AMD and World Labs: An $8.2 Billion Bet on Seeing the Future First

World model. Ask an ordinary image AI for a kitchen and you get a convincing picture: a kettle on the table, a fridge by the wall, sun through the window. Now move the camera behind the fridge. A door may appear that was not there, the table may change length, the chair in the corner may vanish. The model can draw a kitchen without knowing what a kitchen is. (This is an illustration, not a test of any product.)

A world model tries to fix that. It keeps track of where things are, what you should see when you turn, and keeps the whole space consistent. The longer-term aim is to know what happens when a cup is pushed, or how a robot should get round a table ⁸³⁷. The cases World Labs has shown in public are mainly in games, film, 3D design and early robot simulation ¹²¹¹³⁷; reliable use for training robots or simulating factories and roads at scale is still some way off ²⁷³⁷.

World Labs. A small American AI company, founded in 2024, that builds world models. Its product, Marble, turns a sentence, a photo or a video into a 3D world you can move around in ⁸¹².

Fei-Fei Li. A computer scientist at Stanford. In the late 2000s she led the building of ImageNet, a library of more than 14 million photos, each labelled by hand with what it shows ³⁴. It became a shared test for computers that recognise pictures. In 2012 a neural network called AlexNet won that test by a wide margin, and the technology industry took notice of deep learning, the approach behind today's AI ³⁴.

AMD announced the agreement after the US market closed on 28 September. The deal is all-stock and valued at about $8.2 billion; the number of AMD shares will be set by AMD's average share price over the ten trading days before closing ³⁸. AMD expects to complete it by the end of 2026, subject to regulatory approvals ¹²⁴. Fei-Fei Li will become AMD's executive vice president and chief scientist, reporting to chief executive Lisa Su ¹.

Justin Johnson and Ben Mildenhall, two of World Labs' co-founders, will keep leading its team together with Li, as a research organisation inside AMD ². The fourth co-founder, Christoph Lassner, is not named in the announced arrangements ²⁸. Neither company has said what will happen to Marble ³⁵.

World Labs came out of stealth on 13 September 2024 with $230 million from investors including Andreessen Horowitz, NEA and Nvidia's venture arm ⁸. Marble opened to everyone on 12 November 2025; it builds 3D worlds from text, images, video or rough layouts, and exports them with collision meshes a physics engine can work with ¹²³⁷. In February 2026 the company raised $1 billion from investors that included both AMD and Nvidia, as well as Autodesk ⁹.

World Labs itself sorts world models into three kinds: renderers, which produce pictures for people; simulators, which keep a geometrically and physically faithful state of a scene; and planners, which decide what a robot should do next. It notes that robot demonstrations have so far been "confined to heavily constrained laboratory setups" ³⁷. World Labs places Marble between renderer and simulator ³⁷; in our view it is still closer to a very fast set builder than to a trusted set of physical laws.

Then there is Li. She began ImageNet at Princeton in 2007 and presented it in 2009 ³²¹⁵; Stanford's Institute for Human-Centered AI, which she co-founded and co-directs, describes her as its inventor ¹⁴. She was chief scientist of AI at Google Cloud in 2017 and 2018 ¹⁵. Some outlets call her the "godmother of AI" ³⁰; she has said she would never call herself that ³².

The deal pairs two paths into AI. Lisa Su, an engineer, became AMD's chief executive in October 2014 ⁴², in a year it reported a net loss of $403 million ⁴⁰; in September 2026 its market value briefly passed $1 trillion ⁴¹. Hers is an approach from the chip up. Li's ImageNet changed how machines learn to see: an approach from the model down. The acquisition tries to make the two meet in the middle. But generating a beautiful 3D space and helping AMD sell more chips are not the same thing.

Chip companies usually learn what their chips lack from customers: memory runs out, the links between GPUs are too slow, the software is missing a tool. By the time the complaint arrives, the next chip may already be largely fixed. AMD wants that feedback earlier. "Building the compute platforms for the next generation of AI requires a deep understanding of how models are evolving," Su said ¹.

A team building new models hits a chip's limits first: memory size and bandwidth, the links between GPUs, latency, power use and missing software. Inside AMD, World Labs could act as an early, demanding in-house customer for its Instinct accelerators and ROCm software; AMD says the team will help it understand new AI workloads and shape its hardware, software and systems ¹. Li put it from the other side: her team needs to get "closer to the hardware", because without a focused hardware effort "AI is hobbled in efficiency. And scale." ³

The two companies already worked together: they formed a partnership on inference optimisation and training in 2025 ⁴. AMD also says it is not trying to compete with the AI companies that buy its chips; Vamsi Boppana, its senior vice president for AI, said that is "not the intent at all" ⁷. The first test has a twist: RTFM was shown running on an Nvidia H100 ¹³. Whether World Labs' models run as well, or better, on AMD's own accelerators will be one of the earliest signs of whether the bet works.

World Labs follows a run of purchases that moved AMD from chips towards complete AI systems: Xilinx, completed in February 2022 at about $49 billion ¹⁶; the networking-chip designer Pensando for about $1.9 billion in 2022 ¹⁷; the AI lab Silo AI for about $665 million in 2024 ¹⁸; and the server builder ZT Systems for $4.9 billion, completed in March 2025 ³⁶¹⁹. Several smaller AI software teams joined at undisclosed prices, the latest being MK1 in November 2025 ³³. Silo AI, which built the open Poro and Viking language models, was AMD's first model lab ¹⁸; World Labs costs more than ten times as much, and is the first built around world models rather than language models.

Three numbers sit side by side. About $1 billion: World Labs' valuation when it emerged in 2024 ⁵¹¹. About $5 billion: the valuation press reports gave it in February 2026, which the company did not confirm ¹⁰¹¹. About $8.2 billion: AMD's price, some 64% above that reported figure. World Labs has not disclosed revenue, losses or customer numbers, so from outside the price can only be explained by the team, technology, intellectual property and strategic value ³⁵. How many new AMD shares are issued depends on AMD's share price before closing; one estimate put them at about 0.8% of the company ³⁸²⁴.

AMD's market value was about $992 billion at the close on 28 September ²⁰. Its shares were flat to slightly lower after hours and about 1.4% higher before the market opened the next day ²⁴⁷²⁵. Analysts differed. Citi said the deal could give AMD "deeper visibility" into how AI models are evolving; RBC said Nvidia keeps a "significant" lead in spatial AI and physical simulation ²².

World models are also a focus for AMD's rivals. Google DeepMind showed Genie 3, a real-time interactive world model, in August 2025 ²⁶. Nvidia launched its Cosmos world foundation models in January 2025 to generate training data for robots and self-driving cars ²⁷. The broker Stifel said World Labs fills a gap for AMD, which had released only text and video open models ²³.

Large technology companies have also paid heavily for AI teams. Meta invested $14.3 billion in Scale AI in 2025 and hired its chief executive ²⁹. In December 2025 Nvidia agreed a licensing deal with the chip start-up Groq, reported at $20 billion, with Groq's founder joining Nvidia ²⁸.

AMD is not buying a market that already exists. It is buying a chance to help define one. These are the signs that will show whether that worked.

Does it run on AMD? RTFM was shown on an Nvidia H100 ¹³. The measurable test is whether World Labs' models run faster or cheaper on AMD Instinct, and whether that draws outside developers to ROCm.

Was a model lab the gap that mattered most? An independent benchmark by SemiAnalysis in December 2024 found AMD had not yet crossed Nvidia's CUDA "moat", pointing to a "challenging out of the box experience" ³⁹. AMD could instead have put comparable capital and management attention into buying or building software, compilers and developer support; buying World Labs puts a bigger stake on models working hand in hand with hardware.

Lab or product? Neither company has said what happens to Marble ³⁵. World Labs could turn from a start-up serving many kinds of hardware into a lab that mainly shows off AMD's.

Will customers still share their plans? AMD says it will not compete with the AI companies that buy its chips ⁷, but it will now own a model team of its own.

Will the people stay, and will the deal close? With little public financial data, the valuation appears to rest largely on the team, technology, intellectual property and strategic potential ³⁵. The deal still needs regulatory approval, expected before the end of 2026 ¹.

Today World Labs shows the future it imagines on a single Nvidia H100. If, two years from now, those worlds run faster and cheaper on AMD Instinct and bring outside developers into ROCm, AMD will have turned its pair of eyes on the future into chip orders. If not, $8.2 billion may have bought only a front-row seat.

We looked at the articles in our sources about the deal between 28 and 30 September 2026: 37 articles from six countries. Chinese outlets carried more than half of them. This shows only what our sources carried.