Decart’s pitch has been narrow and technical: software that makes it easier for developers to move the same model between different kinds of artificial intelligence chips, and models that generate and simulate realistic three-dimensional environments. To Amazon, the first of those now looks less like a tool and more like infrastructure.
The Wall Street Journal reported on the evening of October 11 that Amazon is in talks to acquire the startup, in a deal that could value it around $7 billion. The negotiations could be wrapped up soon if they do not fall apart. Decart was valued at nearly $4 billion in May, and has raised more than $400 million from investors including Sequoia, Benchmark and Radical Ventures.
Amazon Web Services gave Decart early access to its own chips in 2025, a step consistent with the group’s strategy of selling its Trainium3 processors and reducing its dependence on Nvidia. Buying the company would put the chip effort and the portability layer in the same building.
The strategic case turns on lock-in. Nvidia’s advantage has never rested on silicon alone; it is the software stack that makes its hardware the default target for every model written anywhere. A layer that lets the same workload run across different chips attacks that default from the side, which is worth a great deal to a challenger whose processors need developers more than developers need its processors.
Chip portability has become an urgent subject because the cost of compute has stopped falling fast enough to absorb demand. Companies that train and serve models want the option to move a workload from one vendor’s accelerators to another’s as prices, supply and power availability shift. Software that makes that movement routine weakens the strongest moat in the industry, and the company that owns it sits between every buyer and every seller.
The deal has a second buyer in its recent history. Anthropic, the developer of the Claude chatbot, discussed acquiring Decart for about $6 billion and completed due diligence before deciding in September not to proceed. That decision left the field open, and it carries an unusual symmetry: Amazon has invested billions in Anthropic and is the main source of the computing power it uses.
Decart was founded in September 2023 by three Israeli engineers, brothers Dean Leitersdorf and Orian Leitersdorf, and Moshe Shalev. Funding records show a $21 million seed round in October 2024 and a $300 million round in May 2026 that set the nearly $4 billion valuation, with Radical Ventures, Nvidia, eBay Ventures and Adobe Ventures among the backers. Nvidia’s presence on that list is notable, since the largest chipmaker in the industry also funds companies built to make its chips interchangeable.
The company’s work sits at the junction of two markets that have converged this year. One is inference software, which the company says raises the utilization of expensive graphics processors and speeds up generative media work. The other is world models, systems that generate interactive environments on the fly, which Decart has marketed for robotics, autonomous vehicles, manufacturing and drones, along with tools that edit video live.
That second business explains why the valuation moved so quickly. Training a robot to move through a warehouse requires either a physical warehouse or a simulated one, and simulation is cheaper. Roboticists call the gap between the two the reality gap, and a model that generates physically plausible environments can narrow it. Amazon Web Services released an open-source stack from its own warehouse robots this week, part of a push to let outside companies train machines on data it has collected.
The world-model route is expensive, however, and investors have shown less patience this year for labs without revenue. A buyer with a cloud, a chip line and a robotics division can carry that cost where a standalone startup cannot.
The customer base would also change shape. Decart has sold to developers who want their models to run anywhere, a constituency that includes buyers with no stake in Amazon’s success. Inside AWS, the same software becomes a reason to choose Trainium, and the roadmap bends toward the priorities of one cloud rather than the neutrality that made the company worth buying. Founders who sell on portability often find the buyer’s incentive runs the other way.
Price is the open question. A $7 billion valuation is a large premium over the mark Decart carried in May, and it would rank among Amazon’s biggest acquisitions in recent years, in a period when large technology deals draw scrutiny from regulators in Washington and Brussels. Whether the parties can close without a long review is not something either side has addressed publicly.
For Amazon the logic is straightforward even if the price is not. Cloud customers pick a cloud partly because of the chips it offers, and the ability to switch between them lowers the risk of choosing. Whoever owns the layer that makes switching cheap holds a piece of every purchase decision, a modest-sounding business until the whole industry is arguing about the price of compute.
