Amazon is putting a $7 billion acquisition on the table, targeting a three-year-old startup called Decart. According to The Wall Street Journal, the e-commerce giant is in talks with Decart over a potential valuation of about $7 billion, or approximately 46.9 billion yuan; before that, both NVIDIA and Anthropic had sat at the same table, but their negotiations failed, leaving Amazon to pick up the discarded piece.

Decart was founded in 2023 by Dean, Orian Leitersdorf, and Moshe Shalev. It holds a skill that is particularly valuable and highly coveted by big tech companies today: cross-chip compatibility software. What it aims to solve is an increasingly thorny problem in the AI world — trained models are often locked to a specific vendor's chip, making it necessary to redo a lot of low-level adaptation if they want to run on another brand of silicon. Decart's software acts like a universal adapter for models, allowing them to switch seamlessly between different AI chips. As tech giants compete fiercely for computing power, this ability to flexibly schedule workloads across chips is becoming increasingly critical.

What really caught Amazon's attention is the hard metrics of this approach. Decart's DOS optimization stack claims to make models run up to eight times faster on the same hardware, and its compatibility list includes NVIDIA GPUs, Google TPUs, and even Amazon's own Trainium. In other words, acquiring this company could mean Amazon is aiming to lower the cost per token of Trainium below that of NVIDIA hardware. Amazon isn't starting from scratch in this area — AWS had already given Decart early access to Trainium3 in 2025, and Decart in turn adjusted its model architecture to fit the design of Trainium, meaning the two had already aligned at the fundamental level.

In addition to its "chip-switching" software, Decart also built a real-time world model: Lucy is aimed at live video and immersive experiences, while Oasis serves physical AI and robot simulation. For a giant that wants to sell cloud services and also develop embodied intelligence, this side project is another plus.

The amount of this offer is also worth noting. At $7 billion, it is about 75% higher than Decart's valuation in May of this year — when it was valued at about $4 billion and raised $300 million. According to data from research firm PitchBook, it had previously raised more than $400 million from venture capital firms such as Sequoia and Benchmark. In other words, Amazon is essentially adding a significant premium to the price paid just a few months ago.

Even more dramatic is the context of this relay. As a close AI partner, Anthropic had also targeted Decart, wanting to use its technology to handle more demands on its existing computing infrastructure, but abandoned a $6 billion deal after due diligence. Similarly, NVIDIA's negotiations also ended without success. A company that both failed to acquire is now being taken over by Amazon with a higher offer, effectively taking the seat left empty after the competitors turned away.

However, whether this deal will truly pay off remains uncertain. Analysts point out that it would benefit AWS only if Decart's acceleration effect can be consistently demonstrated in customers' own models, not just within its own video system — once the universality of the cross-model performance proves shaky, the eightfold figure could shrink into an empty joy on the balance sheet. When a company that is based on selling cloud and computing power is willing to spend $7 billion on a software tool that helps models "switch chips," it's clearly not just looking for a tool, but rather a way to lock more customers who are tied to chips into its own ecosystem more firmly.