Anthropic Is Reportedly Paying 7 Billion Dollars for Software That Makes Chips Go Further
Anthropic and Decart are exchanging advanced drafts of a 7 billion dollar acquisition, mostly in shares, with signing targeted for early September. Decart claims models can run up to 8 times faster on the same hardware.
Anthropic and the Israeli startup Decart are reportedly exchanging advanced drafts of an acquisition agreement worth around 7 billion dollars, up from the 6 billion figure Bloomberg reported on 13 August. Most of the consideration would be paid in Anthropic shares, with signing targeted for early September, ahead of Anthropic’s expected public offering. It would be the largest acquisition in the company’s history. Nothing is signed, so treat the whole thing as reported rather than done.
Decart is three years old, has raised about 450 million dollars and employs roughly a hundred people. It builds two things: real-time world models for video and robotics, and software that makes AI models run substantially faster on the chips you already have. Industry estimates put the speedup at up to eight times, though that figure comes from the company’s own positioning rather than independent testing. Decart’s founders Dean Leitersdorf and Moshe Shalev, along with lead investor Sequoia, reportedly preferred Anthropic even though Nvidia, already an investor, is said to have offered a higher valuation. One stated reason: the deal would give Anthropic its first development centre in Israel.
The strategic logic is worth spelling out, because it is the same logic showing up all over the industry right now. Training a model is a one-time cost. Inference, the word for actually running the model to answer a question, is a cost you pay on every single request forever. For any company with real users, inference now dwarfs training in the budget. So a piece of software that makes the same chips serve more requests is not a nice optimisation, it is directly a gross margin. Anthropic reported preliminary second-quarter revenue above 11.5 billion dollars with positive adjusted operating income, its first quarter in the black, and it is heading toward an IPO where margins will be examined closely. Buying efficiency outright, in shares, before the listing, is a fairly legible move.
The caveats are the usual ones for a deal of this shape. Efficiency claims of “up to eight times” almost always describe a best case on a particular workload. Paying mostly in stock means the real price depends on what Anthropic shares turn out to be worth. And large acquisitions immediately before an IPO have a mixed record generally.
What this means for you: if you pay for AI by the token, this is the kind of news that eventually shows up as lower prices, because the same hardware serving more requests is exactly what lets providers cut rates without losing money. If you follow the industry, note the shift: the interesting money is moving from making models bigger to making them cheaper to run. That is usually what happens when a technology stops being a demo and starts being a utility. Nothing to do today, but a useful signal about where the next year goes.
Sources
Source: https://www.calcalistech.com/ctechnews/article/b1evv3aufg
OpenAI Cut Its Flagship Model to Half Price, and the Reason Is Sitting on a Leaderboard
GPT-5.6 Sol is 50 percent off through 18 September on OpenRouter and Vercel's gateway, taking it to 2.50 dollars per million input tokens. It follows cuts to Luna and Terra as Chinese models gain ground.
