Amazon in talks to buy AI startup Decart for around $7 billion
The Wall Street Journal reports that Amazon is in talks to buy Decart, the San Francisco AI startup behind real-time video model Lucy and chip-agnostic inference software, for around $7 billion. The deal is not signed. Anthropic walked away after due diligence, and Decart was valued near $4 billion in May.
Amazon is negotiating to buy Decart, a three-year-old artificial intelligence startup, for roughly $7 billion, according to people familiar with the matter cited by The Wall Street Journal on Sunday, October 11, 2026. If the talks hold, the purchase would rank among Amazon's largest technology deals and deepen the company's push to make its custom Trainium chips a real alternative to Nvidia GPUs.
Neither Amazon nor Decart has confirmed an agreement. Reuters and CNA both stressed that the Journal's sources say a deal could close soon, but that negotiations can still collapse or attract another bidder. That caution matters: Decart has already been through a sale process that pulled in Nvidia and Anthropic before Amazon took the lead.
What Decart actually builds
Decart is not a pure model lab and not a pure chip company. Its public pitch, on decart.ai, is infrastructure plus real-time generative models.
The company's Decart Optimization Stack (DOS) is software that helps AI workloads run faster and switch more easily across chips from Nvidia, Google (TPUs), and Amazon (Trainium). The Wall Street Journal framed that chip-switching layer as the strategic prize for Amazon. GeekWire, summarizing the same reporting, puts it plainly: owning Decart would give Amazon a way to make Trainium easier for Nvidia-native developers to adopt, and to squeeze more performance out of chips Amazon already sells.
On top of that stack sit Decart's own models:
- Lucy: a real-time video and "world editing" system that can transform people, products, backgrounds, and effects live as video is streamed or consumed. Decart markets sub-200 ms latency and interactive experiences for retail, ads, social, and gaming.
- Oasis: a realtime interactive world model aimed at physical AI, including robotics, autonomous vehicles, manufacturing, and drones.
Those products explain why Amazon's consumer businesses also look like a fit. GeekWire notes that streamers already use Decart technology on Twitch (which Amazon owns), and that Lucy can power virtual try-on for online shoppers. The same company that helps AWS sell Trainium could also feed Amazon's media and commerce surfaces with live generative video.
How the price jumped from under $4B to about $7B
In May 2026, Decart raised about $300 million at a valuation of roughly $4 billion, in a round led by Radical Ventures with Nvidia joining as an investor, according to multiple reports that cite PitchBook and contemporaneous coverage. CTech and CNA put total capital raised near $450 million. The Journal's reported Amazon talks at about $7 billion would therefore imply a large premium over that May mark, and sit in the same ballpark as the $6 billion to $7 billion range Anthropic had explored before walking away.
Founders Dean Leitersdorf, Orian Leitersdorf, and Moshe Shalev started the company in 2023. It is based in San Francisco with a major development center in Tel Aviv. Decart has not publicly disclosed revenue. Globes has reported that the business has been profitable almost from the start, while CTech has cautioned that the operating business remains far smaller than the headline valuations under discussion. Customers named in secondary reporting include eBay, which is also described as an investor.
Why Amazon wants this now
Amazon's AI strategy has two bottlenecks that Decart touches at once.
First, chip competition. Nvidia still dominates training and inference in large part because of CUDA and the software ecosystem around it. Porting that work to Trainium normally costs engineering time. Decart's DOS layer is designed to shrink that friction across Nvidia, Google, and Amazon silicon. GeekWire notes that Amazon already showcased Decart running real-time video on Trainium3 at AWS re:Invent: Amazon claimed about 4x faster real-time video at roughly half the cost of GPUs in that demo context. AWS had also given Decart early access to Trainium3 after senior engineers were impressed by the founders, according to SiliconANGLE and CTech.
Second, distribution. Anthropic is one of Amazon's closest AI partners. Amazon has invested about $18 billion in Anthropic, with room for more if milestones are hit, and Anthropic trains and serves Claude on Trainium. A Decart acquisition would put the optimization layer inside Amazon rather than leaving it as a shared supplier that every frontier lab can court.
The deal size also fits Amazon's recent appetite for large bets. GeekWire places a $7 billion Decart purchase behind Whole Foods ($13.7 billion in 2017), Globalstar ($11.57 billion agreed in 2026), and MGM ($8.45 billion in 2022), and far above the roughly $350 million Amazon paid for Annapurna Labs in 2015, the acquisition that seeded today's Trainium effort.
The sale process that got Amazon here
Amazon is not the first suitor.
According to CTech, Nvidia was close to an acquisition before a competing offer emerged. Nvidia later invested in the May round instead of buying the company outright. In August, Bloomberg reported that Anthropic explored a deal around $6 billion. Anthropic then completed due diligence and withdrew in September. Public explanations are thin. Globes, citing sources close to Decart, said the founders resisted Anthropic's push to move much of the Tel Aviv team to San Francisco. Bloomberg also left open that the two companies might still collaborate in other ways.
That history matters for readers for three reasons:
1. The $7 billion figure is not a random round number. It sits on top of a failed Anthropic process in a similar range. 2. Nvidia is both a past suitor and a current investor, which complicates any narrative that frames Decart as purely anti-Nvidia software. 3. Amazon already has a customer and demo relationship with Decart, so an acquisition would formalize a partnership that AWS has already advertised on stage.
SiliconANGLE and the Journal both note that another bidder could still appear, or that talks could fall apart. No signed definitive agreement has been reported.
What Lucy and Oasis say about Decart's bet
It is easy to reduce Decart to "CUDA alternative software." The company's own product pages push a broader thesis: live AI, meaning models that perceive and edit the world at the speed of streaming video and interactive simulation.
Lucy is marketed as production-scale world editing: swap characters, alter products, restyle scenes, and generate effects while content is live. Oasis is positioned for physical AI teams that need interactive, physically plausible environments for robots and vehicles. DOS is the shared runtime underneath. That combination is why valuations moved so fast. Investors are not only paying for a compiler-like layer. They are paying for a stack that ties chip efficiency to consumer-facing generative video and robotics simulation.
For Amazon specifically, Lucy maps onto Twitch and retail try-on, while DOS maps onto Trainium adoption. Oasis is a longer-horizon bet that fits Amazon's warehouse robotics and autonomy ambitions, even if those use cases are not the center of the Journal's deal story.
Open questions the report does not settle
Several gaps remain if you read the Journal, GeekWire, SiliconANGLE, CTech, and CNA side by side.
Confirmation. Amazon declined to comment outside regular hours in some writeups; Decart did not respond. Everything currently rests on people familiar with the talks.
Structure. It is unclear whether Amazon would buy 100% of Decart, keep supporting Nvidia and Google chips after closing, or fold Lucy and Oasis into AWS, Twitch, or a separate product line.
Team location. The Anthropic collapse reportedly involved disagreement over moving talent from Tel Aviv. Any Amazon deal will face the same geography question.
Revenue. Without disclosed ARR or margins, the $7 billion figure is a strategic multiple, not a classic software valuation you can reverse-engineer from public numbers.
Competition law and optics. A deal that helps Amazon pull workloads off Nvidia while Nvidia remains an investor in the target is commercially elegant and politically noisy. Regulators have been watching AI deals more closely; nothing in the Sunday report says a filing has begun.
Why this deal would still matter even if it fails
Even if Amazon and Decart never sign, the reporting already tells you where the AI infrastructure market is heading.
Cloud providers are no longer competing only on raw FLOPs. They are competing on software that makes their silicon usable. Microsoft is pushing Maia for its own products. Google has TPUs. OpenAI has worked on custom silicon with Broadcom. Amazon is trying to turn Trainium from a captive Anthropic advantage into a broader developer default. Decart sits exactly on that seam: multi-chip optimization plus flashy demos that prove the stack works under live constraints.
The failed Anthropic path also shows that acquisition talks in this market are fluid. A company can go from Nvidia interest to Anthropic diligence to Amazon talks in a few months, with the headline price rising from under $4 billion to about $7 billion along the way. That is what a scarce optimization layer looks like when every major AI buyer needs more tokens per dollar and cannot wait for a clean CUDA replacement from scratch.
For now, treat the story as advanced talks, not a closed acquisition. Watch for a formal announcement, a regulatory filing, or a clear denial. Until then, the useful takeaway is simpler: Amazon is willing to put Whole Foods-scale money (in spirit, if not quite in dollars) behind software that makes Trainium feel less like a second-class citizen next to Nvidia, and Decart is the Israeli-founded startup that suddenly sits in the middle of that fight.
Sources
Source: wsj.com