AMD has agreed to acquire World Labs, the AI research company co-founded by Fei-Fei Li, for approximately $8.2 billion in stock. The deal is expected to close by the end of 2026, subject to regulatory approval and other closing conditions. It puts a business question at the center of visual AI: why does a chipmaker want its own model research lab?
World Labs develops world models that represent environments and support tasks such as generating a scene or predicting another view of it. Its work points toward visual tools that let users explore and work within a generated space.
AMD wants that research to help shape its future computing systems. NVIDIA is pursuing a related opportunity through Cosmos. For teams working with visual AI, these investments raise expectations for more capable tools. Whether those tools become reliable and affordable enough for everyday work remains the test.
Why AMD Wants a Model Lab
Understanding What Future AI Will Need
Designing hardware for emerging AI applications requires knowing where models run into problems. A research team building those models can help identify what slows them down and which changes would make them run more efficiently.
World Labs’ researchers would help the company understand how computing needs are changing and inform its future hardware and software.
“Building the compute platforms for the next generation of AI requires a deep understanding of how models are evolving,” AMD chair and CEO Lisa Su said in the announcement.
The companies have already worked together on this. In her account of the agreement, Fei-Fei Li says a technical partnership began the previous year, focused on training models and optimizing how they run on AMD GPUs. The proposed acquisition would bring that collaboration inside AMD.
Making the Hardware Easier to Use
For a team developing a visual application, a powerful processor is one part of the setup. The software must also support the model and help developers run it efficiently. Difficult setup or unreliable software can make otherwise capable hardware expensive to work with.
World Labs already works with AMD on training and running models on its GPUs. Bringing that research team into AMD could give its software engineers more direct feedback on the problems researchers encounter as they develop new models.
AMD could use that feedback to make models easier to run on its hardware. Faster generation and lower costs are potential benefits, but it is too early to say whether the acquisition will deliver them.
NVIDIA Is Building Around the Same Opportunity
NVIDIA is on a similar path, developing world models through Cosmos. Introduced at CES in January 2025, Cosmos gives developers world foundation models they can adapt for physical AI, including robots and autonomous vehicles.
The models come with tools for preparing training data and customizing them for particular tasks. A robotics team, for example, can use recordings of a robot navigating a warehouse to adapt a model to the environment where it will operate. Cosmos also supports generating synthetic scenarios for training and evaluation.
This creates several opportunities for NVIDIA to supply computing power. Preparing video datasets, training models, and generating new footage all require processing resources. NVIDIA offers tools optimized for its GPUs, alongside deployment options and enterprise support. The commercial logic is that helping more teams build these applications can create more demand for the hardware that runs them.
AMD and NVIDIA both invested in World Labs’ $1 billion funding round in February 2026. AMD now plans to buy the company, while NVIDIA is developing Cosmos alongside its investments in other model developers.
What Visual-AI Teams Can Use Today
Exploring a Generated Scene
World Labs’ Marble lets users generate a 3D environment from inputs such as text or images, then explore it from different viewpoints. Users can also download the environment for further editing.
Consider a creative team planning a scene inside a fictional café. The team could generate a setting, then move through it to compare a camera position near the entrance with one behind the counter. This gives them a way to explore framing within the same space before committing to a shot.
Generating each viewpoint as a separate image can introduce differences in the room’s layout. A persistent 3D environment gives the team a shared setting to inspect, although they still need to check its geometry and suitability for production. Being able to navigate a space does not establish that its dimensions match a real location.
Generating Scenarios for Training
NVIDIA Cosmos supports a different workflow: generating footage that developers can use to train and evaluate physical AI systems. NVIDIA describes using Cosmos with controlled 3D scenes built in Omniverse to produce variations of driving environments.
For example, a team developing a driving system could generate versions of a road scene under different weather conditions. Those variations could help expand the situations represented in its training data, including conditions that are difficult to capture repeatedly.
The important question is whether each variation preserves what the task requires. Before using it for training, the team needs to check that the generated footage provides examples that are actually useful.
“Right now world models have really excelled at traversing environments, especially those without a lot of moving elements. However, their biggest gap is around object permanence. When objects are offscreen or occluded, most world models fail to predict how the objects will behave.” - Mark, Sr Project Manager 3D at Wirestock
Why the Training Data Still Matters
What Useful Footage Needs to Show
The task a model needs to learn determines what its training footage should contain. A clear view of an object may help with recognition. Learning how that object moves requires examples that show the action unfolding.
Consider a video of someone placing a box on a shelf. If the footage cuts from the person holding the box to the box resting on the shelf, it leaves out the moment of contact. A continuous shot shows how the person supports the box, lowers it onto the surface, and releases it. For a task involving that interaction, those details matter more than a polished final frame.
“Everything comes down to the metadata and context for a clip. The industry is filled with scraped video data that strips away all context besides the pixels themselves. However, if you are intentional with what data is annotated (especially in synthetic data), you can unlock so much potential in SFT and RLHF training stages.” - Mark, Senior Project Manager 3D at Wirestock
Selecting video datasets for machine learning therefore starts with the intended use. A collection of attractive clips may still omit the interactions a project needs. Licensed visual data providers such as Wirestock offer a reliable way to source material.
Generated Data Needs Evaluation
Synthetic footage can add variations that are costly or difficult to record. Its value depends on whether those variations preserve the action being taught.
In the shelf example, changing the lighting could broaden the training examples. But if the generated box passes through the shelf or floats after release, the clip introduces an incorrect interaction. Review should check the sequence and contact, alongside visual quality.
What Still Has to Be Proven
The larger bet behind the AMD World Labs acquisition is that generating and simulating environments will become a regular part of how companies work. If that happens at scale, these applications could create substantial demand for computing power and influence how chipmakers design future hardware.
Futurum’s analysis argues that memory could become especially important. Generating a longer visual sequence means handling information across more frames. Keeping an environment consistent as someone interacts with it can add further demands. How much information a system can hold, and how quickly it can move that information, can affect generation time and cost.
Futurum sees this as a potential advantage for AMD’s hardware strategy. It also acknowledges that applying video generation benchmark results to world models remains a thesis. A strong result on one video task does not establish how a system will perform across interactive environments.
This raises further commercial questions. Will customers use these tools often enough to support a substantial market? Will closer collaboration between model researchers and hardware engineers make them faster and cheaper to run? And will their outputs meet the accuracy requirements of the work?
What to Check Before Building Around These Platforms
Start with a small pilot drawn from an actual project. Define what a usable result must do before generating anything.
For a creative team, that could mean producing an environment for one planned camera move. Check whether the scene holds up along the full route and whether the team can make the edits the shot requires.
“Make sure to demo the world model in the specific context in which you will need it. Many demos of World Models look very impressive on short clips with the camera slowly moving forward, but if you run your own scenarios you may find that the models will break down in surprising ways. Make sure to test with your intended clip length, camera motion, and environment types.” - Mark, Senior Project Manager 3D at Wirestock
Use the pilot to assess:
Time and effort: Track generation, revisions, cleanup, and the work needed to bring the output into production.
Export options: Check whether the output works in the team’s existing tools and preserves the elements they need to edit.
Licensing: Confirm that the terms allow the intended commercial use.
Hardware and cost: Identify the computing resources required and estimate costs at the expected volume.
Flexibility: Test whether the team could move its assets and workflow to another platform if needed.
Repeat the pilot across several representative scenes. This gives the team a clearer basis for deciding whether the platform can deliver usable results consistently.
Sources
AMD. (2026, September 28). AMD to acquire World Labs to advance the future of AI compute.
Burke, B. (2026, September 30). AMD acquires World Labs to build silicon around world models. The Futurum Group.
Li, F.-F. (2026, September 28). To seek a newer world: Our journey from World Labs to AMD. Substack.
NVIDIA. (2025, January 6). NVIDIA launches Cosmos world foundation model platform to accelerate physical AI development.
Parker-Holder, J., & Fruchter, S. (2025, August 5). Genie 3: A new frontier for world models. Google DeepMind.
Robbins, J. (2026, October 2). AMD’s $8.2B World Labs deal signals buy-over-build era for world models. PitchBook, via Yahoo Finance.
Singal, N. (2026, September 29). AMD agrees to buy World Labs to fill out its AI stack. CIO.
World Labs. (2026, February 18). World Labs announces new funding.
World Labs. (n.d.). Marble.
Why did AMD buy World Labs?
What does World Labs do?
Who is building AI world models?
