They wanted the cheaper tokens

Anthropic walked away from Decart overnight. Bloomberg has the shape: due diligence done, deal off, reasons undisclosed, the two sides may still collaborate. That is the careful version. Secondary wires amplifying an early-Tuesday exclusive they could not open behind the Bloomberg wall are already bouncing the less careful version around the Israeli and crypto press: this was going to be Anthropic's largest acquisition by a wide margin, somewhere around six billion dollars in mostly stock, for a Tel Aviv inference shop last marked at four billion in May, and that it died in the runway lights of an IPO that Morgan Stanley is said to be lining up as lead-left for a filing that could still land next week — that calendar still sitting in yesterday's FT lane.

Decart is not a chatbot company. Dean Leitersdorf and Moshe Shalev, Unit 8200 alumni, built an optimization layer that squeezes more work out of the same GPUs, then stacked world models on top of it: Lucy for live video, Oasis for physical simulation and driving. Nvidia put money in at the four-billion round. Amazon showed up as a customer. Andrej Karpathy's name was on the angel list. The product that mattered for a frontier lab was not the video demo. It was DOS, the stack that makes inference cheaper when every agent-workday is burning six hundred dollars at the median and seven thousand at the ninetieth percentile, numbers OpenAI published on Sunday without meaning to write Anthropic's shopping list.

That is why a six-billion stock deal for a four-billion private mark made a kind of sense in August. Claude's problem is not only intelligence. It is the cost of serving the intelligence at the volumes Coding and Cowork now want. Buying the people who make the chips go further is a shortcut past waiting for the next rack. Calcalist and other Israeli press say Nvidia, already on the cap table, had floated a higher number, and that the founders preferred Anthropic. Preference is not a signed SPA. By early September Anthropic preferred the quiet of a clean prospectus.

A public company story does not want a fresh six-billion integration in the same breath as the roadshow. Dilution, earnouts, "synergies" that have not shown up in a 10-K yet: all of that is sand in the narrative when you are asking the market to believe a two-trillion conversation. Yesterday's brief already had Morgan Stanley frontrunner, Goldman on stabilisation, a public filing possibly as soon as next week. Walking away from Decart is what a company does when the calendar suddenly matters more than the capability. It is also what a company does when diligence finds something it does not want to explain on a Zoom with long-only accounts. Bloomberg will not say which. Neither will Amodei this morning.

Huang, overnight, was on a different sentence. NVIDIA compute is fungible, durable, and highly rentable, he wrote, pointing at three-year-old H100s still climbing in the rental tape, about twenty-two percent month-on-month toward something like three dollars twenty-eight an hour on the note he quoted. The chip that was supposed to be obsolete is still a bond. Inference demand did not wait for Blackwell to finish arriving. That is Decart's whole business case in someone else's mouth: the installed base still has juice if you know how to wring it, and the people who know how to wring it just watched their preferred buyer choose Wall Street over them.

Slovenia became the sixth EU market for Tesla's supervised FSD while nobody was watching Decart. Brockman showed ChatGPT Work learning a person's writing style out of the connected inbox. The week's loud story remains Astra and Pachocki's alien mind. The quiet story is that the lab racing OpenAI on coding just declined to buy the company that makes tokens cheaper, days before it may ask the public markets to value that race. Collaboration is still possible. The acquisition is not. Decart still has its founders, its cash, and an Nvidia that already liked the price once. Anthropic still has an IPO to keep clean. The cheaper tokens will have to come from somewhere else, or from waiting.