Amazon set to acquire Decart for $7 billion: the masterstroke in real-time video AI and compute abstraction
🔎 Why this deal is a game-changer for generative AI
Amazon is in negotiations to acquire Decart for approximately $7 billion, according to the Wall Street Journal and Reuters (October 11, 2026). The Israeli startup, valued at $4 billion last May, has doubled its value in five months. Anthropic had walked away from talks at $6 billion in September after due diligence. This is no simple acqui-hire: Amazon is getting its hands on the software stack that makes it possible to run any video model or world model on any chip — Trainium, Nvidia, AMD — without rewriting the code. The stakes go beyond video: this is compute abstraction at hyperscale.
Key takeaways
- Amazon is in talks to buy Decart for ~$7B (WSJ, Reuters, Oct 11, 2026), nearly double its May 2026 valuation ($4B post-round of $300M led by Radical Ventures with Nvidia).
- Decart owns three strategic assets: Lucy (real-time video editing < 30 ms), Oasis (interactive world model for robotics/autonomous driving), and DOS (Decart Optimization Stack), which abstracts the hardware.
- Anthropic walked away from ~$6B in September 2026 after weeks of due diligence (Bloomberg, Sept 8), leaving the door open for Amazon, which had already granted early access to Trainium3 back in 2025.
- The compute angle: DOS reduces training and inference costs by optimizing chip utilization, regardless of the silicon — a direct lever against Nvidia lock-in.
Recommended Tools
| Tool | Main Use | Price (October 2026) | Ideal For |
|---|---|---|---|
| Decart DOS | Compute abstraction & inference optimization | Quote-based (enterprise license) | Hyperscalers, AI labs, cloud providers |
| Lucy 2.0 | Real-time video editing, try-on, in-video ads | Usage-based paid API | E-commerce, streaming, gaming, advertising |
| Oasis 3 | Real-time physical world simulation | Strategic partnerships | Robotics, autonomous vehicles, manufacturing |
| AWS Trainium3 | High-performance AI training/inference | Hourly instance pricing by region | Foundation model training, fine-tuning |
Why Amazon is betting 7 billion on Decart — Direct Answer
Amazon is buying a hardware abstraction layer that makes its Trainium chips competitive against Nvidia GPUs without asking customers to rewrite their kernels. DOS (Decart Optimization Stack) compiles and optimizes compute graphs for any architecture — CUDA, ROCm, Neuron — while preserving native performance. It's the missing lever to get AWS customers to migrate en masse to Trainium. The 7 billion price tag reflects the scarcity of a stack that works in production at hyperscalers, not in demos.
The strategic lever: silicon independence
AWS has invested billions in Trainium and Inferentia. But as long as the software ecosystem (PyTorch, Triton, custom kernels) remains optimized for CUDA, customers hesitate. Decart solves this problem at the source: DOS takes a standard PyTorch model and generates optimized code for the target — Trainium, H100, MI300 — without human intervention. Amazon thus turns a software disadvantage into a competitive advantage. The pre-existing collaboration (Trainium3 access since 2025 according to EIA Media) proves that the technical integration has already been validated.
From $4 billion to $7 billion in five months: the valuation escalation
The trajectory is dizzying: $0.5B (Series A, $32M led by Benchmark) → $3.1B (August 2025, $100M round) → ~$4B (May 2026, $300M led by Radical Ventures, with Nvidia joining) → ~$7B (Amazon negotiations, October 2026). Each round brought in a strategic investor: Sequoia (seed), Benchmark (Series A), Radical + Nvidia (Series B). Nvidia in the cap table of a company that abstracts away its hardware is a strong signal — either insurance, or a technology watch.
The real revenue behind the valuation
Unlike many pre-revenue AI startups, Decart "generates significant revenue through contracts with cloud providers, AI labs, and hyperscalers that license DOS" (official Decart.ai website). This isn't projection: it's recurring. Customers pay to cut their compute bill by 30 to 50% depending on the workloads. At the scale of AWS or Azure, the economics justify the multiple. Anthropic understood this when it entered talks at $6B — before backing out.
Lucy and Oasis: two pillars for real-time video and Physical AI
Decart doesn't just sell optimization. It owns two flagship models that demonstrate the power of its stack.
Lucy: video editing under the 30 ms bar
Lucy 2.0 is described by Decart as "the only real-time editing world model in production at scale." Cited use cases: virtual try-on, dynamic in-video advertising, live streaming, gaming. Latency < 30 ms opens up use cases that are impossible with traditional diffusion models (seconds to minutes). For Amazon, this means Prime Video, Twitch, and programmatic advertising — billions of hours of video where real-time contextual insertion changes the economics of attention.
Oasis 3: the world model for Physical AI
Oasis 3 is presented as "the first interactive world model designed for Physical AI — interactive, physically accurate, real-time environments for robotics, autonomous vehicles, manufacturing, drones" (Decart.ai/research). Decart has also released a world model simulating hours of photorealistic driving, and collaborates with Comcast on edge inference (< 35 ms). Amazon Robotics, Zoox, and warehouse logistics are natural customers. The acquisition locks in an asset for the next wave: AI that acts in the physical world.
DOS: the abstraction layer that scares Nvidia
The Decart Optimization Stack is the most underrated — and most valuable — asset. DOS operates at three levels: graph compilation (operator fusion, memory layout), kernel auto-tuning (configuration space search via RL), and multi-chip scheduling (pipeline parallelism, compute/communication overlap). All of it exposed via a PyTorch 2.0 / torch.compile-compatible API.
Why this is a game-changer for Trainium
Trainium3 exposes primitives (BF16, FP8, structured sparsity, custom tensor cores) that standard compilers don't fully exploit. DOS targets them natively. The result: according to internal benchmarks cited by Decart, inference on Trainium3 with DOS reaches 85-95% of H100 performance on the same models — at a per-chip cost 3 to 4 times lower. Amazon doesn't need to beat Nvidia on raw silicon; it needs to beat the total cost of ownership. DOS is the lever.
The leverage effect on the ecosystem
If Amazon integrates DOS natively into SageMaker, Bedrock, and the Deep Learning Containers, every AWS customer benefits without any migration effort. It's the equivalent of what CUDA did for Nvidia in 2006-2010: locking in the ecosystem through convenience. But here, the abstraction works against lock-in — it makes the hardware interchangeable. An irony Amazon knows how to turn to its advantage.
The Anthropic Precedent: Why the Deal Fell Through at 6 Billion
Anthropic conducted several weeks of due diligence before backing out in early September 2026 (Bloomberg, Sept 8; Benzinga; SiliconRepublic). Two non-exclusive hypotheses: (1) technical risks around generalizing DOS beyond video/world model workloads (LLMs, MoE, retrieval); (2) a cultural or strategic divergence — Anthropic is betting on its own inference stack (Claude, AWS/Google Cloud partnerships) and perhaps didn't need an abstraction layer it doesn't fully control.
What Amazon Is Buying That Anthropic Refused
Amazon doesn't have the same conflict of interest. AWS sells compute. Lowering the unit cost of inference for its customers increases the volume consumed — the business model is aligned. Moreover, Amazon has the distribution (millions of AWS customers) to amortize the investment within a few quarters. Anthropic, on the other hand, has to recoup its investment through its own models. The failed $6B deal validates the price: a rational strategic buyer almost signed. Amazon is paying a control premium ($1B more) to avoid letting the asset slip away to Google or Microsoft.
AWS and Trainium3: a technical relationship that predates the acquisition
According to EIA Media and Yahoo Finance (via Reuters), "in 2025, AWS's cloud division had given Decart early access to its own Trainium3 chip." This is no small detail: AWS doesn't share its next-gen silicon with just anyone. Decart had time to port DOS, optimize the kernels, and validate the gains. The acquisition formalizes an integration that was already operational.
The signal for AWS customers
Enterprise customers who were hesitant to adopt Trainium ("will my models run?") now have an answer: Decart guarantees it. Amazon can include DOS in SageMaker performance SLAs. It's a massive selling point against Azure (NDv5-series H100) and GCP (A3 Ultra). The AI cloud war is fought on ease of migration — Amazon just radically simplified it.
❌ Common Mistakes
Mistake 1: Mistaking Decart for yet another video generation startup
Decart doesn't compete with Sora, Veo, or Kling. It provides the infrastructure that lets those models run efficiently on any hardware. Lucy and Oasis are demonstrations of the stack, not the main product. The product is DOS.
Mistake 2: Underestimating the technical lead in hardware abstraction
Projects like Apache TVM, MLIR, or Triton exist. But Decart has shipped in production at hyperscalers with measurable gains. The difference between a research compiler and a supported enterprise stack is the same as the difference between an open-source model and a managed service with an SLA.
Mistake 3: Believing Anthropic backed out for purely financial reasons
6 billion for a company valued at 4 billion post-money (May) is a rational price if the asset delivers on its promises. Withdrawing after due diligence suggests technical or strategic discoveries — not a disagreement over the multiple. Amazon, with its own due diligence, judged the risk acceptable.
❓ Frequently Asked Questions
When will the Amazon-Decart deal be finalized?
No official date yet. The "negotiations" (WSJ) still involve due diligence, regulatory approvals (CFIUS likely given the founders' Israeli origins), and closing. Expect 3-6 months if no major obstacle arises.
Will Decart remain independent or be absorbed into AWS?
Probably integrated as an AWS AI/ML team (in the style of Annapurna Labs for network chips). The founders (Dean/Orian Leitersdorf, Moshe Shalev) are used to working closely with AWS (Trainium3 2025). Product autonomy for Lucy/Oasis is likely; DOS becomes a native service.
What impact for current Nvidia customers on AWS?
No immediate change. DOS supports CUDA. Customers can choose to migrate to Trainium to save money, without rewriting their code. It's an option, not a constraint.
Can Microsoft or Google counterattack?
They will — either by developing their own abstraction layer (Microsoft has Olive, ONNX Runtime; Google has XLA/PJRT), or by acquiring a competitor (Modular, OctoML, or internal teams). But Decart has a 2-3 year head start in production. The acquisition locks in this advantage for Amazon.
Does this change the Trainium4 roadmap?
Yes. Amazon can now co-design the software and the compiler in a closed loop. Trainium4 will likely integrate primitives exposed directly by DOS (dynamic sparsity, adaptive quantization), creating a virtuous hardware-software cycle that Nvidia has mastered for 15 years.
✅ Conclusion
Amazon is paying 7 billion to buy itself the key that unlocks its own chips: a software layer that makes hardware interchangeable without sacrificing performance. Decart isn't a video startup — it's the infrastructure that will make Trainium a credible alternative to the H100 for inference at scale. Anthropic's failed deal validates the price; early access to Trainium3 validates the integration. For AWS customers, it's the end of the "but my code is CUDA-optimized" excuse. For Nvidia, it's the first serious software competitor since CUDA. Genesis Mission: $2.4 billion in federal AI compute — Nvidia, AMD and OpenAI commit to American science shows that the compute battle extends beyond the private sector — Amazon has just picked up a key piece.