
Photo: SounderBruce, 2020, via Wikimedia Commons (CC BY-SA 4.0)
Amazon is close to buying a three-year-old AI startup most people have never heard of, and the price keeps going up. On Saturday, Nov. 21, 2026, the Financial Times reported that Amazon is in advanced talks to acquire Decart in a deal valued at about $7.5 billion, which could be announced as early as next week. Neither company has confirmed the talks, and deals at this stage can still fall apart.
The short answer: Decart makes software that squeezes more work out of AI chips and lets the same model move between Nvidia GPUs, Google TPUs and Amazon’s own Trainium chips. It also builds real-time video and "world" models that edit live video and generate playable 3D scenes. For Amazon, the prize is mostly the first part. Nvidia’s lead rests heavily on its CUDA software, and a team that can make Trainium as easy and as fast as a GPU would help Amazon sell more of its own chips. The reported structure, more than $7 billion in cash, stock and deferred payments, with the rest tied to targets, would make it one of the largest acquisitions Amazon has ever made.
Here is what has been reported, what Decart’s products actually do, why Amazon wants them, and what to watch over the next two weeks.
What we know so far
The talks are not new. The Wall Street Journal first reported on Oct. 11 that Amazon was negotiating to buy Decart for around $7 billion. What changed this week is how close the two sides appear to be.
On Nov. 18, the Israeli business paper Calcalist reported that most of the legal issues that had slowed the negotiations were resolved. According to Calcalist, Amazon would pay more than $7 billion through a mix of cash, stock and deferred payments, and part of the money depends on Decart hitting specific targets, which could bring the total to nearly $7.5 billion. The three founders, brothers Dean and Orian Leitersdorf and Moshe Shalev, agreed to stay at Amazon for five years and not to start a competing business, and a large share of their payout would arrive only in the fifth year.
A day later, Globes reported that the talks are being led on Amazon’s side by Nafea Bshara, the vice president who runs Annapurna Labs, the Amazon chip unit that designs Trainium. Globes said the deal was expected to be unveiled at AWS re:Invent, Amazon’s annual cloud conference in Las Vegas, which opens Nov. 30. The FT report on Saturday put the value at about $7.5 billion, matching the top of Calcalist’s range.
What’s still unclear
- Whether it is signed. Every report so far cites people familiar with the talks. Amazon and Decart have not announced anything.
- The exact price. "About $7.5 billion" appears to include performance-based payments. The guaranteed portion is reported as "more than $7 billion."
- What Decart earns. The company has not disclosed revenue. Calcalist cited estimates in the low tens of millions of dollars, with much of it coming from projects for chip companies, which would make the price a bet on the technology and team rather than current sales.
- Whether Decart keeps supporting rival chips. Its software is sold as working across Nvidia, Google and Amazon hardware. Nobody has said whether that continues under Amazon ownership.
What Decart actually makes
Decart was founded in September 2023 and describes itself as an AI research lab built around one question: "how much more can AI do on the hardware that already exists?" It is based in San Francisco with a large team in Tel Aviv, and its about page says it employs "a little over a hundred people" across Tel Aviv, New York and San Francisco. It sells three things.
DOS, the chip-switching layer
The Decart Optimization Stack, or DOS, sits between an AI model and the chip it runs on, tuning the low-level code (kernels, compilers and runtime) so that more of the chip stays busy. Decart says DOS is "hardware-agnostic from the ground up" and can move a model between Nvidia GPUs, AWS Trainium and Google TPUs, so customers can run wherever compute is cheapest or available.
The company’s claims are big. Its product page lists 30 times more video frames per second on Trainium than an unoptimized baseline, eight times more tokens per second on Google TPUs, and about 1,500 tokens per second on OpenAI’s open-weight gpt-oss-120b model. When it announced DOS 2.0 in May, Decart said it delivered "over 1,600 tokens per second for agentic inference (versus an industry average of around 200)." These are Decart’s own benchmarks, and the company has not published the test setups behind them. Decart says it earns "significant revenue" licensing DOS to cloud providers, AI labs and the largest hyperscalers.
Lucy, live video editing
Lucy is a model that edits video while it is still being filmed or streamed. It can swap an outfit, change a background, add a character or restyle a whole scene frame by frame, and Decart says it does this in under 40 milliseconds per frame at up to 100 frames per second. The latest version, Lucy 2.6, released Nov. 5, is pitched at live shopping, where a viewer can see a garment on a person in real time. GeekWire has reported that streamers already use Decart’s technology on Twitch, which Amazon owns.
Oasis, worlds for games and robots
Oasis is the product that first made Decart famous. In October 2024 the company released a playable, Minecraft-style world generated entirely by a transformer model, with no game engine behind it: every frame was produced in response to the player’s keyboard and mouse. Decart says one million people tried it in 72 hours and five million in a week. Its June 2026 successor, Oasis 3, aims at robotics instead, starting with self-driving cars: it generates realistic, interactive driving scenes from a text prompt so autonomous systems can be trained on rare situations, like a whiteout or a child chasing a ball into the road, without waiting for them to happen.
As an Amazon Associate, Aglena earns from qualifying purchases.
Why Amazon would pay $7.5 billion
The short version is Nvidia. Amazon designs its own AI chips, and the current one, Trainium3, is a serious piece of hardware. AWS says each Trainium3 chip delivers 2.52 petaflops of FP8 compute with 144 GB of HBM3e memory, and its largest servers link 144 of them for 362 petaflops. But hardware alone has not been enough to pull most developers away from Nvidia, whose CUDA software has been the default way to build and run AI models for years. As we explained in our look at why GPUs became the engine of AI, that software head start is a big part of Nvidia’s moat. Moving a model built for CUDA onto another chip normally takes extra engineering work.
Decart’s pitch is that it can close that gap. It was one of the first outside companies to run real-time models on Trainium3. When AWS launched the chip at re:Invent in December 2025, its press release named Decart as "achieving 4x faster inference for real-time generative video at half the cost of GPUs." On stage that week, according to GeekWire, Decart’s model turned AWS senior vice president Peter DeSantis into a live cartoon of Amazon CTO Werner Vogels, generated on Trainium3 as he opened his keynote. When Decart raised money in May, Bshara, the executive now reportedly leading the acquisition talks, supplied a quote for the announcement.
They were among the first in the world to optimize on Trainium3 and have embraced our newest software libraries. The results speak for themselves.
Nafea Bshara, vice president, Amazon’s Annapurna Labs, in Decart’s funding announcement, May 18, 2026
In the same statement, Bshara said Decart’s Lucy 2 model ran above 80% "Model FLOPS Utilization," meaning more than four-fifths of the chip’s raw compute was doing useful work. For a cloud provider, utilization is money: every point of idle silicon is capacity Amazon paid for and cannot rent out. Owning the team that gets those numbers would let Amazon build the tuning into Trainium itself, make Trainium easier for customers who wrote their models for Nvidia, and squeeze more out of the chips already in its data centers.

Photo: Tedder, 2023, via Wikimedia Commons (CC BY-SA 4.0)
There is a precedent inside Amazon. In 2015 it bought another Israeli chip startup, Annapurna Labs, for a reported $350 million, according to GeekWire. That team went on to design Amazon’s Graviton processors and its Trainium and Inferentia AI chips. The same group is now reportedly running the Decart deal.
It also fits a wider pattern. Chips are only half of the AI hardware race, and the companies building them keep paying up for the software that makes them usable. OpenAI’s own models were recently used to help improve an AMD GPU design, and investors just lent about $44 billion to buy Nvidia chips for SpaceX. Every gain in how well existing chips are used is worth a lot when the chips themselves are this expensive and this scarce.
How Decart got here
September 2023: Decart is founded by Dean Leitersdorf, his younger brother Orian, now the company’s chief scientist, and Moshe Shalev.
October 2024: Decart comes out of stealth with a $21 million seed round led by Sequoia and releases the original Oasis playable world. A $32 million Series A led by Benchmark follows in December.
August 2025: A $100 million round. In December, Decart’s video model runs live on Trainium3 at re:Invent.
May 18, 2026: Decart raises $300 million led by Radical Ventures, with Nvidia, Adobe Ventures, Toyota Ventures and eBay Ventures joining and Amazon named as a strategic customer. The Journal later reported the round valued it at nearly $4 billion. Decart says it has raised more than $450 million in total.
August to September 2026: Calcalist reports on Aug. 10 that Decart is exploring a sale. Bloomberg reports Anthropic is in talks to buy it for about $6 billion; Anthropic walks away in September after due diligence. Calcalist says Decart then tried to restart earlier talks with Nvidia, which declined.
Oct. 11, 2026: The Journal reports Amazon is in talks at around $7 billion.
Nov. 18 to 21, 2026: Calcalist, Globes and the FT report the deal is nearly done at more than $7 billion, possibly up to $7.5 billion.
How big would this be for Amazon?
Big, but not the biggest. Amazon paid $13.7 billion for Whole Foods in 2017 and $8.45 billion for the MGM film studio in 2022, and in April 2026 it agreed to buy the satellite company Globalstar for $11.57 billion, according to GeekWire’s tally. A Decart deal near $7.5 billion would land just behind those three, and far above what Amazon has paid for most tech startups. It would also be a striking price for a company with a little over a hundred employees and undisclosed revenue, which says a lot about how much big cloud companies now value people who can make AI chips run faster.
For people who simply use AI tools, the effect would be indirect. If Amazon can run the same models on cheaper, more fully used chips, that should eventually show up as lower prices or faster responses for services built on AWS, and as more pressure on Nvidia’s pricing. It could also bring Lucy-style live video effects to Amazon’s own products, from Twitch to shopping, though no one has announced plans for that.
What happens next
The date to watch is Nov. 30, when re:Invent 2026 opens in Las Vegas. It runs through Dec. 4, and Globes reported that Amazon planned to unveil the deal there. If an announcement comes, look for three things: the guaranteed price versus the performance payments, whether DOS will keep supporting Nvidia and Google chips, and whether Amazon folds Decart into Annapurna Labs. A signed deal of this size would also have to be reported to U.S. antitrust regulators before it can close. Until then, every number in this story comes from people familiar with the talks, not from the companies.
Further reading

Chip War: The Fight for the World’s Most Critical Technology
Chris Miller’s prize-winning history of the semiconductor industry, and why control of advanced chips became a contest between companies and countries.

NVIDIA Jetson Orin Nano Super Developer Kit
Nvidia’s palm-sized AI computer for hobbyists and students, a hands-on way to see how CUDA software ties AI models to Nvidia hardware.

Logitech C920x HD Pro Webcam
A long-running 1080p webcam for video calls and streaming, the kind of live camera feed real-time video models like Lucy are built to transform.

Elgato Stream Deck MK.2
A 15-key controller that Twitch streamers use to switch scenes, trigger effects and run their broadcast setup with one press.

Amazon Echo Dot
Amazon’s small smart speaker, an everyday example of AI that runs on Amazon’s cloud and answers by voice in a second or two.
