On Sunday, The Wall Street Journal reported that Amazon is in discussions to acquire AI startup Decart in a deal worth roughly $7 billion. Citing sources familiar with the matter, the Journal said a deal could come together soon, provided the talks do not collapse and no rival bidder emerges. Neither Amazon nor Decart has announced a deal, and both declined to respond to Reuters’ requests for comment on Sunday.
If it materializes, the acquisition would surpass Amazon’s $6.5 billion purchase of the MGM movie and television studio in 2022, making Decart one of the largest deals in Amazon’s history. It would also mark a striking turn in a saga that already has one collapsed suitor: just last month, Anthropic walked away from a deal to buy Decart that could have been worth about $6 billion, after completing due diligence, according to Bloomberg reporting.
Amazon itself has backed Anthropic with roughly $8 billion. That its portfolio lab passed on Decart at $6 billion, only for Amazon to reportedly step in itself at around $7 billion weeks later, says something about how strategically valuable Decart’s technology has become.
What Decart actually does
Decart was founded in 2023 by Dean Leitersdorf, its CEO, and Moshe Shalev. It is best known to the public for its generative video work: Lucy, a model that edits live video in real time, and Oasis, a world model that generates simulated 3D environments used to train and test robots and self-driving systems. But the models are not what a buyer like Amazon is paying for.
Decart’s core infrastructure product is the Decart Optimization Stack, or DOS. The company describes DOS as software that speeds up AI training and inference across Nvidia GPUs, Google TPUs, and Amazon’s Trainium chips, and it licenses the software to cloud providers and AI labs. Decart says its latest stack can process more than 1,600 tokens per second for AI agents.
That portability is the point. As companies scramble for computing power, the ability to move workloads efficiently between different AI chips has become a real economic lever. Nvidia dominates the market for the GPUs that power the AI boom, and much of that dominance rests not just on the hardware but on the software ecosystem that makes Nvidia’s chips the default place to run AI. A stack that lets developers run AI efficiently on alternative hardware, without rewriting their work for each chip, weakens that lock-in.
Why Amazon wants it
Amazon’s interest in Decart did not start last week. Amazon Web Services gave Decart early access to its custom Trainium3 chip last year after meeting the company and coming away impressed by its founders, according to the Journal’s reporting. Trainium3 is Amazon’s in-house AI chip, developed to compete with Nvidia.
The relationship is already commercial. Decart raised $300 million on May 18 in a round led by Radical Ventures, with Nvidia joining as a new investor, and the funding announcement named Amazon as one of Decart’s strategic customers. Decart has a commercial agreement and a joint sales effort with AWS. Its live video model, Lucy 2, runs on Trainium3, and Nafea Bshara, a vice president at Amazon’s chip design unit Annapurna Labs, said in the announcement that Decart was “among the first in the world to optimize on Trainium3.”
PitchBook data put Decart’s valuation at nearly $4 billion in May, up from $3.1 billion when it raised $100 million in August 2025. A $7 billion acquisition price would be roughly a 75 percent markup on that May valuation in less than six months, a sign of how competitive the bidding for AI infrastructure talent and technology has become.
A wider chip strategy taking shape
The Decart talks fit a pattern of Amazon aggressively building its own silicon alternatives. In September, Amazon selected Qualcomm to design custom AI inference chips for its data centers. The order books are booming across the industry: AMD closed above a $1 trillion market cap earlier this month as OpenAI and Oracle chip orders landed. Owning Trainium, partnering with Qualcomm on inference silicon, and potentially acquiring Decart’s cross-chip optimization stack would give Amazon a fuller bench of hardware and the software glue that makes it usable.
The economics explain the urgency. Inference, running AI models for users rather than training them, is becoming the dominant cost center of the AI boom. Whoever controls cheaper, more efficient inference controls the margins of the cloud business. A company that owns both the chip and the software that squeezes performance out of competing chips is in a position to steer a large share of that spend toward its own hardware.
What could still go wrong
Nothing is signed. The Journal’s reporting is explicit that the talks could collapse or that another contender could make an offer, and neither company has confirmed anything. Large tech acquisitions also face real regulatory scrutiny, which has killed or reshaped several deals in recent years.
There is also the question of what Anthropic found during its due diligence in September that made it walk away from a $6 billion price. A buyer doing a deal 17 percent above the price a well-informed, Amazon-backed lab rejected weeks earlier will need to be confident it knows something the first bidder did not, or values the asset differently enough that it does not matter.
What is not in question is the direction of travel. The AI chip wars have moved beyond hardware specs into software portability, and the biggest cloud companies are now buying the software that decides where AI workloads run. Whether Decart ends up inside Amazon or not, the $7 billion figure is a data point in itself: the abstraction layer between chips and models is now priced like a crown jewel.
