Anthropic’s Biggest Deal Yet: $6B for Decart
Anthropic is in early talks to buy Israeli startup Decart for about $6 billion — roughly 50% above the valuation Decart set three months ago. The prize is not the video models everyone knows it for. It is the chip-efficiency software underneath them.
Anthropic is in early-stage talks to acquire Decart, the Israeli startup behind the real-time video model Lucy, in a deal worth roughly $6 billion, Bloomberg reported on Thursday. Nothing is signed, both sides declined to comment, and people familiar with the discussions cautioned the deal could still fall apart. If it closes, it would be the largest acquisition Anthropic has ever made.
The price is the first thing worth noticing. Decart raised $300 million in May at a valuation of about $4 billion, in a round led by Radical Ventures with participation from Nvidia, Adobe Ventures and Atreides Management. That itself was up from $3.1 billion in August 2025. A $6 billion offer three months later is a 50% markup on a company that has raised roughly $450 million in total since it was founded in 2023 by Dean Leitersdorf, Orian Leitersdorf and Moshe Shalev.
Decart is best known publicly for its generative video work. Lucy takes a live video feed and rewrites it in real time — the kind of thing that turns into virtual try-on for fashion e-commerce. Oasis, its world model, generates interactive simulated environments used to train and test robotics and autonomous driving systems. Both are trained on text plus millions of hours of video, with the goal of giving a model a working sense of how physical objects behave.
But the reported destination for Decart's team inside Anthropic says something different. According to the reporting, the staff would join Anthropic's inference and performance organisation — not a research group, not a product group. Decart's less-publicised business is software that makes AI chips run more efficiently. Calcalist reports the technology can push model throughput to as much as eight times the typical rate, and that it works across Nvidia, Google and Amazon silicon rather than being tied to one vendor.
That is the shape of the bet. Training runs are lumpy, scheduled and capital-intensive; inference is continuous, demand-driven and where the bill actually lands once a product like Claude has real usage. Anthropic has spent this year buying its way out of that squeeze from the supply side — a $9.1 billion, 191 MW lease from a bitcoin miner and a set of data centres it does not own. Buying a serving-efficiency team is the same problem attacked from the other end: instead of renting more power, get more tokens out of the power you already have.
The timing matters too. Anthropic is reported to be preparing for an IPO, and unit economics on inference are exactly the line item public-market investors will interrogate. A one-off $6 billion acquisition that permanently bends the cost-per-token curve reads very differently on a prospectus than another multi-year compute lease. Hardware neutrality is a second, quieter benefit: efficiency software that spans three chip vendors is leverage in every future supply negotiation.
For Israel's technology sector, a completed deal would plant a frontier lab's engineering centre in the country outright rather than as a satellite office — a comparison local observers are already drawing to Intel's purchase of Mobileye and what it did for autonomous driving there. Whether any of that happens still depends on talks that both companies will only describe by declining to describe them.
Want AI news before everyone else?
The morning's most important AI stories, straight to your inbox. No fluff.