AMD에 Fei-Fei Li와 그녀의 팀이 합류하며, 주식 기준 82억 달러 규모의 인수가 46일 만에 완료되었다.
Now that's a true meeting of the titans. In an office in San Francisco, AMD CEO Lisa Su and World Labs co-founder Fei-Fei Li sat side by side on bar stools, facing a room full of applauding World Labs employees.
On November 16 (local time), AMD officially announced that it had completed its acquisition of World Labs, the spatial intelligence company Li co-founded. World Labs is now fully integrated into AMD.
In the announcement video, Su, wearing a black hoodie, gestured and chatted with a smile. Beside her, Li had already put on a black jacket with the AMD logo on the left chest. The white caption at the bottom of the video also changed AMD's long-standing slogan to "together we advance_worlds."
To acquire a team founded just two years earlier, AMD offered roughly $8.2 billion in an all-stock transaction.
Following the closing, Li formally became AMD's Executive Vice President and Chief Scientist, reporting directly to Su.
From the signing of the acquisition agreement on September 28 to the completion of all closing steps on November 13, the deal took just 46 days.
**14.98 Million Shares Issued, Second-Largest Acquisition in AMD History Closes**
According to a supplementary Form 8-K/A that AMD filed with the U.S. Securities and Exchange Commission on November 16, the acquisition closed on November 13.
The pace was considerably faster than the year-end timeline the two companies had originally projected.
The closing was carried out as an all-stock private placement. AMD issued 14.978 million common shares to World Labs' former shareholders, for a total valuation of approximately $8.2 billion (about RMB 54.6 billion).
In AMD's acquisition history, the deal ranks second only to its acquisition of Xilinx, which closed in February 2022.
In 2022, Su led AMD in acquiring the FPGA leader Xilinx in an all-stock deal valued at about $49 billion at closing.
That deal set a record for the chip industry and laid the foundation for AMD's embedded and adaptive computing business.
Four years later, Su has again committed $8.2 billion in stock, this time to acquire a team focused on 3D world models.
With the deal finalized, Li, World Labs' co-founder and CEO, has joined AMD's senior leadership as Executive Vice President and Chief Scientist, reporting directly to Su.
World Labs' other two co-founders, Justin Johnson and Ben Mildenhall, will remain with the company.
Within AMD, they will continue foundation model research as an independent frontier research group, working alongside AMD's hardware, software, and systems engineers.
Both are well-known figures in the computer vision community.
Johnson was a PhD student of Li's at Stanford. He helped build CS231n, the course that nearly every AI beginner has taken, and was a core contributor to real-time style transfer and the PyTorch3D library.
Mildenhall, from Google Research, is a co-inventor of NeRF, the 2020 paper that sparked the wave of 3D visual reconstruction research.
*[Left: Justin Johnson. Right: Ben Mildenhall.]*
In celebratory posts published by both companies, World Labs also quoted the line from Tennyson's *Ulysses*, "Come, my friends, 'tis not too late to seek a newer world," echoing the title of Li's earlier open letter.
Financial markets were noticeably more restrained than the warm tone of the companies. When trading opened on Monday, November 16, AMD's stock did not rise. Instead, it fell as much as about 5% amid a broader pullback in the semiconductor sector and AI infrastructure stocks.
Clearly, Wall Street is still waiting to see what first results this newly combined team will deliver in joint software and hardware development.
**From $230 Million to $8.2 Billion: What Did They Build?**
Li co-founded World Labs in 2024. Rather than follow the crowded field of large language models, the company focused on spatial intelligence, which aims to understand and simulate the three-dimensional physical world.
When World Labs emerged from stealth that year, it raised $230 million and was valued at over $1 billion.
In November of last year, World Labs released Marble, its first multimodal world model.
Users can enter text, images, or rough 3D layouts and generate 3D scenes that they can freely explore, edit, and export.
In February of this year, World Labs closed a new financing round of up to $1 billion. Its investors included Autodesk, which contributed $200 million, and Nvidia. AMD also participated.
However, the technology that most clearly demonstrated the company's capabilities was Atlas, a next-generation all-in-one world model released on September 1.
A few days ago, Li presented the technical foundations of Atlas in detail at the CoRL 2026 conference.
While the prevailing approach in large language models is "next token prediction," Atlas is built on the principle "Viewpoint Is All You Need." It predicts the next viewpoint through precise camera pose conditioning.
During training, Atlas simultaneously ingests text, images, video, depth signals, and 3D structural signals. Within a single autoregressive diffusion Transformer architecture, it can switch between three modes: dense reconstruction, sparse reconstruction, and pure text-to-generation.
In a blind test organized by the third-party firm RapidData, testers collected about 500,000 human evaluations comparing the camera-control generation capabilities of leading video and world models.
Atlas ranked first, capable of generating coherent multi-view videos longer than 60 seconds at 1440p resolution.
Perhaps more importantly, it is being applied to robotics.
This year, World Labs also acquired Scenix, a robotics company specializing in real-to-sim and sim-to-real pipelines.
At the event, Li demonstrated robots trained entirely in a virtual simulation environment built with Atlas, then deployed zero-shot into the real physical world.
On November 12, one day before the acquisition closed, World Labs announced that it would make Atlas fully available to the public and developers.
**Taking on Nvidia, AMD Decides to Build Its Own Models**
Just days before the announcement, Su and Li held their first joint public conversation at CoRL 2026.
At the start, Su said that physical AI, particularly robotics, would be a multi-trillion-dollar opportunity in the coming years.
She also revealed that AMD already has a sizable robotics research team, though its overall efforts are still at an early stage.
On stage, Li put forward the view that "AI is compute, and compute is AI."
This closely matches the emphasis in Su's closing announcement on the need for "joint optimization of computing platforms and AI models."
Asked earlier why World Labs did not start with Atlas, Li said half-jokingly, "Because Justin's team needs massive amounts of chips."
Building large-scale world models requires enormous GPU compute.
AMD and World Labs had been partners since 2025, jointly tuning model training and inference on AMD GPUs. However, external partnerships are constrained by commercial boundaries, which makes it difficult to address fundamental technical bottlenecks before chip designs are finalized.
Now that the two companies are one, Li's team can work directly with AMD on the architecture of future Instinct accelerators and the ROCm software stack. They can test real spatial intelligence workloads against hardware designs before those designs are finalized.
AMD has also been expanding its portfolio aggressively. In addition to the $8.2 billion World Labs acquisition, it announced plans to acquire the inference chip company Taalas, recently planned to take a stake in the quantum computing company PsiQuantum, and signed a data center compute leasing agreement with Core Scientific.
Meanwhile, its longtime rival Nvidia spent $12.9 billion in early September to acquire the open-source model platform Hugging Face. In robotics and physical AI, Nvidia has also introduced Cosmos world foundation models, Omniverse, and GR00T.
Facing the tightly integrated hardware-software ecosystem Nvidia has built, AMD has repeatedly emphasized "open AI ecosystems" from its September agreement through the November closing.
AMD aims to offer developers an alternative through open hardware, open software, and its own frontier models.
Competition among chip giants can no longer be decided by spec sheets for individual graphics cards alone.