Amazon Eyes $7 Billion Decart Deal to Break Nvidia's Grip on AI Chip Software

Amazon is in talks to acquire Decart, a three-year-old AI infrastructure startup founded in Tel Aviv, for approximately $7 billion, the Wall Street Journal reported Sunday. The deal, which people familiar with the discussions told the Journal could close within days if talks do not collapse or a rival bidder emerges, would represent one of the largest acquisitions of an AI lab by a major cloud provider and a significant escalation in Amazon's effort to build a viable alternative to Nvidia's dominant GPU ecosystem.

The acquisition would bring Amazon a startup it already knows intimately. Decart's models have been distributed through Amazon Bedrock since December 2025, and its flagship world model, Lucy, runs on Amazon's Trainium3 chips — the 3nm accelerators that Amazon unveiled at re:Invent 2025 and has been working to promote as a GPU alternative across its enterprise customer base. Decart obtained early access to Trainium3 after senior AWS engineers, reportedly impressed by the founders, arranged the arrangement. The outcome of that collaboration — Lucy achieving what Amazon's own chip team, Annapurna Labs, described as greater than 80% Model FLOPS Utilization on Trainium — sits at the center of why Amazon is now willing to pay what amounts to a 75% premium over Decart's $4 billion May 2026 valuation.

Neither Amazon nor Decart commented on the reported deal.

The Real Value Is in the Compiler, Not the Models

Decart's public face is its generative AI products: Oasis, a simulation platform for robotics and autonomous-vehicle development, and Lucy, a real-time video transformation system capable of editing live 1080p footage at 30 frames per second. Both attracted attention when an earlier Oasis demo, a real-time Minecraft-style environment that reached one million users in three days after its October 2024 launch, became one of the fastest-adopted AI consumer products of that year.

But the product that makes Decart interesting to Amazon is neither of those. It is DOS — the Decart Optimization Stack — a software layer that tunes AI models for specific chip architectures and, crucially, does so across multiple chip families. Where traditional AI optimization commits developers to months of work to port a neural network from one chip to another, Decart says DOS reduces that process to a matter of weeks by automating the generation of custom kernels and tuning compiler behavior for the target hardware. The stack runs across Nvidia GPUs, Google TPUs, and Amazon's Trainium accelerators, according to Decart's product documentation.

The significance of this is structural. Nvidia's CUDA ecosystem — the combination of its GPU hardware, the CUDA parallel computing platform, and the cuDNN deep learning library — functions as a soft lock-in mechanism for AI developers. Optimizing a large model for CUDA requires significant engineering effort, and the output of that effort does not transfer to a competing chip without substantial rework. That switching cost is one of the primary reasons that Nvidia controls an estimated 80–90% of the AI training hardware market despite the proliferation of competing accelerators from Amazon, Google, and AMD.

DOS attacks exactly that switching cost. If Amazon can embed DOS's capabilities into the Trainium developer ecosystem, it becomes materially cheaper for enterprise AI labs to move workloads from Nvidia GPUs to Trainium — not because Trainium suddenly outperforms the H100 or B200 in raw throughput, but because the optimization labor that once trapped developers in the Nvidia ecosystem can now be automated and redirected.

The parallels to another recent deal are instructive. In 2026, Qualcomm acquired Modular — the company behind the MAX inference runtime and the Mojo programming language — for approximately $4 billion, giving Qualcomm a software layer that let AI models run efficiently across Qualcomm silicon without chip-specific rework by developers. Amazon appears to be following the same logic: the chip-portability software layer is worth acquiring at a premium because it multiplies the commercial value of the proprietary chip beneath it.

Read more: AMD acquires World Labs for $8.2 billion in AI chip software strategy

How DOS Works Below the API

Decart's optimization approach operates at the low level of the chip stack — closer to hardware than most AI frameworks. For a given neural network architecture, DOS generates custom CUDA kernels or equivalent low-level code targeting the specific instruction set of the destination chip, then tunes communication primitives for multi-accelerator configurations. The framework handles both inference and training workloads, and Decart says it offers milestone-based pilots for hyperscalers and AI labs that want to customize the optimization for their specific workload profiles.

The performance claims, if they hold under independent evaluation, would represent a meaningful efficiency gap. Decart says DOS 2.0, launched May 18, 2026, delivers over 1,600 tokens per second for agentic inference workloads — a figure the company contrasts against an industry average it puts at roughly 200 tokens per second, implying an eightfold throughput improvement. For video and world model workloads, DOS claims full-HD inference at up to 100 frames per second, according to Decart's product documentation.

Model FLOPS Utilization — the fraction of a chip's theoretical compute capacity that is productively used, rather than wasted on memory access bottlenecks, communication overhead, or idle cycles — is the key metric here. The industry standard for large AI workloads typically sits between 40 and 50 percent, meaning that well over half of a chip's compute potential is not reached in a typical training or inference run. Decart claims greater than 80 percent MFU on Trainium, a figure that Amazon's own Annapurna Labs chip team co-stated in the context of Lucy's deployment on Trainium3. AWS and Decart have not published detailed methodology for that benchmark figure.

That co-statement lends the figure more credibility than a purely unilateral vendor claim, though Annapurna Labs is not a disinterested party: AWS has a commercial agreement with Decart and a financial interest in validating Trainium's performance story. No independent third-party benchmark of DOS's MFU figures has been published.

Understanding why those MFU numbers matter requires understanding what chip porting actually involves at the hardware level. When an AI lab trains a large language model on Nvidia H100 GPUs, the engineers writing the training code ultimately produce CUDA kernels — GPU programs optimized for Nvidia's specific instruction set, memory hierarchy, and execution units. A well-written kernel fuses multiple operations (for instance, combining a matrix multiplication with its subsequent normalization step) to avoid loading the intermediate result back to slow DRAM before the next operation can proceed. It tunes blocking strategies so data fits within the GPU's fast on-chip SRAM. It schedules communication between GPUs in a cluster to hide latency. Each of these choices is hardware-specific: the optimal tiling strategy for an H100's memory architecture may be counterproductive on Trainium2, whose memory hierarchy and execution pipeline differ substantially.

Traditional chip porting requires engineers to re-derive these choices from scratch for the new hardware — a process that typically takes a team of experienced GPU programmers six months to a year for a major production model. The result is that even when a competing chip offers better price-to-performance in theory, the cost of accessing that performance in practice deters all but the most committed operators. DOS automates this derivation. According to Decart's own description, the system profiles the neural network's computational graph, identifies the bottlenecks for the target chip architecture, and generates tuned kernels automatically. The company says it also handles multi-chip communication tuning, which matters for any workload distributed across more than a handful of accelerators. The weeks-versus-months claim reflects this automation: where a human engineer must understand the chip architecture well enough to tune against it by hand, DOS encodes that knowledge in the compiler and applies it programmatically.

For comparison, the $4 billion Modular acquisition by Qualcomm — covering similar territory in multi-chip inference optimization — suggests the market values this software category at the higher end of AI startup prices. Amazon's reported $7 billion bid implies it assigns DOS additional strategic leverage specific to the Trainium roadmap.

A Three-Way Bidding War and the Deal That Fell Apart

Amazon is at least the third major technology company to pursue Decart in 2026, and the second to reportedly make an offer that reached the letter-of-intent stage. The sequence matters because it illuminates both the desirability of Decart's technology and the specific conditions Amazon may face.

Nvidia held early acquisition discussions with Decart, according to reports in Israeli technology press, but talks collapsed after larger offers from other parties emerged. Decart's existing relationship with Nvidia as both a customer and an investor — Nvidia participated in the $300 million Series B alongside Radical Ventures, Adobe Ventures, Toyota Ventures, and angel investor Andrej Karpathy — made Nvidia's position unusual: a chip company acquiring a startup that helps companies escape dependence on chips like those it sells.

Anthropic's approach proved more substantive. Bloomberg reported in August 2026 that the AI safety company had entered advanced talks at a valuation of approximately $6 billion — a figure that Reuters subsequently corroborated. Anthropic signed a letter of intent. Then, in September, talks collapsed.

Anthropic's interest in Decart was strategically coherent on its face. The company, which filed a confidential S-1 for an anticipated Nasdaq listing in 2026, has been expanding its inference infrastructure to support both its API business and its Claude.ai consumer product. Acquiring DOS would have given Anthropic a chip-agnostic optimization layer at a moment when the company is deepening its own hardware commitments — Anthropic has made public its work with Google on TPU optimization and has separately been exploring Trainium deployments. A startup that could reduce the engineering cost of moving workloads across chip families would have been useful infrastructure for a company managing multiple chip relationships simultaneously, and the $6 billion valuation would have been defensible against Anthropic's Series E valuation of roughly $60 billion at the time.

The reason the deal collapsed, according to reporting by the Israeli financial outlet The Deep Dive, was a demand from Anthropic that the founders and significant portions of the engineering team relocate from Tel Aviv to California. Decart's founders — Dean Leitersdorf, 26, who holds a PhD in computer science earned at 23; his brother Orian Leitersdorf; and Moshe Shalev, a 13-year Unit 8200 veteran — refused. The company's roughly 100 engineers are based primarily in Israel, and the founders had no intention of dismantling that. Bloomberg separately reported that Anthropic may also have come away from due diligence insufficiently impressed with the technology, or concluded that it could build equivalent capabilities internally or acquire them more cheaply elsewhere. The reporting places the relocation demand as the primary breaking point.

Amazon now holds a structural advantage in this specific negotiation: its deal rationale is primarily about Trainium adoption, not about folding Decart's team into an AI research organization. The question of whether Amazon will also insist on relocating talent has not been reported, but the company's history with AI infrastructure acquisitions — including Adept, which it effectively hired away through a licensing structure — suggests a range of deal structures is possible.

The reported price escalation from Anthropic's $6 billion to Amazon's $7 billion, a roughly 17 percent increase in four weeks, reflects either competitive pressure from additional parties — SpaceX was mentioned in Calcalist reporting as a possible suitor, citing a relationship between Elon Musk and Dean Leitersdorf dating to Oasis's viral launch — or Amazon's willingness to pay a premium to ensure the deal closes before another bidder appears.

What Decart's Technology Actually Does in Production

Decart does not publish financial details, but the picture assembled from its Series B materials and public reporting gives a sense of the operational profile. The company earns what multiple sources describe as "several tens of millions" in annual revenue, primarily from enterprise licensing contracts with cloud providers and AI labs. Its consumer products — Oasis is free to users and has not been monetized — generate no direct revenue. Distribution through Amazon Bedrock since December 2025 has given Decart access to AWS's enterprise customer base without requiring Decart to build its own sales infrastructure.

Lucy is the most commercially mature product. The world model edits live video in real time using a self-anchoring mechanism: rather than independently generating each frame from scratch, Lucy captures its own previous output as a reference frame and uses it to maintain visual consistency across edits. This approach addresses a common failure mode in real-time video generation — accumulated drift, where small frame-to-frame errors compound into visual incoherence — by anchoring each new frame to the model's own recent history. Applications include virtual try-on experiences for retail, in-video advertising, and live-stream effects.

Oasis is technically more complex. The world model uses a transformer-based diffusion architecture trained with a diffusion-forcing scheme: rather than generating entire video sequences in one pass, it produces frames autoregressively, each conditioned on the user's input at that moment. Temporal attention layers interleaved with spatial attention layers allow the model to maintain context from previously generated frames. A ViT-based variational autoencoder compresses the visual signal to a lower-dimensional latent space before generation, which reduces the compute cost of each inference step. Details on the Oasis architecture are available at decart.ai.

In practice, this means Oasis can be steered by user inputs in real time — a design that made the original Minecraft-style demo immediately interactive rather than rendering a pre-computed animation. On an H100 GPU, the model runs at approximately 20 frames per second at 360p. On Trainium3, Decart and Amazon claim a four-times improvement in frame throughput at half the cost of equivalent GPU compute, though neither company has published the methodology or test conditions behind that figure.

For the robotics simulation use case — Oasis 3, released earlier this year — the architecture generates multi-camera driving environments for training autonomous-vehicle systems. Independent testing by Contrary Research found drift over long sessions, control lag, and some physics fidelity issues where objects passed through each other rather than colliding. These are known limitations of autoregressive video generation that affect the technology broadly, not Decart specifically, but they matter for AV simulation where flawed synthetic training data can be worse than no data.

Amazon's Trainium Ambitions and Why DOS Changes the Math

Amazon has been explicit about its chip ambitions. Jeff Bezos said publicly this year that custom silicon has the potential to become Amazon's next durable business pillar, and the company is exploring selling Trainium chips directly to external customers — a model that would transform Amazon from a chip consumer into a chip vendor competing with Nvidia for AI infrastructure revenue. Snowflake recently signed a $6 billion commitment with AWS for cloud infrastructure including its custom Graviton processors and AI accelerators, and Amazon has separately been operating a Cerebras partnership to offer large-format wafer-scale chips for inference workloads.

All of this ambition runs into a practical barrier: enterprise AI developers build their pipelines on Nvidia GPUs, optimize their models for CUDA, and have little incentive to port their workloads to Trainium without a compelling economic reason and a clear path through the porting process. DOS addresses the second part of that equation. If switching from an H100-optimized model to a Trainium-optimized model takes weeks instead of months, and if the resulting Trainium deployment achieves 80 percent or more of the chip's theoretical compute capacity, the economics of switching become plausible.

This is the arithmetic that appears to be driving the $7 billion price tag. The question is not whether Decart's world models and video technology are worth $7 billion on a standalone basis — at "several tens of millions" in annual revenue and a team of roughly 100 engineers, they are not. The question is whether integrating DOS into the Trainium developer ecosystem generates enough incremental Trainium revenue to justify the acquisition cost. For a cloud provider whose AI infrastructure capital expenditure runs in the tens of billions of dollars per year, a software tool that measurably accelerates enterprise adoption of its proprietary chip architecture could represent a return that dwarfs the acquisition price.

The Qualcomm-Modular comparison sets a floor for what the market thinks chip-portability software is worth as a standalone asset. Amazon's premium above that floor reflects Trainium-specific leverage: Decart already runs on Trainium, has achieved credible utilization figures, has an existing distribution relationship through Bedrock, and has demonstrated with Lucy what a production-grade deployment on the chip looks like.

The Independent Inference Optimization Landscape DOS Is Entering

DOS does not operate in a vacuum. The inference optimization space has grown substantially since 2023, and Decart's commercial positioning sits at an unusual intersection: more automated than existing open-source options, but far less hardware-agnostic in its actual deployment record.

The most widely deployed open-source inference engine is vLLM, developed by researchers at UC Berkeley. vLLM introduced PagedAttention — a memory management technique that treats the key-value cache in transformer inference as pages in a virtual memory system, dramatically improving GPU utilization for serving multiple concurrent requests. It is mature, well-documented, and runs primarily on Nvidia hardware, with some support for AMD ROCm. SGLang, also from Berkeley, extends this with structured generation features that speed up constrained decoding for agentic workloads. Neither project claims the multi-chip portability or MFU figures Decart does for Trainium, and both are built and maintained primarily for Nvidia-class silicon.

Nvidia's own TensorRT-LLM occupies a different position: it is Nvidia's in-house inference optimizer, tightly integrated with TensorRT's quantization and kernel tuning capabilities, and effectively proprietary to Nvidia GPUs. It produces highly optimized inference on H100s, but it exists explicitly to keep enterprise customers on Nvidia hardware. Google maintains XLA and the OpenXLA project, the compiler infrastructure used to optimize for TPUs and, increasingly, for other accelerators via MLIR-based lowering. XLA is powerful but requires significant engineering effort to apply outside Google's own stack.

DOS's commercial differentiation relative to these alternatives rests on two claims: that it achieves comparable or better MFU to hand-tuned Nvidia solutions on non-Nvidia hardware, and that it reduces the engineering effort for porting by automating what the open-source tools require manual expertise to accomplish. Neither claim has been independently benchmarked across all three chip targets the company advertises. What is documented is the Trainium deployment with Lucy — one production workload, on one chip family, co-validated by a commercially interested party. That is enough for Amazon to make an acquisition case; it is not yet enough for the broader industry to treat DOS as a proven cross-platform solution.

The consolidation risk matters here precisely because the independent alternatives remain thinner than the technology's strategic importance suggests they should be. If DOS moves inside Amazon and Modular moves inside Qualcomm, the remaining neutral inference optimization tools are largely open-source projects optimized for Nvidia hardware — which is exactly the competitive position Nvidia would prefer to see preserved.

The Founders' Question and What Comes Next

The deal's most uncertain variable is the one that already killed it once. Decart's founders are based in Tel Aviv, built the company around Israeli engineering talent drawn partly from Unit 8200 connections, and declined to move when Anthropic asked. Whether Amazon will structure a deal that preserves the Tel Aviv R&D center, or whether it will attempt some version of the relocation approach that Anthropic could not complete, is not reported.

Deals in which Israeli tech founders retained operating independence after acquisition — several precedents exist — tend to close at a modest discount to deals that require relocation, because the buyer accepts ongoing operational complexity in exchange for retaining the team. If Amazon is willing to pay $7 billion for an operational structure that leaves most of the engineering in Tel Aviv, that premium itself signals how seriously it takes the DOS technology.

The Wall Street Journal's characterization that a deal could close within days means the resolution of those details is close, or that the parties have already agreed to structure the deal in a way that avoids the relocation impasse that ended the Anthropic talks. Either way, the speed of the reported timeline, coming just weeks after Anthropic's departure, suggests Amazon entered the negotiations with a clear position rather than exploratory interest.

The deal will also face regulatory scrutiny. A transaction above $119.5 million triggers mandatory Hart-Scott-Rodino Act filing in the United States, and a $7 billion deal involving one of the nation's largest technology companies will almost certainly receive substantive review from the Federal Trade Commission. The FTC's approach to big-tech AI acquisitions has shifted toward structural skepticism over the past two years: the agency scrutinized Microsoft's partnership with Mistral, investigated Nvidia's acquisition of Run:ai, and reviewed Amazon's own Anthropic investment. Amazon structured at least one prior AI relationship — its arrangement with Adept, the agent-software company — as a talent and licensing deal rather than a full acquisition, a structure some observers interpreted as designed to reduce merger-notification exposure. Whether Amazon pursues a similar licensing structure with Decart, or moves to a full acquisition despite the regulatory timeline, will affect both the deal's closing speed and its final financial terms. The European Union's Digital Markets Act and the UK Competition and Markets Authority are also likely to review a transaction of this size involving a gatekeeper-designated cloud provider.

For the broader AI infrastructure market, the outcome will signal whether chip-portability software is consolidating into hyperscaler ownership — which would limit its availability as a neutral tool — or whether independent alternatives will continue to develop. Modular is now inside Qualcomm. If DOS moves inside Amazon, the remaining independent inference optimization players will face a landscape in which two of the field's most credible assets are proprietary to chip companies with competing interests.

What Decart's founders decide to do with a potential $7 billion outcome, and where their next research push goes after the acquisition, may matter as much to the future of real-time AI video and physical AI simulation as the deal structure itself. Dean Leitersdorf has said repeatedly that the company's goal is to build one of the deepest AI labs in the world from Israel. How that ambition survives inside Amazon's cloud infrastructure business remains the most open question in a deal that has not yet been signed.

The AI infrastructure acquisition wave of 2026 — AMD's purchase of World Labs, Qualcomm's Modular deal, and now Amazon's reported move on Decart — is concentrating the tools that control chip access inside the companies that also control the chips. Each deal individually has a clear rationale. The cumulative effect is a narrowing of the independent software layer that once allowed AI developers to build against any hardware they chose. If the trend continues, the realistic path to competitive inference on non-Nvidia hardware will increasingly run through a hyperscaler's proprietary stack. That outcome is precisely what Amazon is betting on — and precisely what the rest of the industry should be watching most closely.