AI

Anthropic’s biggest deal ever wouldn’t buy a smarter model. It would buy cheaper tokens.

Bloomberg reported Thursday that Anthropic is in talks to acquire Israeli startup Decart for about $6 billion — roughly a 50% premium to the $4 billion valuation Decart carried in May. Decart’s core product makes AI chips run more efficiently. Read that twice.

N Noah · The Sharp Brief · August 14, 2026 · 4 min read

Anthropic is in talks to buy Decart for around $6 billion, Bloomberg reported Thursday. It would be the largest known acquisition in the company’s history by a wide margin. The talks are not finished and the deal could still collapse; neither side is commenting.

Decart is three years old. Founded in 2023 by Dean Leitersdorf and Moshe Shalev, both Unit 8200 alumni, it has raised roughly $450 million in total — including a $300 million round in May led by Radical Ventures, with Nvidia joining as a new investor, at a valuation of about $4 billion. A $6 billion price tag is a roughly 50% markup on a number that was set three months ago.

The company builds two things. One is a set of world models — systems that simulate physical environments for applications like autonomous driving and commerce. The other, and the one that explains this deal, is software that squeezes more work out of the same silicon: making inference and training run more efficiently on the chips you already have.

This is a margin purchase

Frontier labs have spent two years buying capability. Anthropic appears to be buying throughput. If your compute is fixed by supply contracts and power, the only lever left is how much useful output each chip produces per second — and the only way to move that lever fast is to buy the people who already moved it.

The context helps. Anthropic has signed up for two gigawatts of AMD’s Helios racks, is hiring a custom silicon team, and is widely reported to be preparing for public markets. Every one of those moves points the same direction: the company is trying to own more of its own cost structure before anyone gets to look at the books.

Our take: A $6 billion cheque for an efficiency layer tells you what the binding constraint is at a frontier lab in 2026. It isn’t ideas and it isn’t talent — it’s cost per token. When the smartest buyers in AI start paying model-sized money for infrastructure software, the market has quietly stopped being a capability race and started being a gross-margin race.

The auction nobody admits to

Decart has been circled for weeks. Earlier reporting put Nvidia in advanced talks before a bigger bid arrived; Amazon and Nebius have both been named. A report that SpaceX was closing on the company was called “fake news” by Elon Musk on X. What survives all that noise is the price: multiple serious buyers converged on roughly the same $6 billion number for a 100-person company with no consumer product.

What to watch

Advertisement

Get the day, decoded — at 7 PM ET

The Sharp Brief: AI, money, business & performance in five sharp minutes. Free.

Free bonus: subscribe today and The 2026 AI Playbook (PDF) lands with your welcome email.

Recommended by 5+ newsletters across AI, markets & business.