Reuters reported on October 11 that Amazon is in talks to acquire AI startup Decart for around $7 billion, citing a Wall Street Journal report based on people familiar with the matter. No agreement has been announced, the talks could still fall apart or attract another bidder, and neither company has confirmed a deal. Even so, the reported price and the buyer matter because this is not just another AI startup auction. It is a test of whether control over the software layer that makes different AI chips usable is becoming as strategic as the chips themselves.
That is the real question behind the headline: is Amazon trying to buy a model company, or the bridge that could make its custom Trainium chips more competitive with Nvidia?
What Decart actually has
Decart, founded in 2023 by Dean Leitersdorf, Orian Leitersdorf and Moshe Shalev, sits across several parts of the stack. Reuters reported that the company raised $300 million in May in a round led by Radical Ventures, with Nvidia joining as a new investor, and that PitchBook valued it near $4 billion at the time. A reported $7 billion sale price would therefore represent a steep step-up, though both numbers are approximate and the deal terms remain unsettled.
The company’s own materials describe three connected products. Its Decart Optimization Stack, or DOS, is meant to improve training and inference performance across Nvidia GPUs, AWS Trainium, Google TPUs, AMD accelerators and other chips. Lucy is a real-time video-editing and world-model product. Oasis is an interactive world model for physical AI, aimed at generating controllable environments for robotics and autonomous-vehicle training.
Those model and application teams could be part of Amazon’s interest. But the strategic center of gravity appears to be DOS. In AI infrastructure, the bottleneck is often not access to silicon alone. It is the engineering work needed to port models, tune kernels, manage compiler behavior, and hit acceptable latency and cost on a given accelerator. A model that runs well only on one chip deepens lock-in. A stack that can be retuned across chips makes more hardware choices viable.
That is why Decart already matters to Amazon beyond a simple investment return.
Why AWS would care
Amazon and Decart are not strangers. AWS is already a strategic customer, and reporting says Decart received early access to Trainium3 in 2025. AWS customer materials also say Decart’s early Trainium testing delivered as much as four-times higher frame throughput and two-times better cost efficiency than top GPUs on a specific workload, with latency dropping from 40 milliseconds to 10 milliseconds. Decart’s own research site describes Lucy running at 101 frames per second on AWS Trainium in a particular optimization result.
Those figures are useful as clues, not universal scorecards. They are vendor-reported results tied to specific workloads, not proof that Trainium broadly outperforms Nvidia across enterprise AI. But they do reveal the mechanism that makes this reported deal strategically important: software tuning can make custom silicon look far more competitive than raw ecosystem comparisons suggest.
For AWS, acquiring Decart could reduce one of Trainium’s biggest disadvantages against Nvidia. Nvidia’s lead is not only about hardware. It is also about the maturity of the software ecosystem, the breadth of existing model support, and the depth of engineering familiarity across customers and developers. If DOS can cut the time and talent needed to make a workload run well on Trainium, Amazon gains a lever it cannot get by ordering more chips alone.
That could matter directly to enterprise buyers. If more workloads can move onto Trainium without a painful rewrite, AWS gets a stronger pitch on cost and capacity. Scarce or expensive accelerator supply becomes more fungible. Enterprises gain another path besides defaulting to Nvidia-heavy configurations.
The catch: neutrality is part of the value
The strongest argument for Decart is also the source of the biggest risk. Decart markets DOS as hardware-aware but hardware-agnostic. That neutrality is what makes it valuable in a world where companies want leverage across clouds and accelerators.
Inside Amazon, the incentives would change. AWS could keep DOS broadly available and use its cloud distribution to expand adoption. But it could also prioritize Trainium first, bundle the stack tightly with AWS services, limit access, or steer the roadmap toward Amazon’s own workloads. Any of those choices might improve Trainium utilization while reducing the cross-platform value that attracted customers in the first place.
That tension matters because portability is not just a technical feature. It is a pricing tool. The more credible it is for a customer to move a model between Nvidia, Trainium, TPUs and AMD accelerators, the more negotiating power that customer has over infrastructure cost. If a neutral bridge becomes an in-house AWS advantage, enterprises may get better performance on Amazon but lose some bargaining power and exit flexibility.
This is also where the unresolved parts of the Decart sale process matter. Bloomberg previously reported that Anthropic explored buying the company at around $6 billion before walking away after due diligence in September. Nvidia had also pursued Decart. Public reporting does not explain why Anthropic stopped, and that gap should temper any rush to treat a high headline price as validation of every product claim or commercial assumption.
What to watch if the deal progresses
If Amazon does reach an agreement, the first test will not be the purchase price. It will be operating posture.
Enterprise buyers should ask a short list of practical questions. How much staff time does DOS actually save when porting a real production model? Does model quality hold up after optimization, or only benchmark speed? How much ongoing compiler and kernel maintenance is required across hardware generations? What is the total cost after cloud, storage, networking and support charges, rather than accelerator pricing alone? And if a customer later wants to leave AWS, how portable is the optimized workload in practice?
Investors should watch for something slightly different: whether Decart creates durable usage and margins, or only eye-catching benchmark moments. The reported $7 billion valuation makes sense only if Amazon would be buying more than a one-time tuning advantage. It would need technology and talent that can repeatedly make alternative accelerators usable enough to change buying behavior.
That is why this reported transaction fits a broader 2026 pattern. AMD has agreed to buy World Labs for $8.2 billion, and Nvidia’s reported talks with Reflection AI have already drawn attention. The common theme is not simply that chip companies want model companies. It is that control over models, optimization software and developer workflows increasingly shapes which silicon gets used.
Amazon can design more chips. What it cannot automatically buy is a software ecosystem that makes customers comfortable choosing them. If the Decart talks turn into a deal, the smartest way to read the price is not as proof that Decart is indispensable today. It is as a possible payment for control over the compatibility layer that could decide how much real choice the AI infrastructure market still has tomorrow.
By
By
By
By
By
