Amazon is in active negotiations to acquire Decart, an Israeli AI startup, at a valuation of approximately $7 billion — a deal that, if finalized, would hand Amazon the inference optimization software it needs to make its own custom AI chips finally viable against Nvidia’s dominant hardware ecosystem. SiliconAngle confirmed Amazon in acquisition talks with Decart at that figure on October 11. The Wall Street Journal reported the talks Sunday, citing people familiar with the discussions, and said an agreement could be wrapped up within days.
The deal would be Amazon’s largest pure AI software acquisition and one of the most expensive exits in Israeli technology history. It would also conclude a months-long bidding war in which Nvidia, SpaceX, and Anthropic each circled the three-year-old company before Amazon — which had already built a commercial relationship with Decart through a $300 million strategic partnership formed in May 2026 — moved to the front of the line.
What Decart Has That Amazon Needs
Decart’s primary product is not a language model or a flashy consumer application. It is the Decart Optimization Stack — called DOS internally — a vertically integrated software platform that makes AI models run dramatically faster on existing hardware. The Decart Optimization Stack announced May 2026 combines hardware-aware model design, proprietary compilers, custom kernel tooling, and inference optimization into a single platform that runs across Nvidia GPUs, Google TPUs, and Amazon’s own Trainium chips.
What makes DOS strategically valuable is what it automates. When an AI developer wants to run workloads on a non-Nvidia chip — including Amazon’s Trainium — they face weeks or months of specialized engineering work to port their models to that hardware efficiently. Nvidia’s CUDA platform, built over 15 years, has accumulated the libraries, frameworks, and developer tooling that make Nvidia hardware the path of least resistance for most AI teams. AWS’s competing Neuron SDK lags behind CUDA maturity in ways that matter to enterprise customers. DOS compresses that porting process from months to weeks, according to Decart, removing the primary barrier to Trainium enterprise adoption.
Amazon has confirmed at least one aspect of DOS’s Trainium performance: the company stated that running Decart’s Lucy2 model on Trainium3 chips achieved 80% Model FLOPS Utilization on Trainium3 — a metric that measures how efficiently software exploits a chip’s theoretical maximum performance, and a figure that chip engineers consider excellent for any non-Nvidia hardware. Decart claims DOS delivers more than 1,600 tokens per second for agentic AI inference, compared with an industry average of approximately 200. The company also claims 8x speed and 100x cost efficiency gains against baseline systems on equivalent hardware — figures that are company-stated and have not been independently verified by third-party benchmarkers.
Why Inference Is the Battleground That Matters Now
Understanding the strategic value of DOS requires understanding where AI economics have shifted. Training an AI model — the resource-intensive process of teaching a neural network on vast datasets — is an occasional investment. Inference, the continuous process of actually serving that model to users, is where the costs accumulate. Industry research indicates inference can account for 80% to 90% of an AI system’s lifetime operating cost, and the AI inference market growing to $255 billion by 2030 from approximately $106 billion in 2025.
Per-query inference costs have already fallen dramatically — from roughly $20 per million tokens in 2022 to approximately $0.07 per million tokens by 2025, driven largely by optimization advances rather than new hardware. But as AI products scale from thousands of users to millions, inference bills still represent one of the largest line items in an AI product company’s operating budget. Software that reduces those bills by a factor of eight — or even a more conservative factor of three or four in production deployments — is not a peripheral tool. It is potential infrastructure for AI at scale — competitive infrastructure for any company that runs AI at production volume.
Amazon Web Services faces this dynamic as both a cloud provider and AI product developer — as the infrastructure platform running other companies’ inference workloads, and as an AI product company in its own right. Its competitive position in cloud AI depends increasingly on being able to offer customers better inference economics than rivals, and those economics depend on whether customers can efficiently use Trainium chips rather than defaulting to Nvidia hardware at higher cost.
Trainium’s Problem — and Decart’s Proposed Solution
Amazon has invested heavily in Trainium as a custom AI silicon play, with Jeff Bezos describing chips as next pillar alongside AWS, Prime, and Marketplace. Trainium3 reportedly delivers 4x performance improvement and 40% better energy efficiency compared with prior generations. The hardware case for Trainium is real. The adoption case has been harder to make.
The obstacle is the CUDA ecosystem. Nvidia has spent 15 years building the software environment — compilers, libraries, debugging tools, training frameworks — that makes running AI on its GPUs straightforward for any competent ML engineer. AWS’s Neuron SDK, the Trainium equivalent, is improving but widely acknowledged as still catching up. AWS engineers reportedly understand internally that CUDA familiarity keeps developers on Nvidia hardware even when Trainium’s price-performance metrics are favorable.
DOS addresses this gap directly. Rather than asking engineers to learn Neuron SDK deeply, DOS automates the hardware-specific optimization work — the kernel selection, the memory layout decisions, the compute scheduling — that previously required months of manual engineering. For an enterprise customer considering shifting inference workloads from Nvidia to Trainium, DOS turns that migration from a major engineering project into a manageable transition.
That is the specific strategic logic of a $7 billion acquisition: Amazon would be buying the software key to unlocking the hardware investment it has already made. DOS becoming a first-party AWS tool would mean every Trainium customer gains access to the optimization stack that currently requires engaging with Decart separately. The result, if it works as described, could be a meaningful pull toward Trainium adoption across AWS’s enterprise base.
There is also a broader platform logic embedded in the acquisition. DOS does not work only on Trainium — it runs on Nvidia GPUs and Google TPUs as well. An Amazon-owned DOS that continues serving the full hardware ecosystem would give AWS a software layer that sits between AI developers and all major chip platforms. That is a more powerful position than any single-hardware play.
The Bidders Decart Outlasted
The $7 billion figure sits 75% above Decart’s most recent funding valuation of approximately $4 billion, established in a $300 million round five months ago. That round drew a notable roster of strategic investors: Nvidia, which had reportedly also explored acquiring Decart outright; eBay Ventures; Adobe Ventures; and Toyota Ventures. Returning investors Sequoia Capital and Benchmark also participated, along with individual backers including Andrej Karpathy, the former Tesla AI director and OpenAI co-founder who has become one of the more visible technology investors in applied AI infrastructure.
The acquisition path to Amazon has been circuitous. Nvidia reportedly explored acquiring Decart earlier in 2026 before a larger bid materialized. CTech in Israel first reported in August that Decart’s founders were weighing a $6 billion to $7 billion sale, with SpaceX also described as circling. Bloomberg reported on August 13 that Anthropic, the Claude developer, had entered advanced acquisition talks at a reported $6 billion figure — a deal that would have embedded Decart’s engineering team inside Anthropic’s inference and performance organization.
Then, on September 8, Bloomberg reported that Anthropic had completed due diligence and walked away without explanation. Neither company commented. Analysis at the time focused on the timing: Anthropic was simultaneously preparing for an initial public offering with marketing expected to begin in mid-October, and absorbing a stock-based acquisition worth $6 billion to $7 billion immediately before going public would have introduced significant balance sheet complexity. Whatever Anthropic’s internal reasoning, its withdrawal left Decart available — and Amazon, already holding a commercial relationship with the startup, moved to close the gap.
Three Years, $450 Million, $7 Billion
Decart was founded in 2023 by three engineers from Israel’s Unit 8200 intelligence unit: brothers Dean and Orian Leitersdorf and Moshe Shalev. Dean Leitersdorf, who completed a computer science doctorate at the Technion at age 23 — a record at the time, since broken by his younger brother Orian — has described his ambition as building “the Google or Apple of AI.” The company employs approximately 100 people and has offices in San Francisco, New York City, and Tel Aviv.
Decart’s public breakthrough came in October 2024, when it and chip startup Etched released Oasis: an AI-generated playable version of Minecraft that ran entirely inside a neural network with no traditional game engine. The demo spread quickly and established the company’s reputation as a serious world-model developer. A world model, in AI terminology, is a system that builds an internal representation of an environment and predicts how that environment changes in response to actions — enabling real-time generation of interactive simulated spaces rather than merely static images or short video clips. The commercial applications extend to robotics training environments, autonomous vehicle simulation, and synthetic data generation for industrial AI.
Oasis has since been updated to Oasis 3, and the third product line — Lucy — is a real-time video model capable of modifying live video feeds as they stream, with sub-30 millisecond latency real-time video. Together, the product lines give Decart its dual identity: infrastructure software solving inference economics at scale, and a generative AI frontier developer capable of the kind of real-time simulation that the next generation of physical AI applications requires.
The company raised a seed round at above $500 million valuation in late 2024, followed by $100 million at $3.1 billion in August 2025, and then the $300 million May 2026 round. Total capital raised stands at more than $450 million raised total. The proposed $7 billion acquisition price would put Decart among the most expensive per-employee AI acquisitions on record, reflecting an assessment that the team — with its deep background in low-level programming, compiler design, and GPU optimization — is as valuable as the products they have already built.
What This Deal Says About Where AI Value Lives
The succession of major technology companies bidding for a 100-person, three-year-old startup reflects a specific judgment about where AI value is accumulating in 2026. Through 2023 and much of 2024, the dominant competition was at the model layer — who had the best language model, the most capable reasoning system, the most useful image generator. Through 2025 and into 2026, the emphasis has shifted to inference economics: the cost and speed of running models at scale for real users.
A model that costs a dollar per query serves a fundamentally different — and much smaller — set of applications than one that costs a cent. Software that compresses that cost gap without requiring new hardware is not a peripheral tool; it is economic infrastructure for entire AI industry.
Decart and Amazon declined to comment.
Frequently Asked Questions
What does Decart’s DOS software actually do?
DOS — the Decart Optimization Stack — is a vertically integrated software platform that makes AI models run faster and more cheaply on existing hardware. It uses hardware-aware model design, proprietary compilers, custom kernel tooling, and inference optimization techniques to extract more performance from chips without requiring new hardware purchases. The same stack runs on Nvidia GPUs, Google TPUs, and Amazon Trainium chips. Decart claims DOS delivers more than 1,600 tokens per second for agentic AI inference, compared with an industry average of roughly 200 — and reduces overall inference costs by a factor of eight on equivalent hardware. These figures are company-stated; independent third-party benchmarks have not been published.
Why did Amazon pay a $7 billion premium for a company valued at $4 billion just five months earlier?
The 75% premium reflects what DOS could mean to Amazon’s Trainium chip strategy specifically. Amazon has invested heavily in Trainium as a lower-cost alternative to Nvidia’s GPUs, but Trainium’s adoption has been limited by the difficulty of porting models to it — work that currently requires months of specialized engineering. DOS automates that porting process, turning a months-long migration into a weeks-long one. If Amazon integrates DOS as a first-party tool for AWS customers, it could dramatically accelerate Trainium adoption across its cloud base — which would deliver returns far exceeding the $7 billion acquisition price.
Why did Anthropic walk away after reportedly bidding $6 billion?
Anthropic completed due diligence in early September 2026 and withdrew without public explanation. The most widely discussed theory is that timing drove the decision: Anthropic was simultaneously preparing for an initial public offering with marketing expected to begin in mid-October, and a large stock-based acquisition immediately before going public would have created significant balance sheet complexity. Neither Anthropic nor Decart has confirmed the reason for the withdrawal.
What is a world model in AI, and why does it require specialized inference software?
A world model is an AI system that builds an internal representation of a physical or simulated environment and predicts how that environment changes in response to actions. Unlike language models that predict the next text token, world models work on sensory data — video frames, sensor readings, spatial coordinates — and must generate continuous, interactive simulated spaces in real time. That real-time requirement is computationally demanding: interactive applications like Oasis require response times under 30 milliseconds to feel seamless, which means inference must be far more optimized than standard AI serving. DOS was built specifically to meet those latency requirements, making inference optimization not an optional enhancement for world models but a prerequisite for making them usable in commercial applications at all.
