Amazon is in talks to acquire the AI company Decart in a deal that could be worth about $7 billion, the Wall Street Journal (WSJ) reported on October 11, 2026. The report is based on people familiar with the matter, and no acquisition agreement has been confirmed. Decart is already an Amazon customer and runs its video AI models on AWS's custom Trainium3 chip. If the relationship moves from partnership to acquisition, what Amazon would gain is a set of video-generation models and the engineering capability to extract performance from chips for those models.
A $7 billion negotiation puts video and compute infrastructure in focus
Decart was founded in 2023 and announced a $300 million raise led by Radical Ventures on May 18, 2026. NVIDIA also joined as a new investor, and the company says its total funding now exceeds $450 million. According to the WSJ, citing PitchBook data, the company's valuation in May was about $4 billion. The $7 billion figure refers to the size of the reported acquisition talks and should be distinguished from past funding amounts or valuations.
Decart's business consists of the Decart Optimization Stack (DOS), which speeds up AI training and inference; Lucy, which edits live video; and Oasis, a world model that generates environments. The company builds both the models and the systems that run them. That combination is a useful key for reading the acquisition report.
Whether the deal will happen is still unclear. On September 8, Bloomberg reported that Anthropic had considered acquiring Decart for about $6 billion but dropped the idea after conducting due diligence on its assets and business. The WSJ likewise says the situation could change if negotiations collapse or another buyer emerges. There is a gap between reports of talks toward a high-priced deal and a company actually coming under new ownership.
DOS handles optimization between chips and models
Decart says DOS 2.0 supports NVIDIA GPUs, Google TPUs, and Amazon Trainium. It runs Decart's own Lucy and Oasis, and the company also licenses it to cloud providers and AI developers. Customers get optimization technology that works across different chips, but describing it simply as general-purpose conversion software doesn't capture what it does.
DOS's official description lists kernel development and compiler tuning tailored to each customer's workloads and evaluation metrics. A kernel is an individual computation that runs on a chip, and a compiler is the mechanism that converts a program into a form the chip can execute. Beyond getting a model to run, the work involves tuning which computations to group together and where to place data to make things faster.
Decart also changes the model side to suit the chip. In its technical explanation of Lucy 2, the company says it measures fine-grained accelerator performance and designs the model architecture based on the results. It is a development approach that tunes models and hardware together, rather than simply porting an existing model to another chip.
That difference matters for developers using Trainium. Even if a trained model can run, it cannot use the chip's full capability if the compute units are waiting for data to arrive. DOS aims to reduce such waiting time and wasted processing. Support for multiple chips should not be read as a guarantee that models can be ported without tuning or that every chip delivers the same performance.
Cutting compute and communication time for live video
Lucy 2.5, which Decart announced on July 16, is described as a model that edits people, clothing, and surroundings in a live stream at 1080p and 30 frames per second. Unlike returning a finished video later, it has to keep updating the footage in response to viewers' movements and instructions.
Displaying 30 frames per second means a cycle of about 33 milliseconds (1,000 ms divided by 30). That is the display interval, not the model's inference time or the round-trip latency including communication. Still, it shows that keeping latency low in live video involves burdens beyond computation.
The Lucy 2 explanation describes combining multiple computations into larger kernels to reduce the overhead of launching operations and memory transfers. Rather than moving data back and forth to HBM, the high-bandwidth memory, it keeps data close to where computation happens. Decart also tunes the WebRTC mechanism that sends and receives video to cut extra buffering and communication delay. Speed is not determined by the compute unit's processing power alone.
Lucy 2.5 adds quantization, which lowers the numerical precision used in computation, and a technique that skips unnecessary attention computation. Attention determines which information in the video to reference. The model combines efforts to reduce its workload with efforts to run processing efficiently on the chip.
However, video that looks physically natural does not prove that physical phenomena are being calculated accurately. Decart explains that Lucy 2 is a diffusion model that does not depend on explicit 3D meshes or a physics engine. The company is also expanding its use to generating data for robotics, but how convincing the footage looks and how effective it is for real-world training are separate questions that need to be verified independently.
The AWS relationship already spans chips through sales
In Decart's May 18 announcement, Amazon appears as a strategic customer. Decart has a partnership with AWS built on a commercial agreement and joint sales, and also works with Annapurna Labs, AWS's chip development unit. Sorting the relationships described in the announcement by role—fit to the compute platform, running the model, and selling to customers—gives the following.
| Point of contact | What has been disclosed | What it means for Amazon |
|---|---|---|
| Chip optimization | Decart tailored its model architecture to Trainium's design | Development experience in making full use of its own chip's characteristics |
| Running the model | Lucy 2 runs on Trainium3 | A real-time video use case in actual operation |
| Sales to customers | Commercial agreement and joint sales with AWS | A route to deliver the models to AWS customers |
The facts in the table are based on Decart's May 18, 2026 announcement; the classification of roles and meanings is ours. The connection between Decart and AWS spans three layers: model optimization for Trainium3, running Lucy 2, and a commercial agreement with joint sales. This is work already under way through the partnership, and the table does not show any post-acquisition organization or contracts.
AWS itself has a development environment for running models on Trainium. The December 2, 2025 announcement of Trn3 UltraServers describes PyTorch integration in the Neuron SDK, as well as features giving deep access to the chip so that engineers pushing for performance can tune kernels. In weighing DOS's value, it is necessary to look at how far specific models and workloads can be optimized on top of this environment.
In a comment accompanying Decart's May announcement, Nafea Bshara, vice president at Annapurna Labs, said Lucy 2's "model compute utilization" exceeds 80%. This is a metric for how much of the chip's theoretical capability the model's computation uses; it does not mean costs fall 80% for every customer. It applies to Lucy 2, and detailed measurement conditions were not given.
On AWS's customer case study page, Decart CEO Dean Leitersdorf says initial tests cut latency from 40 milliseconds to 10 milliseconds. However, the GPU model compared and the Trainium generation are not specified. This figure cannot be treated as the same measurement as Lucy 2's utilization rate, nor as evidence that Trainium is superior for every model.
Under Amazon, what happens to support for other companies' chips?
NVIDIA has invested in Decart, and Amazon is a strategic customer. DOS supports both companies' chips as well as Google's TPUs. Decart is thus a company that holds both a position of drawing out performance across chip makers and a position of tuning deeply for a specific chip.
If Amazon acquires it, chip design and model development could proceed in even closer coordination. For example, when improving a video model, engineers could restructure computation based on the chip's memory configuration and feed the results back into the next hardware or software improvements. A key technical question is how far ownership could deepen the co-design that has already been disclosed.
On the other hand, it also raises the question of how Decart would keep providing technology to customers using GPUs or TPUs. Focusing on Trainium could advance improvements to Amazon's own chip, but the DOS business would change depending on how it handles the development experience gained across multiple chips and its relationships with outside customers. There has been no report that support for other companies' chips will end or become exclusive; this is an issue that would arise if an acquisition happens.
Assessing whether $7 billion is reasonable would require revenue, contract terms, and the effects of integration. The information published so far cannot justify that price. If a formal agreement is announced along with a policy on supporting other companies' chips, and optimization extends to workloads from customers beyond Lucy, it will become possible to evaluate concretely what development capability AWS would gain and how much DOS is worth as technology usable across chips.