Amazon is putting forward a $7 billion acquisition offer on the negotiating table, targeting a startup called Decart that has been established for only three years. According to The Wall Street Journal, this e-commerce giant is in talks with Decart for an acquisition worth approximately $7 billion, equivalent to 46.9 billion yuan. Prior to this, NVIDIA and Anthropic also sat at the same table for negotiations, but those talks failed repeatedly, allowing Amazon to take advantage of this abandoned opportunity.
Decart was co-founded by Dean, Orian Leitersdorf, and Moshe Shalev in 2023. It holds a highly valuable and attractive skill in the AI world: cross-chip adaptation software. What it aims to solve is a growing problem in the AI world—models trained by developers are often tightly bound to a specific manufacturer’s chips. To run on another silicon platform, a lot of underlying adaptation work is required. Decart’s software acts like a universal adapter for models, enabling seamless switching and free movement between AI chips from different brands. As tech giants fight fiercely for computational advantages, this underlying capability to flexibly schedule workloads across chips is becoming increasingly important.
What truly attracts Amazon is the hard indicators of this strategy. Decart’s DOS optimization stack claims to enable models to run up to 8 times faster on the same hardware. The list of compatible devices includes NVIDIA GPUs, Google TPUs, and Amazon’s own Trainium. In other words, by acquiring this company, Amazon aims to reduce the cost per token of Trainium to a level lower than that of NVIDIA hardware. It did not start from scratch in this technology—AWS gave early access rights to Trainium3 to Decart as early as 2025, and Decart, in turn, adjusted its model architecture to fit the Trainium chip design. The two parties have already worked together at the underlying level.
In addition to that “chip replacement” software, Decart also set up real-time world models: Lucy is designed for immersive experiences like live video, while Oasis serves physical AI and robot simulation. For a giant company that wants to sell cloud services as well as develop embodied intelligence, this subsidiary line of business is also a plus.
The amount of this offer is also worth considering. $7 billion is about 75% higher than the valuation Decart set in May this year—at that time, it was valued at around $4 billion with a financing of $300 million. According to PitchBook data from research institutions, it had previously raised over $400 million from venture capitalists such as Sequoia and Benchmark. In other words, Amazon is adding a significant premium to the ticket price it had a few months ago.
Even more dramatic is the background of this succession. As a close AI partner to Amazon, Anthropic also targeted Decart, aiming to use its technology to make existing computing infrastructure handle more demands, but abandoned a deal worth about $6 billion after due diligence; NVIDIA’s negotiations also failed. The companies that neither got were now taken over by Amazon with a higher offer, meaning they ended up taking the empty seat right after their competitors turned around.
However, whether this deal can truly pay for itself remains a question. Analysis suggests that it will be beneficial to AWS if Decart’s acceleration effect is consistently reflected in the customers’ own models, rather than just in their own video systems—once the universality across models fails, the 8-fold figure might turn into nothing but a disappointment on the financial statement. When a company based on selling cloud and computing power is willing to spend $7 billion to acquire software that allows models to "change chips," what it wants is clearly not just a tool, but to more firmly integrate those customers whose services are tied to chips on its cloud into its own ecosystem.