Why Anthropic May Pay $6 Billion for Decart
20 August, 2026
Bloomberg says the Claude developer is in talks to acquire Israel’s Decart. Its optimization technology could address one of Anthropic’s biggest costs: compute
Above: Decart’s two flagship models: Oasis, which generates interactive worlds in real time, and Lucy, which enables real-time live video transformation. Both run on the company’s DOS optimization infrastructure. Source: Decart AI
AI company Anthropic is in talks to acquire Israeli startup Decart AI for approximately $6 billion, according to Bloomberg. Reuters separately confirmed through a source that discussions are underway, although no deal has been finalized and there is no certainty that one will be reached.
If completed, it would be a massive transaction for a company that only in May raised $300 million at a valuation of approximately $4 billion. Nvidia and SpaceX have also previously been reported as showing interest in Decart. But the more interesting question than the price tag is what exactly Anthropic sees in Decart that could be worth billions of dollars.
At first glance, Decart may look like a generative AI company specializing in video and “world models.” The company developed Lucy, which enables users to transform live video through text instructions — replacing characters or backgrounds and adding objects and effects, for example — while maintaining a frame rate of 30 frames per second.
It also developed Oasis, an interactive world model designed, among other applications, for simulations involving robots and autonomous vehicles. Unlike a conventional simulation engine, the model itself generates the environment and how it evolves in response to actions taken by a user or AI agent.
But these products tell only half the story. Underneath them lies what may be the technology of greatest interest to Anthropic: the Decart Optimization Stack, or DOS.
Decart’s premise is that a significant portion of the computing power available in AI accelerators is not being used optimally. DOS is a software layer that performs deep optimization across the stack — from hardware-aware model design to kernels, memory management, compilers and inference optimization.
The technology supports multiple architectures, including Nvidia GPUs, Google TPUs and AWS Trainium. The objective is straightforward: extract more work from the same chip, allowing AI models to run faster and at lower cost.
Decart already offers these capabilities as a service to chipmakers, hyperscalers and AI labs, allowing customers to bring their own workloads and have kernels and compilers optimized for them. It has also launched Cogito, an inference layer for open-source language models.
According to Decart, its technology can in some cases deliver speeds roughly five times faster than a typical inference provider and exceed 1,000 tokens per second in certain configurations. These are company-reported figures, but they illustrate the kind of asset Anthropic may be targeting.
Getting More From Every GPU
For Anthropic, the technology addresses one of the biggest challenges in the AI business: the enormous cost of compute.
As Claude evolves from a model that simply answers questions into a system of AI agents that perform lengthy tasks, write code, operate tools and repeatedly call the model to “think” about their next step, inference demand rises dramatically. Any improvement in throughput or accelerator utilization can therefore translate directly into substantial cost savings.
Reuters estimated that if Anthropic’s compute spending were to reach approximately $56 billion under a scenario in which the company generates $100 billion in revenue, even a 10% improvement in compute efficiency could theoretically be worth $5.6 billion annually. Viewed from that perspective, a $6 billion price tag for a company capable of materially improving infrastructure efficiency begins to look considerably less surprising.
There is another clue. According to reports about the negotiations, Decart employees are expected to join Anthropic’s Inference and Performance organization. If that happens, it would be a strong indication that the primary objective is not to turn Claude into a video-generation model, but to bring Decart’s expertise in AI model optimization directly into Anthropic’s infrastructure.
From Unit 8200 to Claude’s Infrastructure
Decart was founded in 2023 by Dr. Dean Leitersdorf and Moshe Shalev, who met while serving in Unit 8200, the Israeli military’s elite technology and intelligence unit.
Leitersdorf, the company’s CEO, holds three computer science degrees from the Technion – Israel Institute of Technology, including a PhD, while Shalev serves as chief product officer. The company has raised funding from prominent investors including Sequoia, Nvidia, Benchmark, Radical Ventures and Zeev Ventures.
In its latest funding round in May, Decart explicitly presented DOS as the infrastructure underlying its two main model families: Lucy for interactive experiences and Oasis for Physical AI.
That may be the best way to understand Decart: not as a video AI company that happens to have developed optimization technology, but as an AI optimization company that built real-time video and world models to demonstrate what its infrastructure can do.
The company describes its ambition simply: to make AI operate “at the speed of reality.”
For Anthropic, which is competing in a market where the cost of running a model is becoming almost as important as the model’s capabilities, that technology could be worth a great deal of money.
Posted in: News
Posted in tags: Decart AI
