An all-stock deal worth about US$8.2 billion
On 28 September 2026 AMD posted a statement on its website saying it had signed a definitive agreement to acquire World Labs. The transaction is an all-stock deal valued at roughly US$8.2 billion (8.2 billion US dollars), and is expected to close before the end of 2026, subject to regulatory approval and customary closing conditions.
World Labs was founded in 2024 and works on spatial intelligence, the ability for AI to understand the three-dimensional world, navigate it and act inside it. The company has about 70 people. Late last year it released its first product, Marble, a foundation model that generates 3D worlds from an image or a text prompt. Earlier this year it closed a US$1 billion Series C round.
After the acquisition, World Labs founder Fei-Fei Li joins AMD as executive vice president and chief scientist, reporting directly to CEO Lisa Su.
Two women, two very different journeys
Li was born in Beijing in 1976 and grew up in Chengdu before moving to New Jersey with her parents at 16. She earned a physics bachelor’s degree from Princeton and a PhD in electrical engineering from Caltech, then taught for years at Stanford, where she is the Sequoia chair professor. In 2009 she created ImageNet, the dataset widely credited with igniting the deep-learning revolution. She briefly served as chief scientist of AI and machine learning at Google Cloud in 2017, returned to Stanford, and co-founded the Institute for Human-Centered AI (HAI). In 2024 she started World Labs around the idea of spatial intelligence: if large language models taught machines to read and write, the next step is letting them understand the physical world.
Lisa Su was born in Tainan in 1969 and moved to the United States at three. She earned her bachelor’s, master’s and doctorate at MIT, all in electrical engineering, working on silicon-on-insulator technology. After stints at Texas Instruments and IBM she joined AMD in 2012 and became CEO in October 2014. At the time AMD’s market value was a little over US$3 billion, it was posting losses and the shares hovered around two or three dollars, widely written off as a dying also-ran. Her bets on the Zen architecture, on turning TSMC’s leading process into a weapon, and on cutting non-core businesses turned AMD within a decade into a trillion-dollar challenger to Nvidia.
Su’s plain explanation of the deal: the more you understand the end-to-end flow, the better the system you can build. She said AMD bought World Labs to get the world-class talent Li had assembled, and to connect it to AMD’s hardware, software and systems. Li was more direct: the company’s ambition is growing, and so is its hunger for compute. By teaming up with AMD, she said, they create a chance for themselves, for AMD and for the whole ecosystem to speed up the virtuous loop between software and hardware development.
Not an impulse: two years of groundwork
This was not a sudden move. AMD had been an investor in World Labs for two years before the acquisition. Across the three and a half years from January 2023 to 29 September 2026, AMD made 37 investment or acquisition moves touching 33 companies. The pace quickened every year: 4 deals in 2023, 9 in 2024, 9 in 2025, and 15 in the first three quarters of 2026 alone.
By sector the concentration is striking. Artificial intelligence accounted for 28 of the 37 deals, or 76 per cent. The rest spread across advanced manufacturing, enterprise software, automotive, healthcare and frontier tech. In stage terms AMD favoured later rounds, but its early bets were enormous: a US$650 million seed round into Recursive Superintelligence, a US$1.1 billion Series A into River AI, a US$700 million Series A into Hark. These are not small options bets; they look like buying the most important boarding pass in the narrowest window.
Geographically, 23 of the 33 companies are in the United States. In Europe, AMD backed ARM, Stability AI and Wayve in Britain and Teramount and Classiq in Israel. In Asia it touched Moreh and Upstage in South Korea, Turing in Japan, Video Rebirth in Singapore and Uniphore in India. The individual tickets in Korea, Japan, Singapore and India are modest, mostly US$20 million to US$80 million, but together they map a clear intent: those countries are all pushing their own sovereign AI, and compute autonomy commercially means wanting an alternative to Nvidia. AMD is buying the narrative behind them.
What the money actually bought
At the model layer AMD spread its bets deliberately. It backed Cohere twice at US$500 million each, Hugging Face’s US$235 million Series D, Scale AI’s US$1 billion Series F, Liquid AI’s US$250 million Series A and Stability AI’s US$76 million Series B. In inference software it funded RadixArk, builder of SGLang, one of the most important open-source inference engines, with a US$100 million seed, plus Fireworks AI and Wafer. In hardware it reinvested twice in Ayar Labs, an optical I/O maker, at US$155 million and US$150 million, backed photonic-compute firm Celestial AI, joined ARM’s US$735 million IPO cornerstone round and funded Teramount. And across world models, AMD missed none of them: World Labs, Luma AI, Runway, Odyssey, Video Rebirth and Wayve, eight deals in total.
Seventeen times at the same table as Nvidia
Unpacking the co-investor lists reveals something counter-intuitive: Nvidia appears in 17 of AMD’s 37 rounds, across 14 companies. In other words, close to half of AMD’s moves put it at the same table as its most direct rival. The earliest was Hugging Face’s 2023 Series D; ARM’s IPO cornerstone round read like a group photo of the semiconductor industry; Ayar Labs had AMD, Nvidia and Intel all in both its D and E rounds. Cohere, RadixArk and Hark each had both chip giants on the cap table. And in World Labs, both AMD and Nvidia were in the B and C rounds from the start, until AMD took the whole pot for US$8.2 billion.
The reason is simple. A corporate venture stake is rarely exclusive, and neither side expects a few hundred million dollars to lock a company in. What both want is the same thing: information. Knowing what these model companies are building, what they train on, where the bottleneck sits and where they go next is worth as much to Nvidia as to AMD, perhaps more to AMD as the chaser who cannot afford an information gap. Nvidia plays defence, protecting its CUDA ecosystem; AMD plays offence, making sure it can at least see.
Why spend like this
Nvidia’s moat was never the GPU; it was CUDA, two decades of software, operator libraries and developer habits that a latecomer cannot storm head-on. AMD has the chips but an empty ecosystem, and could not fill in the operators one by one in twenty years. So it invested instead, trading capital for time and equity for knowledge. With 37 deals it built a chain: at the bottom, silicon and interconnect; above that, compute and scheduling; then models; then distribution; then shipping channels. Every layer has an anchor, and often more than one. It is not chasing a hit rate but coverage.
The cleverest part of the play is that the investing itself is merger due diligence. The biggest risk in a traditional acquisition is information asymmetry. If you first take a minority seat and watch for a year or two, that risk shrinks. World Labs is the first full loop: AMD entered at the B round eight months after the company was founded, followed it to C, observed it for two years, then bought it outright. This was not spotting and pouncing; it was watching enough and then pouncing.
There is also a demand angle. The companies AMD funds need compute, and many are already AMD GPU customers or prospects. Li herself made the point: World Labs’ ambition is growing and so is its compute need, and the deal joins the two bottlenecks, a model company constrained by compute and a chip company that does not know how much future compute models will demand.
What to watch next
Three and a half years, 37 investments plus one acquisition, about 136.8 billion yuan (roughly US$19 billion) deployed. The question is no longer who AMD will back, but which of the 33 it has already backed becomes the next wholly owned target. The profiles that fit: reinvested two or more times, sitting at a key node, not yet locked by a giant. Ayar Labs took two rounds, Luma AI two Series C rounds in a year, Cohere two Series D rounds, and all three are on the list. If AMD’s thesis is that world models and physical AI are the next compute surge, it will probably buy again, and likely something it has already watched for a year or two.
In September 2024 AMD first appeared on World Labs’ cap table, in a round that raised only US$230 million and barely registered in the news. Two years later it bought the company for US$8.2 billion.
Editor’s note: This is an adapted translation of the original 36Kr report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://36kr.com/p/4004155204178053.