Anthropic's $6B Decart talks: buying down its own token cost

Anthropic is in talks to acquire Israeli startup Decart for about $6 billion, which would be its largest known acquisition. The reported reason isn't a product line or a customer list. It's throughput: Decart's technology makes chips run models more efficiently, and Anthropic wants its existing infrastructure to absorb more demand. When a frontier lab spends $6B on inference cost rather than on capability, that tells you where the pressure is — and what your own bill is pinned to.

What actually happened

Bloomberg reported the talks today, with details also carried by Fortune:

• ~$6 billion, which would be Anthropic's largest known deal. It is not finalized and talks could fall through. Both companies declined to comment.
• Decart was founded in 2023 by three Israeli engineers — brothers Dean and Orian Leitersdorf, and Moshe Shalev.
• It raised $300M in May 2026 at close to a $4B valuation, up from $3.1B in August 2025. Backers include Radical Ventures, Nvidia, Sequoia, Benchmark, and Adobe Ventures. A $6B price is roughly a 50% premium on the May mark.
• Decart builds real-time world models — its Lucy model generates live video overlays — plus software that cuts the cost of training and serving models by using chips more efficiently.
• The reported plan is to fold Decart's work into Anthropic's inference and performance organization to absorb surging demand.

Treat the strategic framing as reporting from people familiar with the matter, not a published roadmap. The dollar figure and the funding history are the firm parts.

Why your vendor's inference cost matters for your business

You don't negotiate your token price. You inherit it. The published rate you pay is a retail markup on what it costs the lab to run one forward pass on hardware it rents — and that cost floor is set by silicon supply, power contracts, and how many tokens per second the lab can squeeze out of the GPUs it already has. A $6B bet on the third variable is a lab telling you the first two aren't getting cheaper fast enough.

The near-term read for operators is boring and useful: stop modeling your AI line item on the assumption of steady price cuts. Labs are buying efficiency because they need margin, not because they plan to hand it to you. Some of that gain shows up as capacity — fewer rate limits, better latency at peak — which is worth real money if your automations run interactively. Some of it shows up as the price simply not going up.

There's a second-order effect worth pricing. Consolidation at the inference layer means the companies that were selling everyone a cost advantage get absorbed into one buyer's stack. If your architecture assumes you can swap providers when one gets expensive, that assumption depends on the cheap-inference vendors staying independent. Fewer of them is fewer escape hatches — which is exactly why the abstraction layer belongs in your code, not in your vendor's SDK. Same argument we made when Schneider bought Cognite and when Stripe took a stake in OpenRouter.

Sources: Bloomberg, Fortune.