AMD Acquires World Labs for $8.2 Billion; Fei-Fei Li to Serve as AMD Executive Vice President and Chief Scientist

$8.2 billion has already provided a value anchor for the entire world model track.

Author: Xing Lijuan

Editor: Cen Feng. $8.2 billion, an all-stock acquisition of World Labs, founded by Fei-Fei Li. This is AMD's second-largest acquisition in history, second only to the roughly $50 billion deal for Xilinx in 2022. Xilinx is a veteran chip company of more than 30 years, while World Labs has been around for only two years. After the news broke, the first reaction among industry insiders was unanimous: Lisa Su is benchmarking against Nvidia's Cosmos and robotics platforms—buying people, buying model insight, buying a ticket into the era of physical AI. But Su is not the only one placing a bet in this deal. Fei-Fei Li is betting too. And what she is staking is her own next twenty years. Over the past two decades, she has repeatedly done the same thing: at every technological inflection point, define one step ahead "what is the next thing worth doing." But the first two times, she was only responsible for setting the questions, leaving the work of answering to others. This time, she has seated herself at the answering desk.

01

Three times "digging pits": what's different? Fei-Fei Li, hailed as the "godmother of AI," is perhaps best known in the industry for being best at "digging pits" (setting questions). Over the past twenty years, she has dug three of the most critical pits: ImageNet, AI democratization, and spatial intelligence. Looked at individually, each was a precise judgment about the direction of the AI industry. But placed side by side, the role she played was completely different.

▎The first "pit": purely setting the question, with the entire deep learning community answering Around 2007, Fei-Fei Li was an assistant professor at Princeton. The most critical problem in computer vision at the time was generalization—algorithms worked in the lab but failed outside it. Most people chose to improve algorithms; Fei-Fei Li chose another path: building data. At the time, almost no one in academia was optimistic about this choice. Senior professors at Princeton advised her not to do it, saying it was "too ambitious"; her applications for federal funding were repeatedly rejected, and some even criticized the project as "the shame of Princeton." Not only that, she also gambled her academic future. If ImageNet failed, she would likely not get tenure. This dataset did not solve any algorithmic problem. It only raised one question: by what standard do we measure visual intelligence? That question hit the nail on the head at the time. Without enough data, all algorithmic innovation is "building towers on sand." When it was released in 2009, it barely caused a ripple in academia. Because ImageNet was a "pure question-setting" kind of pit—someone had to be willing to fill it. It was not until 2012, when Hinton and Krizhevsky's team swept the ImageNet challenge with AlexNet, that the deep learning revolution was ignited. The spark that brought AI into the era of deep learning revival was AlexNet. Fei-Fei Li successfully defined the underlying infrastructure of the AI industry. But she was only the "question setter"; no matter how glorious ImageNet was, the people who filled the pit were Hinton, who won the Turing Award, Krizhevsky, who captured the first wave of dividends, and the entire deep learning community.

▎The second "pit": AI democratization—she wanted to answer, but the exam paper was not hers At the end of 2016, Fei-Fei Li joined Google on an academic sabbatical as chief scientist of Google Cloud. This time the pit she dug was called "AI democratization." Make AI no longer the exclusive patent of big companies, but a tool that small and medium-sized enterprises, traditional industries, and academic researchers can all use. In March 2017, she announced the acquisition of Kaggle at the Google Cloud Next conference; that same year AutoML was brought to market, Contact Center AI entered alpha testing, and the Google AI China Center was established in Beijing. The product line was rolling out, and the direction was not wrong. But the way of filling the pit went wrong. Because filling the pit required three kinds of power: decision-making power over the product commercialization path, negotiating power over the direction of customer contracts, and control over internal public opinion. But Fei-Fei Li's position at Google Cloud was chief scientist; she could influence research direction, but she could not influence how this commercial machine operated. What truly collapsed this pit was "Project Maven." After it was exposed that Google was providing technical services for the project, internal protests quickly escalated. And the line Fei-Fei Li emphasized in an internal discussion email—"avoid any mention of or allusion to AI at all costs"—was interpreted by employees as covering up for the company's cooperation with the military, and she was pushed to the center of public controversy. A person who joined Google Cloud under the banner of "AI democratization" was ultimately pushed away from that position by a storm of public opinion over AI ethics. During her more than 600 days at Google Cloud, she set the question of "AI democratization" and wanted to answer it, but the question setter, the invigilator, and the paper collector were not her. A scientist's professional influence cannot replace commercial and underlying decision-making power. Defining a direction and making that direction truly land have always been two completely different things.

▎The third "pit": putting herself in the answer seat

In April 2024, Li Feifei joined forces with Stanford vision researcher Justin Johnson, NeRF paper first author Ben Mildenhall, and NeRF follow-up work researcher Christoph Lassner to found World Labs, formally diving into entrepreneurship. This time, the proposition was a sentence she had repeated in countless public talks: the universe is not made of words, but of real things. No matter how powerful LLMs become, they still live on a two-dimensional screen: they do not understand occlusion, do not understand gravity, and do not understand that a cup falling off a table will shatter. "Spatial intelligence" is the next frontier of AI. This was her third "pit-digging," but unlike the previous two: she was no longer just the one setting the question. Before formally answering, someone first had to have ammunition. The World Labs founding team itself was a trump card. Justin Johnson is a researcher at the Stanford Vision and Learning Lab, with long-term accumulation in image captioning, visual reasoning, and generative models; Ben Mildenhall is the first author of the NeRF paper, which redefined the technical route for 3D scene reconstruction; Christoph Lassner is one of the core researchers of NeRF follow-up work. The four formed a complete technical lineup from 3D reconstruction to generative models. The academic advisory team was equally heavyweight: Wu Jiajun, an assistant professor at Stanford and a Tsinghua Yao Class alumnus, whose research direction is physical scene understanding; and Wu Yiren, a professor at UC Berkeley and Malaysian Chinese, who studies 3D vision. On the ammunition side, the company completed a seed round in the month it was founded, led by a16z and Radical Ventures, at a valuation of about $200 million; on February 18, 2026, it completed a $1 billion Series C round led by Autodesk, with participation from AMD, Nvidia, Emerson Collective, Fidelity, and others, at a post-money valuation of about $5.4 billion. The valuation went from $200 million to $5.4 billion in less than two years. The money raised also had a clear purpose: the training cost of spatial intelligence is far higher than that of text models, requiring large-scale compute, long training times, and parallel advancement of multiple product lines. Without enough ammunition, this pit cannot be filled. While stockpiling ammunition, the product line was also advancing.

In December 2024, the first model LWM (Large World Model) was released, capable of generating an interactive 3D world from a single image.

In November 2025, the first commercial product Marble launched, supporting the generation of explorable, editable 3D scenes from text, images, video, and 3D layouts.

In July 2026, it acquired the robotics simulation company SceniX, connecting world models into the robotics reinforcement learning loop.

On September 2, 2026, the multimodal world model Atlas was released. This multimodal world model supports pixel-level camera control, sparse-view 3D reconstruction, and robot pretraining. Every step was about turning the "world model" from a demo into an engineered product. Every step had a different form of compute demand: from real-time inference on a single H100, to large-scale training of multimodal autoregressive diffusion Transformers, to the hard requirements of robotics simulation for memory bandwidth, parallel computing, and low-latency inference. Funding was increasing and products were advancing, yet one question was becoming increasingly clear: the further world model R&D progressed, the less it looked like a purely software problem. Li Feifei has acknowledged on multiple occasions the special compute requirements of spatial intelligence: models need to understand 3D structure and motion, understand physical laws, and require very large memory, massive parallel computing, and fast inference. Atlas reconstructs 3D scenes from 1 to dozens of photos and generates videos up to 1 minute long at 1440p resolution. These tasks involve a large amount of geometric computation and sparse reconstruction, and general-purpose GPU architectures are not ideal for such tasks. Li Feifei chose to move closer to AMD. In January 2026, Su Zifeng invited Li Feifei onto the CES keynote stage. Li Feifei demonstrated on the spot that Marble, after optimization on AMD MI325X, was more than 4 times faster, and revealed that she "can't wait to see it continue to improve on MI450." However, when the model team and the chip team sit at the same card table and "converse" with real training workloads, a structural contradiction is exposed: World Labs is still on the side of "making demands," and AMD is on the side of "meeting demands." Between the demand side and the supply side, there are still layers of bargaining, prioritization, and cross-company collaboration. Once demand is diluted by bargaining, what the model team wants most may not be what the chip team prioritizes most. At this point, Li Feifei should already have realized that to truly land the answer to the spatial intelligence question, this issue would have to be solved sooner or later. The key is, how to solve it?

02

A higher-dimensional solution: cutting into hardware discourse power. On September 28, 2026, AMD announced the acquisition of World Labs for approximately $8.2 billion in an all-stock deal. After the transaction closes, Li Feifei will become AMD's Executive Vice President and Chief Scientist, reporting directly to Su Zifeng. On the surface, this acquisition is AMD catching up on physical AI. But if you only see this level, you underestimate Li Feifei's own strategic planning. In this deal, World Labs' 70-person team is not being swallowed up, but is being incorporated into AMD as a relatively independent research unit. Li Feifei herself reports directly to Su Zifeng, while Justin Johnson and Ben Mildenhall continue to lead model R&D. She is not going to be a vice president whose role is diluted by the reporting chain, but rather is bringing the model team to directly connect with the chip roadmap. In her open letter "To Seek a Newer World," Li Feifei mentioned the most critical reason behind this decision: "We need to expand the scale and scope of our research and integrate more closely with hardware. Without dedicated investment in hardware R&D, both the efficiency and scale of AI will be constrained. For the goals we want to achieve, it will also remain trapped in the digital world." Su Zifeng also explained this very clearly: "Building a computing platform for the next generation of AI requires a deep understanding of how models will evolve. Li Feifei and the World Labs team bring research leadership and model expertise, and by working together, we can use these insights to develop the hardware, software, and systems that will drive the next generation of AI." Translated, it comes down to one thing: from this day on, World Labs' models are no longer just customers making requests to AMD; they are now sitting on the side that writes the requirements. This is not just narrative rhetoric, but Li Feifei's entry point into hardware discourse power. In the past, she defined "what problems are important." This time, what she wants to define is "what the tools for solving this problem look like." And "getting closer to hardware" is a very key answer she has given. But getting closer to hardware is not the endpoint. The next step to consider is how to verify whether this path works.

03

This time, the evaluation standard is no longer citation counts. This time, the evaluation standard is different from any previous academic achievement. The first is the technical closed loop. The AMD MI450 series is based on the CDNA 5 architecture, with FP4 compute reaching 40 PFLOP and FP8 compute reaching 20 PFLOP, double that of the MI350 series; memory has been upgraded from 288GB HBM3e to 432GB HBM4, with bandwidth reaching 19.6 TB/s, 2.45 times higher than the previous generation. MI450 will begin mass production and delivery in the second half of 2026. If the performance improvements of Atlas and Marble on MI450 continue the 4x speedup curve seen on MI325X, that will show that the path of "model requirements reverse-defining hardware architecture" works. Conversely, if MI450 is merely an ordinary process upgrade and has not made architectural adjustments for sparse computing, multi-view geometry, and video generation under camera control conditions for spatial intelligence, then the logic of "models reverse-defining hardware" has only been completed halfway: the model team has entered the door, but has not obtained chips tailor-made for it. The second is ecosystem traction. If World Labs' models only run well on AMD's own hardware, that is just a custom optimization. True success would be getting other teams working on spatial intelligence and embodied intelligence to also actively choose AMD hardware. As Li Feifei emphasized in the official announcement, the goal is to "continue building an open AI system covering hardware, software, models, and data platforms." The word "open" deserves attention. NVIDIA's dominance in the AI compute market largely comes from its deep ties with major AI labs around the world—anticipating the characteristics of next-generation models in advance from real training workloads. This time, AMD chose to directly incorporate a top model team into its own system. But World Labs also promises to maintain a strategy of open models and cross-platform deployment. This means AMD must prove that after incorporating World Labs into its system, it is not narrowing the ecosystem but expanding it. If other teams turn to AMD because World Labs' models run better on AMD, then her choice will truly have shifted the competitive landscape. If everyone just watches a launch event and then goes back to buying NVIDIA, then the $8.2 billion buys only a ticket to entry, not an ecological niche. Beyond the two objective standards, there is an even more fundamental question that needs answering: how far can Li Feifei, who has chosen to join forces with AMD, actually take spatial intelligence? The success of ImageNet twenty years ago did not lie in how perfect the dataset itself was, but in making the entire field realize that "data-driven" was the direction, and in pushing the whole community to keep trying, competing, and breaking through around this problem, ultimately making deep learning the consensus of the AI industry. Now, Li Feifei herself has taken the answer seat. What she wants to build is an open ecosystem from hardware to models, from platforms to developers. If this can truly be implemented, turning spatial intelligence from a research direction of a few teams into a technology agenda that the entire industry takes seriously, then this pit will truly have been filled.

04

Closing remarks. And before all of this is verified, the $8.2 billion AMD has offered has already provided a value anchor for the entire world model track. Song Kuan, founder of Zhitianxia, who went all in on 3DGS back in 2023, judges it this way: AMD's acquisition has, for the first time, shown China's compute companies, cloud computing vendors, and even the broader internet giants just how large the future computing market for world models and spatial intelligence can be. This track now has a quantitative reference for evaluation. The price itself also tells the story. World Labs was valued at $5.4 billion after its Series C funding round in February 2026. AMD's all-stock acquisition at $8.2 billion is about 52% higher than the previous round's valuation. Su Zifeng herself was an angel investor in World Labs, and this time she continued to increase her stake, and at a premium. This shows that the market's judgment of World Labs' value is, on the whole, still moving upward. $8.2 billion is a number that will be cited by every team working on world models, and referenced by every investor looking at this track. And looking back on Li Feifei's twenty years, what she has truly cared about has always been defining "what is the next thing worth doing." Twenty years ago, in an office at Princeton, facing colleagues' doubts and rejected funding, Li Feifei decided to dig a pit called ImageNet, and then let the entire deep learning community answer the question. Ten years ago, she walked into Google Cloud, wanting to dig a second pit called "AI democratization." She set the question and also wanted to answer it, but the person who wrote the exam paper was not her, the invigilator was not her, and a public opinion storm over Project Maven pushed her out of the exam room. Leiphone Leiphone Leiphone Twenty years later, she has dug another pit called "spatial intelligence." This time, she is the one setting the question, and she is also the one answering it. And the criteria for judgment next will no longer be just paper citation counts, but the performance curve of the soon-to-be-delivered MI450, the speed of developer adoption, and whether "spatial intelligence" can, like "deep learning," transform from a scholar's judgment into a consensus of an era.

To make it easier for everyone to connect and share conference news, we have set up a dedicated IROS 2026 attendee exchange group! Join the group and you will unlock these "benefits"?

Conference intelligence station: submission DDLs, format specifications, review progress... key milestones delivered as soon as they arrive, so you never miss a deadline again and can prepare your submission with confidence.

Peer tea chat: peers from all over are in the group, sharing insights, bouncing around ideas, and building connections—we are on the research journey together.

Frontier supply station: the latest research results and hot directions synced in real time, helping you clarify your submission thread in line with the conference theme and fueling your paper inspiration.

On-site support team: (exclusively opened during the conference): no need to panic once you arrive at the venue—

Topic co-reading: hot sessions and Keynote discussions, so you can catch up even if you miss them;

Poster compilation: highlights from on-site posters organized for you, one-click save so nothing is missed;

Meetup plaza: dinner meetups, interview meetups, exchange meetups... whether you want to chat, collaborate, or meet in person, just start it in the group.

? Group entry portal: scan the QR code to join or add WeChat Luyoyo_2026, with the note: ECCV + unit/school + name + direction.