Amazon is in talks to buy Decart for $7B, WSJ says

Amazon may be about to pay $7 billion for the plumbing that lets AI models run on any chip. The chips are the reason.

The Wall Street Journal reported on Sunday that Amazon is in talks to buy Decart, an Israeli AI startup, in a deal that could value it at about $7 billion. A deal could come within days, the Journal reported, unless talks collapse or another bidder appears, according to SiliconANGLE's account of the report. No company has commented on the record.

Decart sells chip portability

Decart is two businesses. It is best known for real-time video and world models, Oasis and Lucy, which generate and simulate 3D environments for uses like video games and robotics training.

The part Amazon wants is probably the other one. Decart's software, which SiliconANGLE calls DOS, tunes neural networks for specific processors. SiliconANGLE says it cuts that work from months to about two weeks and lets companies move AI workloads between chips when needs or supply change.

CTech reports that Decart's technology runs on Nvidia, Google and Amazon hardware, and that industry estimates put the speedup at "up to eight times faster" depending on workload. That number is an estimate, not a benchmark anyone published.

The Amazon ties are already deep. CTech says Decart's real-time video model ran four times faster on Amazon's Trainium chip in December 2025, after which Amazon became a strategic customer and gave Decart early access to Trainium3.

Decart raised $300 million in May at a $4 billion valuation, with total funding past $400 million, per SiliconANGLE.

$7 billion would be 1.75 times that valuation, five months later.

This is a Trainium story

Here's the thing about Amazon's chips. They only win your workload if your model runs well on them. Every week a team spends porting a model to a different chip is a week it does not spend shipping.

Software that makes the port fast changes the math. If AWS can promise that your model runs well on Trainium in two weeks instead of two months, more of you will rent Trainium. CTech makes the same argument: better software could let AWS get more compute out of the chips it already has.

There is a precedent. Amazon bought Israeli chip designer Annapurna Labs in 2015 for $350 million, CTech notes, and that deal laid the foundation for its in-house silicon.

Decart has also been shopped before. SiliconANGLE notes that Bloomberg reported last month that Anthropic walked away from a reported $6 billion deal after due diligence. Nobody has said why. Nvidia was also involved in earlier acquisition talks, CTech reports.

What it means for you

If you rent GPUs, nothing changes this week. Nothing is signed.

If you are choosing a cloud for inference next year, this is the thread to pull. The fight over chips is now also a fight over the software that moves models between them, and the prize is your workload. That cuts both ways. Portability you can use to cut costs is portability AWS can use to win you over.

Ask your provider a plain question: how long would it take to run our model on your own silicon, and who does the work?

What to watch

Whether a deal is announced this week, as the Journal suggested it could be. Whether Decart's tools stay available on Nvidia and Google hardware afterward, or become an AWS-only advantage.

Nvidia is reportedly in talks over Reflection AI. Amazon wants Decart. The big buyers are shopping for software now.

Watch who buys the next one.