Anthropic Offers $6 Billion to Acquire Decart, Betting on AI Compute Optimization

When a leading AI lab drops $6 billion on a three-year-old startup, the immediate assumption is that they are buying a flashy new model or a consumer application. But Anthropic’s advanced negotiations to acquire Israeli startup Decart for exactly that sum are not about adding a new chatbot to its roster. Instead, the maker of Claude is purchasing the plumbing required to make its existing models financially sustainable. The proposed acquisition, which would be the largest in Anthropic's history, targets Decart's core asset: a vertically integrated optimization stack designed to make AI training and inference drastically faster and cheaper.[1][5]

The $6 billion price tag represents a 50% premium over Decart’s $4 billion valuation from just three months ago, when the company raised $300 million in a round backed by Nvidia and Radical Ventures. If finalized, Decart’s engineering team will be absorbed into Anthropic’s inference and performance organization. The aggressive move comes as Anthropic reportedly prepares for an initial public offering and faces the crushing infrastructure costs required to serve surging enterprise demand for its Claude models.[1][2][3][5]

The mechanics of the deal highlight a shifting reality in the artificial intelligence sector: the primary bottleneck is no longer just algorithmic breakthroughs, but hardware utilization. Decart’s flagship product, the Decart Optimization Stack (DOS), operates at the lowest levels of computing architecture to improve Model Flops Utilization (MFU). MFU measures how much of a chip's theoretical computing power is actually being used during an AI workload. By optimizing memory management, compilers, and kernels, Decart's software allows AI developers to extract peak performance from existing hardware across Nvidia GPUs, Google TPUs, and Amazon Trainium chips.[1][6]

"If you can raise MFU from roughly 25% to 50%, it effectively doubles the usable compute capacity of an existing fleet without adding more chips or drawing more power," noted Harvard engineering professor Vijay Janapa Reddi. At the gigawatt scale where Anthropic operates, that efficiency translates directly into billions of dollars in unlocked capacity and deferred capital expenditure. For a company spending heavily to secure computing clusters, acquiring the technology to make those clusters run twice as efficiently is a strict mathematical calculation.[1][5]

Beyond raw compute optimization, the acquisition also hands Anthropic a foothold in the emerging field of "world models." While Decart is fundamentally an infrastructure company, it has publicly demonstrated its optimization capabilities through generative video models like Oasis and Lucy. These models are capable of modifying live video feeds and simulating interactive physical environments in real time with sub-35 millisecond latency. While Claude currently excels at text and code generation, integrating Decart’s real-time video architecture could accelerate Anthropic’s push into multimodal AI and autonomous agents that interact with dynamic visual environments.[5][6][7]

The negotiations also underscore the escalating arms race for AI infrastructure talent. Decart was founded in late 2023 by Dean Leitersdorf and Moshe Shalev, who met in Israel's elite military intelligence unit 8200. The startup quickly attracted attention from major tech players, with reports indicating that Nvidia and SpaceX had both explored acquiring the company before Anthropic stepped in with a higher offer. By securing Decart, Anthropic not only gains critical optimization technology but also establishes its first major research and development hub in Israel, tapping into one of the world's densest AI talent ecosystems.[3][4][5]

While the deal remains in the preliminary stages and could still fall apart, it signals a maturation in how frontier AI labs allocate capital. Rather than solely chasing the next generation of massive models, companies are increasingly investing in the unglamorous but vital software layer that makes those models economically viable to operate at scale. For Anthropic, a $6 billion bet on efficiency is a clear indicator that the next phase of the AI boom will be defined as much by profit margins and compute economics as by raw intelligence.[1][2][5]

The timing of the acquisition aligns with Anthropic's broader corporate strategy as it reportedly moves toward the public markets. Having confidentially submitted an S-1 registration filing to the U.S. Securities and Exchange Commission earlier this year, Anthropic is under pressure to demonstrate a sustainable business model. By bringing Decart's optimization stack in-house, Anthropic can present prospective public investors with a clear roadmap for reducing its reliance on expensive hardware upgrades and improving the unit economics of every API call made to Claude.[1][2]