WSJ: Amazon in talks to buy Decart for $7 billion — 75% above its May valuation
The Wall Street Journal reports Amazon is in advanced talks to acquire Israeli AI company Decart for roughly $7 billion — 75% above its valuation from May. What's actually for sale isn't world models, but software that makes AI models run efficiently on chips that aren't Nvidia's.
The bid that isn't signed yet
The Wall Street Journal reported on October 12, 2026, according to accounts relayed by Walla! Money and The Jerusalem Post, that Amazon is negotiating to buy Decart for approximately $7 billion. The talks are said to be advanced and a deal could be near — but it is not final, and the report explicitly leaves open that the conversations could collapse or that another buyer could emerge (Walla!/Jerusalem Post via WSJ). Business Standard reports the talks are underway with no agreement confirmed (Business Standard).
Neither company has commented. Amazon declined to comment, citing its policy of not addressing speculation, and Decart did not respond to Benzinga's inquiry (Benzinga via AOL). The entire story therefore rests on WSJ's sources, relayed through secondary outlets.
The price is remarkable in itself: the $7 billion figure represents a roughly 75% increase in value compared with May of this year, when Decart was valued at around $4 billion (8355efa8).
What Decart actually makes
Decart was founded in 2023 by Dean Leitersdorf, Moshe Shalev and Orian Leitersdorf, and is publicly known for real-time generative video and interactive world models, including Oasis and Lucy (2f2011c4). But what reportedly makes the company worth $7 billion to Amazon is not the models themselves.
In May 2026, Decart announced its Decart Optimization Stack (DOS): a platform combining hardware-aware modeling, its own compilers, tools for custom-tailored kernels and inference optimization in a single solution that runs across Nvidia GPUs, Google TPUs and Amazon's own Trainium chips (40efa2ca). In short, DOS is the software that makes it possible to move a model from one chip architecture to another without rewriting everything — work that would otherwise take months.
That is what explains the price surge. Nvidia's dominance in AI is not just about faster chips but about CUDA — the software ecosystem nearly all AI development is built on. A software stack that makes models just as efficient on other hardware attacks that lock-in directly.
The numbers — and where they come from
It is important here to distinguish confirmed figures from claimed ones:
- 80% Model FLOPS Utilization: Amazon itself has confirmed that Decart's Lucy2 model achieved 80% MFU on Trainium3 — a measure of how efficiently the software exploits the chip's theoretical peak performance. According to TechTimes, chip engineers consider this very good for any hardware that isn't Nvidia's (40efa2ca). This is the only performance figure confirmed by Amazon.
- 1,600+ tokens per second: Decart states that DOS delivers over 1,600 tokens per second for agentic inference, against an industry average of around 200. The company also claims eight times faster speed and one hundred times better cost efficiency compared with baseline systems on equivalent hardware. These figures are company-stated and not independently verified by third-party benchmarkers (40efa2ca).
The difference between the two categories of numbers is the story's core: Amazon's confirmed MFU figure suggests Decart's software really does get Trainium running near its ceiling, while the spectacular throughput figures are, for now, only Decart's own accounting.
The price history: Nvidia, Anthropic — now Amazon
The Amazon report is the latest chapter in a dense prehistory. According to the WSJ account, Decart's founders previously rejected an acquisition offer from Nvidia and instead chose Anthropic's offer, made in stock (8355efa8). Then, in September, Bloomberg reported that Anthropic withdrew from talks to buy the company for around $6 billion after completing due diligence (2f2011c4). About a month later, Amazon is in the picture — with a reported bid roughly $1 billion above the amount Anthropic walked away from.
The funding history shows how quickly the value has moved: in May 2026, Decart raised $300 million in a Series B led by Radical Ventures, with Nvidia among the investors, at a valuation of nearly $4 billion. The company has raised over $400 million in total from investors including Sequoia Capital, Benchmark and Radical Ventures, according to reports citing PitchBook data (596df40e, 2f2011c4).
That Nvidia is both an investor and a previously rejected bidder, and that Anthropic withdrew after scrutiny, makes the picture murky: there is no public explanation for why due diligence led Anthropic to drop out — and yet Amazon apparently believes the company is worth more than that amount.
Why Amazon cares
The context is that AWS already has a relationship with Decart. As early as 2025, AWS gave the company early access to Trainium3, the chip Amazon developed in-house (8355efa8). With the confirmed 80% MFU, Decart has demonstrated something Amazon needs: that Trainium can be exploited efficiently enough to be a genuine alternative when customers would otherwise default to Nvidia.
For Amazon, the math is in practice an inference equation. The more models that can run cheaply on Trainium instead of Nvidia's GPUs, the more of the AI workload AWS can keep in its own cloud — at lower cost and with less dependence on a single chip vendor.
The open question: neutral software or locked to Trainium?
The entire strategic value rests on one unanswered question: will DOS still support rival hardware after an acquisition? As The Motley Fool puts it, the value for Amazon shareholders lies in whether a chip-agnostic software layer strengthens Amazon's cloud offering — and it does so only if Decart's tools keep working on rivals' chips. Locked to Amazon's own silicon, developers will route around it (846ed1a7).
Neither company has commented, and no source can yet confirm how this would be handled. That is the real uncertainty in the deal: if Amazon buys Decart and makes DOS Trainium-exclusive software, it has bought a tool with no market power. If it leaves it neutral, it has bought a foothold in the infrastructure beneath the entire industry — including its competitors'.
What remains to be seen
The deal could collapse, as Anthropic's $6 billion agreement did. The reported history of previous bidders rests on secondary sources and cannot be independently verified. And Decart's own performance figures still lack independent verification. The documented picture is nonetheless clear enough: a company valued at $4 billion in May is now, according to Wall Street Journal's sources, worth around $7 billion — because ownership of the software that moves AI models between chip architectures may be worth far more than the models themselves.