Amazon Wants to Buy the Software That Makes Chips Interchangeable

Two months ago, Decart was the company Anthropic wanted. Now it may be the company Amazon gets.

The Wall Street Journal reported on October 11 that Amazon is in talks to acquire Decart for roughly $7 billion, citing people familiar with the matter. A deal could come soon, the Journal said, unless talks fall apart or another suitor emerges. Amazon declined to comment and Decart did not respond, according to Reuters' write-up of the report.

We covered the Anthropic side of this story in August, when Bloomberg reported a deal near $6 billion and the logic was margins: a lab heading toward a listing buying more work per GPU. Bloomberg then reported on September 8 that Anthropic had finished due diligence and walked away. Nobody has said why.

Amazon is a different buyer with a different problem, and the price says it values the asset more, not less, after a suitor looked under the hood and left.

The product that matters

Decart is best known for Lucy, a model that edits live video in real time, and Oasis, which generates simulated environments for training robots and self-driving systems. Neither is why this is a $7 billion conversation.

The core product is DOS, the Decart Optimization Stack. It lets developers run AI workloads on Nvidia GPUs, Google TPUs or Amazon's Trainium without rebuilding for each one. SiliconANGLE, as cited by Tech Funding News, reported that tuning a model for a new chip used to take months and DOS can cut that to a couple of weeks.

That sentence is the whole deal. The hardest thing about selling a chip that is not made by Nvidia has never been the silicon. It is the months of porting work every customer has to do before the silicon is useful, and the CUDA software ecosystem that makes staying on Nvidia the path of least resistance.

Decart is Amazon's best Trainium story

Decart did not start out hardware-agnostic. Calcalist reported in December 2025 that the company ran entirely on Nvidia GPUs until an Amazon representative approached CEO Dean Leitersdorf at a conference and talked the team into testing Trainium.

The results became a stage moment. At re:Invent in December 2025, AWS said Decart was getting 4x faster frame generation at half the cost of GPUs on Trainium3. Leitersdorf told Calcalist the best Decart had managed on GPUs was 26 frames per second, and on Trainium it was approaching 101. Those are vendor figures, not independent benchmarks, but they were the kind of figures Amazon needed someone outside Amazon to say.

Calcalist also reported that Decart had become one of Amazon's largest AI-chip customers after Anthropic. In May 2026, according to Tech Funding News, Amazon became a strategic customer under a commercial agreement and joint go-to-market plan.

So Amazon is not buying a stranger. It is buying the startup that proved, in public, that a frontier-grade real-time workload could leave Nvidia for Trainium and get faster and cheaper on the way out.

Why the price went up

Decart raised $300 million in May at a valuation near $4 billion, in a round led by Radical Ventures with Nvidia joining as an investor. Tech Funding News put the reported $7 billion at about a 75% premium to that mark, inside five months.

For Anthropic, Decart's software was a cost lever on its own inference bill. Useful, but bounded by how much compute one lab buys.

For Amazon, the lever is the size of the Trainium market. Every customer that can move a workload onto Trainium in two weeks instead of six months is a customer AWS can price against Nvidia instances. The value of DOS to Amazon scales with every AWS customer, not just one.

There is precedent for Amazon paying for exactly this kind of leverage. In 2015 it bought Israeli chip designer Annapurna Labs for a reported $350 million, a price Amazon never confirmed. Annapurna's engineers went on to build Graviton, Inferentia, Nitro and Trainium. Annapurna gave Amazon its own chips. Decart would give those chips a faster on-ramp.

The neutrality question

Here is the tension. DOS is valuable because it is chip-neutral. The pitch to a developer is: write once, run on whatever silicon is cheapest this quarter.

Under Amazon ownership, that promise gets harder to believe. The obvious move is to keep supporting Nvidia and Google TPUs, because a tool that only targets Trainium is just another proprietary SDK, and developers already have one of those from Nvidia. But "supports every chip" and "works best on Trainium" can both be true at once, and the second one is where an owner's incentives point.

Google faces the mirror problem. A widely used layer that makes TPU workloads portable would sit inside a competitor's cloud. Nvidia, which Calcalist reports was in advanced talks to buy Decart itself before higher bids arrived, would watch a company it invested in become the toolkit for moving customers off its hardware.

Leitersdorf summed up the founder's view to Calcalist last year: "Customers don't care what chip we use." Amazon is betting $7 billion that customers will care a great deal about how cheaply they can switch.

What to watch

The deal is not signed, and this story has already lost one buyer after diligence. Three things will show what Amazon is actually buying.

Whether TPU and GPU support survives. If DOS keeps shipping first-class Nvidia and Google targets after close, Amazon is buying an industry tool and betting Trainium wins on price. If those targets go quiet, it was a moat purchase.

What happens to Lucy and Oasis. Real-time video and robot-training simulation fit neatly with AWS's Bedrock catalogue and Amazon's robotics ambitions. Keeping them would suggest Amazon wants the models as well as the stack.

Whether the speed-ups hold up outside a keynote. Dealroom has called Decart's claims of 8x speed-ups and 100x cost efficiency self-reported and hard to verify independently. Anthropic had access to the numbers and left. Amazon has its own data from running Decart on its chips for a year, which may be the one thing it knows that the last bidder did not.