both companies declined to comment. One person familiar with the discussions said the two may still find other ways to work together, and that detail turns out to be the most useful sentence in the story.
The talks first surfaced in August at about the same valuation, with the usual caveat that nothing was signed. Now nothing is.
Most of the coverage since August has described Decart as a world-model company, which is how it is known publicly: real-time video generation, interactive environments, the demos that circulate. That is the visible product. The acquisition rationale, per the reporting, sat somewhere much duller. Decart also sells an optimisation layer meant to get more throughput out of the same silicon, across both training and inference. The company’s own marketing talks about squeezing performance out of every chip. Anthropic was not shopping for a new product category. It was shopping for a cheaper cost of serving customers.
What six billion dollars was actually buying
Anthropic has spent the past year buying compute the ordinary way, by contract and at scale; the $35 billion cloud agreement with Lambda is the headline example. A Decart purchase would not have added a single chip to that pile. It would have changed the ratio between the capacity already committed and the demand that capacity can absorb.
That distinction matters more than it sounds. For a company selling tokens, inference efficiency is not an engineering statistic, it is gross margin. A few points of throughput per accelerator flow straight to the line that a listing prospectus gets judged on, and they flow there permanently, without renegotiating a single supply agreement. If you believe compute cost is the binding constraint on the business for the next several years, paying a premium to own the thing that relaxes it is defensible arithmetic.
Which is what makes the walk-away interesting.
Why this diligence was more informative than most
Large AI acquisitions usually rest on claims that diligence cannot really test. Team quality, roadmap credibility, strategic option value: these are judgements, and reasonable buyers reach opposite conclusions from the same data room. Deals in that category collapse for reasons that tell you almost nothing about the asset.
An efficiency thesis is the rare exception. The buyer already has the workloads, the hardware, and the baseline numbers. A vendor claiming a multiple on tokens per second either reproduces that on the buyer’s own traffic or it doesn’t, and the test runs in weeks rather than quarters. Anthropic is about as well equipped to run it as any organisation on earth.
So a completed diligence process followed by a decision not to proceed carries more signal here than the equivalent outcome would in a talent deal. It does not follow that the technology fails. The far more common result is that gains turn out to be real and narrow: strong on certain model architectures, weaker on the ones you actually run in production; impressive against an unoptimised baseline, less so against the kernels your own team already shipped last quarter. An acquirer can conclude the software works perfectly well and still decline to pay $6 billion for the increment it adds to work already done in-house.
The reporting does not establish which consideration ended the talks, and price remains a live explanation on its own. Several outlets have quietly promoted “walked away after diligence” into “diligence found something,” which the sourcing does not support. Hold that one loosely.
The IPO clock cuts both ways
Decart raised $300 million in May, led by Radical Ventures, with Nvidia, Adobe Ventures, Valor Equity Partners and Atreides Management joining existing backers including Sequoia, Benchmark and Zeev Ventures. That round put the company close to $4 billion, up from $3.1 billion the previous August. A $6 billion purchase, reportedly structured mostly in stock, would have been roughly a 50% step-up four months later.
Stock is the awkward part. Anthropic is expected to begin marketing an IPO as early as mid-October, with a listing targeted before the November midterms. Paying in equity weeks before that equity gets priced by public investors means setting an internal mark on your own currency at precisely the moment you least want to argue about what it is worth. It also means walking onto a roadshow with a fresh, unintegrated $6 billion acquisition in the story.
A listing candidate wants two things that point in opposite directions here: the leanest possible cost of revenue, and an acquisition history that looks disciplined. Decart improved the first and complicated the second.
There is a further wrinkle circulating in secondary coverage, so far unconfirmed by the original reporting, that Nvidia, already a Decart investor, had put forward a competing offer the founders passed on in favour of Anthropic. If that holds up, the sequence is considerably less comfortable for Decart than a simple failed negotiation.
What to watch from here
Two things are observable and will settle most of the ambiguity. The first is whether a commercial arrangement between the two companies appears in the coming months; the “may still collaborate” line points that way, and a licensing or partnership deal would suggest the technology cleared and only the price did not. The second is where Decart prices next, whether in a round or a sale. A mark meaningfully below the May valuation would support the reading that diligence turned up something structural. A mark at or above it would suggest Anthropic simply declined to pay a 50% premium in its own pre-listing shares, which is a statement about Anthropic’s stock rather than about Decart’s software.
One deal died. The cost problem it was meant to solve did not.
