Amazon in Advanced Talks to Buy AI Startup Decart for About $7 Billion

The Wall Street Journal reported on Sunday, October 11, 2026 that Amazon is in advanced talks to acquire Decart, an Israeli AI startup, in a deal that could value the three-year-old company at around $7 billion. The Journal, citing people familiar with the discussions, said an agreement could be reached soon and that another bidder could still emerge. SiliconAngle carried the same figure the following day. Neither Amazon nor Decart has confirmed the talks, and both declined to comment.

Related reading: Amazon's AWS Grew 39% While Free Cash Flow Fell to -$42.4 Billion · Nvidia Puts $3 Billion Into Meta's Equity Sale and Expands the Muse Partnership.

The price would make it one of Amazon's largest acquisitions, and one of the biggest exits in Israeli tech. But the logic behind the deal is narrower than the headline number. Amazon has spent years and billions building Trainium, its custom AI accelerator, and the chip's hardware case is not the problem. The problem is software. Most AI models are written for Nvidia's CUDA platform, and moving that work to Trainium normally takes months of engineering. Decart sells the software that shortens the move.

What Decart actually sells

Decart's central product is the Decart Optimization Stack, which the company calls DOS. It is not a model. It is a platform that combines hardware-aware model design, proprietary compilers, custom kernel tooling and inference optimisation into one stack, and it runs across Nvidia GPUs, Google TPUs and Amazon's Trainium chips, according to Tech Times. The pitch is portability: a team can take a model built for Nvidia and get it running efficiently on other silicon without rebuilding it from scratch. Decart says DOS compresses a process that normally takes months into weeks, a claim about its own software rather than an independently measured result.

That claim matters more every year, because the money in AI has shifted from training to inference. Training a frontier model is a large, occasional expense. Serving that model to users runs continuously, and industry research cited by Tech Times puts inference at 80% to 90% of an AI system's lifetime operating cost, with the inference market projected to grow to $255 billion by 2030 from about $106 billion in 2025. Per-query costs have fallen sharply, from roughly $20 per million tokens in 2022 to about $0.07 by 2025, but at the scale of millions of users the bill is still one of the largest line items an AI company carries. Software that cuts that bill by any meaningful factor is worth more than another speedup on the training side.

The CUDA moat is a software problem

Amazon is not the only company trying to loosen Nvidia's grip on AI compute. Microsoft unveiled the third generation of its Maia chip and said it plans to make Maia the main hardware for running AI in its own products starting in 2027, GeekWire reported. Google runs its TPU programme, and OpenAI has designed a chip with Broadcom. All of them face the same obstacle. Nvidia spent about 15 years building CUDA: the compilers, libraries, debugging tools and frameworks that make its GPUs the path of least resistance for an ML engineer. That ecosystem, more than the silicon, is why Nvidia keeps its share of the accelerator market even when a rival's price-performance looks better on paper.

AWS's equivalent, the Neuron SDK, is improving but still trails CUDA in maturity, and Tech Times reports that AWS engineers understand internally that CUDA familiarity keeps developers on Nvidia hardware. DOS is meant to sit in that gap. Rather than requiring engineers to learn Neuron deeply, it automates the kernel selection, memory-layout and scheduling work that a port normally demands. For an enterprise weighing a shift of inference workloads from Nvidia to Trainium, that turns a major engineering project into a manageable transition.

What $7 billion would buy is the software layer that makes a chip Amazon has already built easier to sell.

The ties between the two companies are not new. AWS gave Decart early access to Trainium3 in 2025, and the pair partnered in December 2025. At AWS re:Invent in Las Vegas that month, Decart's real-time video model turned AWS senior vice president Peter DeSantis into a live cartoon of Amazon CTO Werner Vogels on the keynote screens, an animation generated on Trainium3, GeekWire reported. Amazon said at the time that Decart was running real-time video four times faster on Trainium at half the cost of GPUs, and the company said running Decart's Lucy2 on Trainium3 reached 80% Model FLOPs Utilisation, a measure of how efficiently software uses a chip's theoretical peak.

Who else wanted Decart

Amazon is late to this particular queue. Bloomberg reported on August 13, 2026 that Anthropic, the Claude developer, was in advanced talks to acquire Decart for about $6 billion, which would have been its largest known acquisition. Nvidia had held acquisition talks earlier, CTech reported, and SpaceX was described as circling the company. Then, on September 8, Bloomberg reported that Anthropic had completed due diligence and walked away, without a public explanation.

The most common explanation offered at the time was timing. Anthropic was preparing an initial public offering, with marketing expected to begin in mid-October, and a large stock-based acquisition just before listing would have complicated its balance sheet. Globes, citing sources close to Decart, reported a different factor: the founders resisted Anthropic's push to move much of the team from Tel Aviv to San Francisco. Neither version has been confirmed by either company.

The bidding history is part of why the price sits where it does. Decart raised $300 million in May 2026 in a round led by Radical Ventures, with Nvidia joining as a new investor, at a valuation of nearly $4 billion, per PitchBook data cited by the Journal. A $7 billion price is roughly a 75% premium to that round, five months later, for a company that has not disclosed its revenue. Tech Times reported that the round also drew eBay Ventures, Adobe Ventures and Toyota Ventures, along with returning backers Sequoia Capital and Benchmark.

What Amazon gets beyond the chips

Amazon would not be buying DOS for Trainium alone. The same stack runs on Nvidia GPUs and Google TPUs, so an Amazon-owned version could sit between AI developers and every major chip platform, a bigger position than any single-hardware play. Amazon also has businesses the models fit. Decart's Lucy video model can show online shoppers how clothes look on them, and Twitch, which Amazon owns, uses real-time video technology of this kind. The company's Oasis world model, released with the chip startup Etched in October 2024 as an AI-generated playable version of Minecraft, points at the simulation and synthetic-data work now central to robotics.

The team is a large part of the purchase. Decart was founded in 2023 by three engineers from Israel's Unit 8200 intelligence unit: brothers Dean and Orian Leitersdorf and Moshe Shalev. Dean completed a computer science doctorate at the Technion at 23, a record later broken by Orian. The company has about 100 employees and offices in San Francisco, New York and Tel Aviv. GeekWire notes that Decart has not disclosed revenue but has been profitable almost from the start, according to the business publication Globes, and that its customers include eBay, itself an investor.

For scale, Amazon paid $13.7 billion for Whole Foods in 2017 and $8.45 billion for MGM in 2022, and agreed to buy the satellite company Globalstar in April 2026 for $11.57 billion. Its chip business began with the 2015 purchase of the Israeli chipmaker Annapurna Labs for a reported $350 million, a team that went on to design Trainium. A $7 billion software deal would be far larger than the one that started it.

Caveats

The first caveat is the most basic: this is a report, not a deal. The Wall Street Journal's account rests on people familiar with the talks, no agreement has been signed, and the sources said a rival bidder could still appear. Amazon and Decart have said nothing publicly.

The second is that the central performance claims are Decart's, not independent findings. The 1,600 tokens per second, the 8x speed and the 100x cost-efficiency figures come from the company, and no third party has published benchmarks. Amazon's own 80% MFU figure for Lucy2 on Trainium3 is a company measure of one specific model, not a general test of DOS across arbitrary workloads.

The third is price against revenue. Decart has not disclosed its revenue, and CTech reported that the business remains far smaller than the valuation under discussion. At $7 billion for about 100 employees, the price is largely a bet on the team and on software the buyer believes it can spread across its cloud.

The fourth is neutrality. DOS works across Nvidia, Google and Amazon chips, and an open question is whether it keeps supporting rivals' hardware once Amazon owns it. A version restricted to Trainium would be less useful to customers, and a less credible neutral tool, which cuts against the reason for buying it in the first place.

What to watch

The next signal is confirmation. If the talks produce a signed agreement, the filing and the final price will be the first check on the $7 billion figure. If they collapse, Decart returns to a market with a documented group of suitors.

Then comes the software question. Watch for whether Amazon commits to keeping DOS available on Nvidia and Google silicon, or folds it into the Neuron SDK as a first-party optimisation tool. That answer decides whether Amazon bought a neutral layer between developers and chips, or an accelerator for one.

The last is regulatory. A $7 billion acquisition involving a US cloud provider and an Israeli startup would face a review, and the close date will show how quickly it moves. The longer test, whether the price was worth paying, will show up in porting times and inference bills on Amazon's cloud. Neither is visible yet.

Cover image: "Amazon Spheres from 6th Avenue, April 2020.jpg" by SounderBruce, CC BY-SA 4.0, via Wikimedia Commons. Amazon's Seattle headquarters.