Anthropic is reportedly in talks to buy Israeli startup Decart for $6 billion, its largest deal yet

[Anthropic](#/company/anthropic) is in early talks to acquire Decart AI, an Israeli startup building chip-efficiency and real-time video-generation software, for roughly $6 billion, Bloomberg reported August 13, citing people familiar with the matter. If it closes, it would be Anthropic's largest acquisition to date. Nothing is signed yet, and Bloomberg's sources caution the talks could still fall apart before a deal is reached.

Decart is three years old. Dean and Orian Leitersdorf and Moshe Shalev, engineers who met in Israel's Unit 8200 military-intelligence corps, founded it in 2023 and emerged from stealth in late 2024 with what the company says was profitability within three months of launch. Its stack has three parts: DOS, an optimization layer that raises inference throughput across Nvidia GPUs, AWS Trainium and Google TPUs; Lucy, a real-time video-editing model used for virtual try-on and live advertising; and Oasis, a world-generation model aimed at training and testing robots and self-driving systems. Decart says DOS runs full-HD video inference at up to 100 frames per second and pushes agentic inference to roughly 1,600 tokens per second, against what it puts at a roughly 200-token-per-second industry baseline — a company-reported figure, not one this newsroom has independently verified.

The company closed a $300 million round in May 2026, led by Radical Ventures with Nvidia, Adobe Ventures and Toyota Ventures participating, at a valuation nearly $4 billion — up from $3.1 billion the previous August. A $6 billion acquisition price would sit about 50% above that May mark, though the comparison is apples to a slightly different fruit: a venture round values minority shares traded among existing and new investors, while an acquisition price is what a single buyer would pay for the whole company.

Oasis, Decart's least-discussed product in this week's coverage, is worth a beat on its own: a "world model" that generates interactive, physically-consistent environments in real time, aimed at training and testing robots and self-driving systems without the cost of running them in the physical world first. World models are the same category behind Google DeepMind's Genie line and a growing slice of robotics-lab research spending industry-wide — the pitch is that simulated environments are cheaper and safer to fail in than warehouses and roads. Anthropic has no public robotics or embodied-AI program today, which is part of why Bloomberg's sourcing points at DOS, the efficiency layer, as the actual target rather than Oasis or Lucy.

The strategic logic, per Bloomberg's sourcing, is about capacity rather than Decart's consumer-facing video tools: Anthropic wants DOS folded into its own inference stack to help its existing infrastructure absorb more demand as adoption keeps outrunning the compute it can bring online. Decart CEO Dean Leitersdorf has been blunt about the pace driving deals like this across the industry: “This market is moving so fast. We get to reinvent every different aspect of the economy in the next 18 to 24 months.”

That capacity crunch is also the throughline connecting this deal to Anthropic's own recent moves. In May 2026, Anthropic closed a $65 billion Series H at a $965 billion post-money valuation — up from the $183 billion its Series F set eight months earlier — with CFO Krishna Rao saying the round would help the company “serve the historic demand we are experiencing.” Days later, Anthropic confidentially submitted a draft S-1 to the SEC, a step toward an IPO the company has said depends on market conditions and carries no committed timeline. And the same week this newsroom covered [Theseus Infrastructure](#/article/anthropic-macquarie-gic-theseus-infrastructure), the data-center venture Anthropic is building with Macquarie and Singapore's GIC rather than owning outright — a company buying compute capacity from every direction it can reach, of which an efficiency-software acquisition would be one more.

Two things this timeline does not settle. First, whether the deal happens at all, at $6 billion or any other number — talks this early routinely change or collapse, and neither company has issued its own statement. Second, whether Anthropic is paying for Decart's chip-efficiency software, its video and world-model research, or its team; Bloomberg's sourcing points at infrastructure, but an acquiring company rarely says in public which asset it valued highest.

Anthropic isn't the only frontier lab in this position. OpenAI confidentially filed its own draft S-1 in June 2026, and both companies are now racing toward public markets while still spending at acquisition scale — a $6 billion purchase is a large line item for a private company to absorb five weeks after its last funding round closed, even one that just raised $65 billion. Set against other 2026 deals this newsroom has tracked, like [Cohere's pending purchase of Aleph Alpha](#/company/cohere) or [Groq's chip-architecture license to Nvidia](#/company/groq), a $6 billion efficiency-software acquisition is large but not unusual for a year in which compute-adjacent M&A has become routine at the top of the industry.

Nothing here is Decart's or Anthropic's own account of the talks — neither company has confirmed the reported price or that a deal is imminent, and the entire acquisition thread rests on sourcing Bloomberg attributes to people familiar with the matter rather than an on-record statement. The two things that are confirmed, in both companies' own words, are what put Anthropic in a position to spend $6 billion in the first place (the Series H) and what it's spending toward next (the S-1) — the acquisition talks are the reported bridge between them.