Amazon Reportedly Nears a $7 Billion Deal for AI Startup Decart

The company is said to be in advanced talks as it looks to loosen Nvidia’s grip on AI computing

Amazon is in advanced talks to buy the AI startup Decart for around $7 billion. Nothing is signed. The numbers could shift, and the deal could still fall apart. But even as a report, it says plenty about where the big cloud companies are putting their money.

Most of the AI conversation still revolves around chatbots and who has the smartest model. This story is about something less glamorous. It’s about the chips and the software that decide how well those models actually run.

What Decart actually does

Decart was valued at nearly $4 billion in May 2026, when it raised $300 million in a round led by Radical Ventures. Nvidia took part too, which is a little ironic given what Amazon appears to want from the company. A $7 billion price would add roughly $3 billion in value within a few months, an increase of about 75%.

The company has two sides. One is generative AI models. The other is software that lets AI workloads run across different types of chips. On the model side, Decart is known for Lucy, which edits video in real time, and Oasis, which generates interactive simulated environments. 

Neither is about typing a prompt and getting a paragraph back. They produce scenes and worlds that react as you use them. The company also builds models that generate and simulate three-dimensional spaces, which matters for robotics, autonomous driving research, and any field where a machine has to learn a space before it enters the real one.

The second side is probably what drew Amazon in, and it gets less attention.

Why the Chip Software Matters Most

Nvidia’s GPUs dominate AI training and inference. Part of the reason is hardware, but a large part is habit. Developers have spent years writing code around Nvidia’s tools, and moving off that platform can mean rewriting pieces of a project, retesting performance, and accepting risk that nobody on the team wants to explain to a manager. So people stay, even when a cheaper option exists.

Decart’s software reportedly lets developers move between artificial intelligence chips and tune workloads across different hardware. If that works well in practice, it shrinks the cost of leaving Nvidia. For a company that sells its own chips, that’s a big deal. A fast processor nobody wants to program for doesn’t earn much.

Where Trainium Fits

Amazon has been designing its own chips for years, including Trainium for AI training. The logic is simple. Depending on one outside supplier for the most important part of modern computing is expensive, and it leaves you exposed when supply tightens or prices climb.

Building the chip is the easier half, though. Getting customers to trust it is harder. Amazon Web Services can offer strong silicon, but a developer with a deadline will often pick the familiar option. Software that smooths the switch could change that math, turning Trainium from something teams consider into something they actually use.

Cost matters here as well. Running heavy AI applications is pricey, and cloud providers feel pressure to improve performance without inflating customers’ bills. Anything that squeezes more out of existing hardware goes straight to margins.

The Robotics Angle

It would be a mistake to file Decart under chip tools and stop there. Amazon runs one of the largest logistics and robotics operations anywhere, and simulation is a natural fit. Training machines in a realistic virtual space is cheaper and safer than doing it in a live warehouse. Autonomous driving research works the same way, since simulated roads let engineers test rare or dangerous situations without putting anyone at risk. The report doesn’t say how Amazon would use these models, so this part is informed guesswork. But the fit is easy to see, and it may explain why the price is as high as it is.

Anthropic Looked & Passed

One of the more interesting details is that Anthropic reportedly explored buying Decart too, at a valuation of about $6 billion, and then decided not to go ahead. The report doesn’t give a reason. There are several plausible ones. The price might have felt too steep. The strategic fit might have looked weaker up close. Or the company may have preferred to spend its money elsewhere. None of that is confirmed. What we do know is that at least two major players took a serious look, which suggests Decart’s technology has real substance behind it.

A Race for The Whole Stack

Zoom out and the deal fits a pattern that’s been building for a couple of years. The biggest tech companies want to own every layer of the AI stack, and they’re pursuing it through acquisitions, investments, and partnerships all at once.

Compute is the center of that race. Advanced models need enormous processing power, and the supply of top chips is limited. Whoever can offer a credible alternative, or make existing hardware work harder, gains leverage over both rivals and customers.

Chips get most of the headlines, but optimization software is just as important. A faster processor does little good if the code running on it wastes capacity. As models get larger and costlier to operate, companies that handle that middle layer well could capture a surprising amount of value.

Nvidia isn’t sitting still either. Its investment in Decart suggests it sees something worthwhile in the startup, even though the same technology could eventually make Nvidia’s own chips easier to replace. That kind of overlap is common in this industry. Today’s partner is often tomorrow’s competitor.

What it Could Mean for Developers

If Amazon closes the deal and builds Decart’s tools into AWS, developers might get an easier route to Trainium and other chips. That could bring lower costs and more choice. Smaller companies that struggle to get the latest Nvidia hardware could benefit most.

There’s a catch, though. Will the software stay neutral and keep supporting a wide range of chips, or will it start leaning toward Amazon’s own silicon? Customers who value flexibility will want a clear answer. Acquisitions can change a product’s direction quickly, and developers who rely on Decart today will be watching closely for any shift.

Risks Worth Keeping in Mind

The deal isn’t final, and the price could move. The report relies on sources, and neither company has publicly laid out a plan. Big technology acquisitions can also attract regulators, especially when a dominant cloud provider buys into a fast-growing field like AI.

Integration is another hurdle. Startups move fast and take risks that are harder to sustain inside a giant company. Keeping the engineers and researchers who built the technology often decides whether an acquisition succeeds or turns into an expensive lesson. At $7 billion, that matters even more.

Then there’s the price itself. Paying close to double the recent valuation leaves little room for error. Amazon would be betting that Decart’s tools will matter far more in a few years than they do today. That could prove right. It could also prove expensive.

What to Watch Next

First, confirmation. A signed agreement would force both companies to explain their plans, and that would answer a lot of questions about chip support and product direction. After that, watch AWS. New tools that make it easier to move workloads onto Trainium would suggest the purchase is doing its job. How existing Decart customers are treated will say a lot as well. Rivals will be paying attention. If chip-flexible software turns out to be valuable, other cloud providers and chipmakers may go looking for similar deals.

The Takeaway

Amazon’s reported interest in Decart goes beyond one startup. It points to an effort to cut reliance on Nvidia, make custom chips easier to adopt, and own the software that determines how AI hardware performs. The video and simulation models add a second possible payoff, especially for robotics.

Whether the deal happens, and at what price, is still unknown. But the reasoning behind it is easy to follow, and the contest over the foundations of AI computing is nowhere near finished.