Amazon's $7 Billion Bid for Decart Is Really About Escaping Nvidia's Grip

The Wall Street Journal reported Sunday that Amazon is in talks to acquire Decart, an Israeli-founded artificial intelligence startup, in a deal worth roughly $7 billion. Sources familiar with the matter told the Journal that an agreement could come together soon, unless the talks collapse or another bidder emerges. Neither Amazon nor Decart has confirmed anything, and both declined to comment on the report.

If it closes, the deal would be Amazon's largest acquisition ever, surpassing the $6.5 billion it paid for the MGM movie and television studio in 2022. But this is not a media play. It is a bet on the layer of the AI stack that is quietly becoming the most strategically valuable of all: the software that decides which chips AI models run on. (Chip portability is suddenly the industry's favorite battleground: DeepSeek open-sourced a CUDA rival.)

The startup at the center of the bid

Decart was founded in San Francisco in 2023 by Dean Leitersdorf and Moshe Shalev. Its main infrastructure product is the Decart Optimization Stack, or DOS: software that the company says speeds up AI training and inference across Nvidia GPUs, Google TPUs, and Amazon's own Trainium chips. Decart licenses DOS to cloud providers and AI labs, and says its latest stack can process more than 1,600 tokens per second for AI agents.

On top of the stack, Decart builds two generative models. Lucy edits live video in real time, and Oasis generates simulated 3D environments used to train and test robots and self-driving systems. The company's latest funding round, $300 million in May led by Radical Ventures, brought Nvidia in as a new investor and valued Decart at nearly $4 billion, up from $3.1 billion when it raised $100 million in August 2025. PitchBook data puts total fundraising above $400 million, with backers including Sequoia Capital and Benchmark.

This is not a media play. It is a bet on the software that decides which chips AI models run on.

Why chip portability is the prize

The technical pitch sounds mundane: software that lets an AI model run well on chips it was not originally built for. Today, models are typically built and tuned for a specific processor, and moving them to another chip means performance-tuning work that costs time and money. Decart's software is designed to remove that burden, making chips effectively interchangeable.

That ability has become strategically precious. As AI demand has surged, Nvidia's GPUs have become both scarce and expensive, and relying on a single chip supplier leaves cloud providers and AI labs exposed to supply shocks and price increases. The ability to shift workloads flexibly across Nvidia, Google, and Amazon silicon is, for a cloud giant, the difference between resilience and exposure.

Amazon is deep in this fight. It introduced Trainium, its first AI training chip, in 2020, and now offers it through AWS to customers including OpenAI, Anthropic, and Uber. In April, the company said revenue commitments tied to Trainium had surpassed $225 billion, and it is now working to sell the chips directly into other companies' data centers, not just its own cloud. In September, Amazon selected Qualcomm to design custom AI inference chips for its data centers. Trainium3, the newest generation, was built to go head to head with Nvidia hardware.

Decart is already entangled with that effort. The startup was among the first in the world to optimize for Trainium3, according to Nafea Bshara, a vice president at Amazon's chip design unit, Annapurna Labs. AWS gave Decart early access to the chip last year after company officials met the founders and came away impressed. Decart calls Amazon one of its strategic customers, and the two have a commercial agreement with a joint sales effort.

The Anthropic twist

Amazon is not the first suitor. In August, Anthropic explored acquiring Decart in a deal that could have been worth about $6 billion, but dropped the pursuit in September after due diligence. That is notable because Amazon has invested roughly $8 billion in Anthropic, making its backer's new bid an interesting sequel. Whatever gave Anthropic pause, Amazon apparently saw the math differently, or wanted something different from the deal.

The reported $7 billion figure would represent a jump of roughly 75 percent from Decart's May valuation of nearly $4 billion, suggesting either a competitive bidding environment or a steep strategic premium for the chip-portability layer and the world-model research.

What this signals

The Decart talks fit a broader pattern: the AI wars are moving down the stack. The early battles were over models; the current ones are over compute, chips, and the software between them. A cloud provider that controls the optimization layer gains leverage over which silicon its customers, and its own vast AI operations, run on. For Amazon, Decart could make Trainium the path of least resistance for AI workloads, chipping away at Nvidia's dominance without requiring customers to rewrite their software.

The world models are the wildcard. Lucy and Oasis point toward simulation and robotics training, areas where Amazon, with its warehouse robotics empire and autonomy ambitions, has natural interest. Real-time video editing and synthetic environments for robot training are not side projects. They are the raw material of the next generation of physical AI.

For now, it is only talks. Acquisitions of this size can still fall apart, and the Journal's sources were careful to note the contingency. But the logic is already visible: in the race to control AI infrastructure, the most valuable startup of 2026 might be the one that frees you from having to care which chip you run on.

References

The Wall Street Journal, reported via PYMNTS, "Amazon Weighs $7 Billion Purchase of AI Startup Decart," October 11, 2026. Seoul Economic Daily, "Amazon in Talks to Buy AI Chip-Portability Startup Decart for $7 Billion," October 12, 2026. AIStockWire, "Amazon (AMZN) reportedly in talks to buy Decart for $7B," October 2026.