Amazon in advanced talks to buy Decart for $7 billion to lower Trainium costs

The Wall Street Journal reports Amazon is in advanced talks to acquire Decart for about $7 billion after earlier negotiations with Nvidia and Anthropic collapsed. The startup, founded in 2023 by Dean and Orian Leitersdorf and Moshe Shalev, develops software that lets developers move AI workloads across different chips — a capability Amazon sees as key to driving the cost per token of its Trainium silicon below that of Nvidia hardware.

Decart's core product is DOS, an optimization suite the company claims accelerates models up to 8x on the same hardware, including Nvidia GPU, Google TPU, and Amazon's own Trainium. If those numbers hold up in real-world tests, Amazon could offer AWS customers a genuine economic alternative to the Nvidia ecosystem without swapping hardware. Amazon's advantage is not accidental: AWS gave Decart early access to Trainium3 as early as 2025, and the company tailored its model architecture to Amazon's specific chip design.

Decart is not stopping at inference optimization. The company is also building real-time world models: Lucy for immersive experiences such as live video, and Oasis for physical AI and robotics simulation. Both products serve as proof points for the optimization stack; they run on DOS and demonstrate what the hardware can deliver when software knows how to squeeze it. The open question is whether that acceleration translates to third-party customer models, or only to Decart's own video and simulation systems.

The discussed price implies a roughly 75% premium over the valuation Decart set in May, when it raised $300 million at a valuation of about $4 billion. Anthropic, a close Amazon AI partner, wanted the technology to help its existing compute infrastructure meet growing demand but walked away from a $6 billion deal after due diligence. Nvidia also explored an acquisition and dropped out. The fact that two strategic players declined after deep examination should raise a red flag — or at least remind us that the gap between an impressive demo and production at global scale has not yet closed.

The deal makes sense for AWS only if the speedups Decart demonstrates on its own systems hold when running third-party models on the stack, from Llama to proprietary enterprise models. Without proof that the optimization translates to a lower cost per token across a variety of real workloads, Amazon would be buying mainly a talented team and IP that has not yet been proven in mass production. Talks have not yet produced a signed agreement, and as Anthropic and Nvidia showed, due diligence is where deals like these fall apart.