AMD Buys a Ticket to the 3D World: $8.2 Billion for World Labs
$8.2 billion—about 55 billion RMB, all stock. That is the price AMD put on the table on September 28, 2026, for World Labs, the company Fei-Fei Li founded in April 2024. The deal is expected to close before the end of 2026, subject to regulatory review.
To calibrate the scale: this is AMD's second-largest acquisition ever, behind only Xilinx (~$49 billion, closed in 2022). It represents roughly 0.8% of AMD's near-$1 trillion market cap, and comes in about 64% above the $5 billion valuation floated during January negotiations.
Where the Two Key People Land
After closing, Fei-Fei Li joins AMD as Executive Vice President and Chief Scientist, reporting directly to Chair and CEO Lisa Su. World Labs' other core founding members will continue leading world model research with their teams.
Su's framing: building a computing platform for the next generation of AI requires deep understanding of how models evolve. Li's framing: joining AMD gives her team the resources and engineering depth to accelerate research and help define the infrastructure the next AI era needs.
Assessment: the motive behind a chipmaker buying a model lab comes down to information asymmetry. AMD's biggest customers are the AI labs deciding which models to train next—and those customers won't tell you what their models need before your hardware roadmap is finalized. Running your own lab internalizes that intelligence.
What World Labs Actually Built
A world model is essentially a flight simulator for AI: the model learns how a space is laid out, so a user or robot can move through it from any angle and ask "what happens if I do this?"
In its first two years, the company shipped three generations of products:
| Time | Product | What it does | |---|---|---| | 2024 | LWM (Large World Model) | Generates interactive 3D worlds from a single image | | Later | Marble | Builds 3D worlds from text, images, video, or 3D layouts | | Sept 2026 | Atlas | Touted as the first multimodal world model: pixel-level camera control, 3D reconstruction, video re-shooting, and replicating real scenes as robot training environments |
The funding ladder is equally clear. In September 2024, a $230 million round valued the company at about $1 billion; in February 2026, a $1 billion round reportedly valued it at roughly $5 billion post-money. Shareholders include a16z and Radical Ventures—as well as Nvidia, AMD, and Autodesk.
Li's own condensed thesis: in a November 2025 essay she called language models "wordsmiths in the dark"—fluent, but cut off from the physical world. That line is the shortest path to understanding what World Labs is doing.
Earlier this year she appeared on stage with Lisa Su to demo Marble, generating a 3D scene from a handful of images. Her pitch: agents—robots, vehicles, tools—can learn first in physically grounded digital worlds before deploying to the real one, which is far safer.
The Competition Is Already There
Nvidia already sells open-weight world models—Cosmos among them—alongside its chips. Until now, AMD only had text and video models. This acquisition fills exactly that gap.
The two companies had prior ties: in 2025 they collaborated on training and inference optimization of World Labs' models on AMD GPUs. So this isn't a first handshake—it's a partnership becoming an internal division.
Other names on the same bill: CNBC's report listed parallel deals—OpenAI bought Jony Ive's AI device startup io for ~$6.4 billion last year; Nvidia spent ~$33 billion combined on Groq and Hugging Face; Meta paid $14 billion in 2025 for a minority stake in Scale AI and hired founder Alexandr Wang to build a new AI unit.
Assessment: what these deals buy is people and time, not revenue. The $8.2 billion price attaches to a lab only two and a half years old with a product line still iterating rapidly.
The Bill Isn't Settled Yet
The announcement places the payoff on a roadmap—no product dates, and consumer lines like Ryzen and Radeon are out of scope. World Labs operates independently until closing.
Assessment: the real open question is the last link in the chain: whether spatial model research can change which chips data centers buy. Su's announcement talks about roadmaps, and there are years between a roadmap and a purchase order.
Same-day context: Su is also scheduled to speak at OCP 2026 on open full-stack hardware for agentic workloads. Same direction—but so far that's an agenda item, not a product.
The concrete checkpoint: wait for the deal to close at end of 2026, then wait for the first Atlas results running on AMD hardware. Only then does the $8.2 billion get its first verifiable number.
Sources: AMD announcement and CNBC reporting, 2026-09-28; Securities Times and China Securities Journal on the deal amount; World Labs product pages.
Caveats: the $8.2 billion figure is an all-stock consideration that will fluctuate with AMD's share price; the $5 billion valuation, 64% premium, and 0.8% market-cap share come from media reports without official AMD confirmation; other acquisition figures are cited from the same CNBC report.