She built the database that taught machines to see. She ran artificial intelligence research for the United States government during the Obama administration. She cofounded a $1 billion startup that spent two years attempting to give machines the ability to understand — not just describe — the physical world. Now, Dr. Fei-Fei Li is joining AMD.
On Sunday evening, AMD announced it would acquire World Labs, the spatial-intelligence startup Li founded in 2024, in an all-stock transaction valued at approximately $8.2 billion. Li will join AMD as executive vice president and chief scientist, reporting directly to CEO Dr. Lisa Su. The deal is expected to close by year-end, pending regulatory approval.
The announcement lands as the AI industry’s competitive dynamics are shifting from pure text and code generation toward a harder, more consequential frontier: physical AI. Building machines that can navigate hospitals, manufacture cars, unload shipping containers, and rescue people from burning buildings requires more than a large language model. It requires something World Labs has been quietly building for two years — a system that can model, simulate, and reason about the physical world as it actually behaves.
What World Labs Actually Built
World Labs’ flagship product is called Marble. It generates high-fidelity simulations of reality — photorealistic, physically accurate environments that can be used to train robots without exposing real hardware to the dangers of real-world trial and error. When a robot learns to pick up a fragile object inside Marble, it is building intuitions from thousands of simulated attempts rather than thousands of broken dishes.
The technical category World Labs occupies is called “world models” — AI systems that go beyond predicting the next word or the next pixel to constructing an internal representation of how space, physics, causality, and time interact. Li has argued since founding the company that this is the missing piece between today’s impressive but brittle AI systems and tomorrow’s genuinely autonomous ones.
“True artificial general intelligence requires grounding in physics and the ability to reason about data beyond text,” Li has said. The core insight behind World Labs is that language models, however sophisticated, are modeling the world through its shadow — words and descriptions — rather than the thing itself.
Why AMD Spent $8.2 Billion
At first glance, a semiconductor company buying an AI research lab looks like a strategic detour. AMD makes chips. World Labs trains models. Why not just sell more chips to whoever trains world models?
The answer, according to AMD CEO Dr. Lisa Su, is that understanding what frontier workloads actually need at the hardware level is becoming AMD’s most critical competitive capability. “Building compute platforms for next generation AI requires deep understanding of model evolution,” Su said in the official announcement.
Analyst Matt Kimball of Moor Insights & Strategy put it bluntly: “This isn’t about AMD buying a model company. This is about being able to understand the requirements of compute and memory” as AI evolves.
The physical AI transition is accelerating faster than most semiconductor roadmaps anticipated. World models and robot-training simulations impose radically different demands on hardware than transformer-based text generation does. By bringing Li and her team inside AMD, the company gains a live laboratory — a team building the exact workloads that will define chip requirements for the next five years.
AMD estimates the physical AI market could reach $200 billion by 2035. That projection matters not just for World Labs’ model sales, but for what it implies about the chips that market will consume.
The NVIDIA Problem AMD Is Trying to Solve
For the past three years, NVIDIA has systematically built what critics call an AI moat: CUDA software, the DGX hardware platform, and more recently a full physical AI stack including Cosmos (a world model), Isaac and Omniverse (robotics simulation and development platforms), and GR00T (a foundation model for humanoid robots). The result is an end-to-end ecosystem that is genuinely difficult to replicate.
AMD has made significant gains in AI chip market share through its MI300X accelerator and upcoming MI400 series. But the software ecosystem remains a weakness, and the physical AI layer — world models, simulation infrastructure, robotics training pipelines — is an area where AMD had essentially nothing.
World Labs changes that. When Li joins AMD’s leadership team as chief scientist, she is not arriving as an external partner or an affiliated researcher. She is becoming part of AMD’s strategic core, with a team that has spent two years working on the exact problems physical AI is trying to solve.
Brendan Burke of Futurum Group frames this as primarily a talent acquisition: “You can’t hire a more accomplished researcher than Dr. Fei-Fei Li, and access to talent is the biggest advantage in AI training.” That is not wrong — but the technology and the IP Li’s team developed at World Labs are themselves formidable assets.
The Stakes for Robotics
The timing of this acquisition is not coincidental. The robotics industry is in the middle of a massive inflection. Tesla’s Optimus and Cybercab programs are generating significant demand for synthetic training data. Figure AI, Physical Intelligence, Agility Robotics, and a dozen other companies are competing to build the first genuinely useful humanoid workforce. All of them need simulation infrastructure. All of them need world models that accurately represent how objects behave when manipulated, how floors feel underfoot at different speeds, how liquids respond to tilting.
Marble is built for this exact moment. And AMD, by owning World Labs, is now positioned to offer the hardware-software combination that these robotics companies will need at scale — not as a CUDA-dependent NVIDIA customer, but as an AMD-powered alternative.
For AMD, the physical AI layer is also a way to frame the chip competition differently. Selling MI400 accelerators purely on benchmark performance is a defensive game against a dominant NVIDIA. Selling MI400 accelerators as part of an integrated physical AI platform — with world models, simulation software, and the team that built them — is a different and potentially more compelling offer.
A Pioneer Returns to Hardware
There is something philosophically apt about Li’s destination. She built her career on the conviction that intelligence requires grounding — that you cannot understand the world without engaging it at the level of physical, visual, perceptual reality. Her creation of ImageNet in 2009 was motivated by that same conviction: that machine learning systems needed to confront the actual complexity and diversity of the real world, not just the controlled simplicity of benchmark datasets.
Now, in taking her spatial intelligence work inside a semiconductor company, she is completing a kind of loop. The most direct way to ground AI in physical reality may be to build the hardware alongside the models — to let the research inform the chips and the chips inform the research in a feedback loop that no pure-software company can replicate.
AMD and World Labs had already formed an inference optimization partnership in 2025. That collaboration reportedly gave both teams an early read on what the combined entity could achieve. The $8.2 billion acquisition formalizes what that partnership had already begun to suggest: that the future of physical AI runs better when the model builders and the chipmakers are the same team.
The AI industry’s next battle is being fought not in language, but in space, physics, and motion. AMD just paid $8.2 billion to enter it with one of the most credentialed combatants in the field.
