AI news

AMD announces World Labs acquisition closing; Fei-Fei Li becomes EVP and chief scientist

By 克雷西3 viewsSource: 量子位
Lisa Su and Fei-Fei Li at the World Labs acquisition announcement

AMD announced completion of its World Labs acquisition on November 16, with Fei-Fei Li reporting directly to Lisa Su. Regulatory filings place closing on November 13 and record approximately 14.98 million common shares issued to former shareholders.

Editorial correction (November 17, 2026): AMD’s announcement and 8-K/A distinguish signing, announcement and closing dates. The agreement was signed September 26 and closed November 13, not signed September 28 and closed forty-six days later. Verification source: https://ir.amd.com/financial-information/sec-filings/content/0001193125-26-462340/d105780d8ka.htm .

Two leading figures have truly come together!

In a San Francisco office, “semiconductor queen” Lisa Su and “AI godmother” Fei-Fei Li sat side by side on stools, before a room of applauding World Labs employees.

On November 16 local time, AMD formally announced completion of its World Labs acquisition, bringing Li’s spatial-intelligence startup fully into AMD.

Article illustration

In the announcement video, Su gestures and smiles in a black hoodie. Beside her, Li wears a black jacket bearing AMD’s logo.

White subtitles beneath the celebration adapt AMD’s longstanding slogan to "together we advance_worlds".

Article illustration

AMD offered approximately $8.2 billion in stock for the two-year-old team.

Following closing, Li became executive vice president and chief scientist, reporting directly to Su.

The 8-K/A states signing on September 26, the original 8-K filing on September 28 and closing November 13. Signing and disclosure are different dates; September 28 cannot be used as the signing date to calculate elapsed time.

14.98 million shares: AMD’s second-largest acquisition closes

AMD’s November 16 Form 8-K/A filing with the SEC confirms completion on November 13.

That was earlier than the originally expected year-end deadline.

The transaction used a private all-stock issuance. AMD issued 14.978 million common shares to former World Labs shareholders, valued around $8.2 billion (approximately 54.6 billion yuan).

Article illustration

In AMD’s acquisition history, the deal ranks second only to Xilinx, completed in February 2022.

Su led that all-stock acquisition of the FPGA leader, worth approximately $49 billion at closing.

It set a semiconductor-acquisition record and underpinned AMD’s subsequent embedded and adaptive-computing portfolio.

Article illustration

Four years later, Su again offers stock, this time $8.2 billion for an AI team specializing in 3D world models.

After closing, World Labs cofounder and CEO Fei-Fei Li joins AMD’s leadership as executive vice president and chief scientist, reporting to Su.

World Labs’ other two cofounders, Justin Johnson and Ben Mildenhall, also remain.

They will continue to advance foundation-model R&D as an independent frontier-research organization within AMD, while also collaborating with AMD's hardware, software and systems engineers.

Both are familiar figures with extensive computer-vision credentials.

Johnson earned his Stanford PhD under Li, helped build the widely followed CS231n course and was central to real-time style transfer and PyTorch3D;

Mildenhall, from Google Research, coinvented the 2020 NeRF paper that energized 3D reconstruction.

Article illustration
△Left: Justin Johnson. Right: Ben Mildenhall.

In the celebratory posts released simultaneously by the two companies, World Labs also borrowed the line from Tennyson's Ulysses, “Friends, let us seek a newer world together,” echoing the title of Fei-Fei Li's earlier open letter.

Markets reacted more coolly than the companies’ celebrations.

After Monday’s November 16 opening, AMD initially fell around 5% amid a broader semiconductor and AI-infrastructure pullback.

Article illustration

Wall Street evidently awaits the new team’s first hardware–software integration results.

From $230 million to $8.2 billion: what did they build?

Li cofounded World Labs in 2024, bypassing the crowded LLM field for spatial intelligence that understands and simulates the 3D physical world.

On emerging from stealth that year, it raised $230 million at a billion-dollar valuation.

Last November, it released its first multimodal world model, Marble.

Text, images or rough 3D layouts can generate freely explorable, editable and exportable scenes.

In February this year, it raised another $1 billion, with Autodesk contributing $200 million alongside investors including Nvidia and AMD.

The release that most clearly showcased its technology was the September 1 next-generation general-purpose world model Atlas.

Article illustration

Li explained Atlas’s technical foundation at CoRL 2026 a few days earlier.

Where LLMs emphasize next-token prediction, Atlas’s idea is “Viewpoint Is All You Need”: predicting the next viewpoint under precise camera-pose conditions.

During training, Atlas took in text, images, video, depth signals and 3D structural signals together. Under the same autoregressive diffusion Transformer architecture, it can switch among three modes: dense reconstruction, sparse reconstruction and generation from text alone.

In a blind test initiated by the third-party organization RapidData, the evaluators collected around 500,000 human evaluations to compare the camera-controlled generation capabilities of major video and world models.

Atlas ranked first and can generate coherent viewpoint videos exceeding sixty seconds at 1440p.

Robotics applications matter even more.

World Labs also acquired Scenix this year, specializing in real-to-simulation and simulation-to-real robotics pipelines.

In a live demonstration, Fei-Fei Li showed robots completing their training entirely within virtual simulation environments built by Atlas, then being deployed directly to the real physical world in a zero-shot manner.

On November 12, the day before closing, World Labs announced full public and developer access to Atlas.

Competing with Nvidia, AMD develops models in-house

A few days before the closing announcement, Su and Li appeared together at CoRL 2026 for their first onstage conversation.

Su opened with the view that physical AI, particularly robotics, represents a multitrillion-dollar opportunity in coming years.

She disclosed that AMD already has a sizable robotics research team, although the overall strategy remains early-stage.

Li described “AI is computing, and computing is AI” as her view.

That matches Su’s closing-announcement emphasis on jointly optimizing computing platforms and AI models.

Asked earlier why World Labs did not build Atlas immediately, Li half-jokingly answered that Justin’s team needed enormous numbers of chips.

Scaling world models requires substantial GPU resources.

Previously partners, AMD and World Labs had jointly tuned model training and inference on AMD GPUs since 2025.

External partnerships retain commercial boundaries, making it harder to expose and fix fundamental problems before chip tape-out.

With the two companies now part of the same family, Fei-Fei Li's team can directly bring Atlas and subsequent frontier models into the architectural design of AMD's future Instinct-series chips and ROCm software stack. Before the hardware blueprints are even finalized, they can stress-test them with real spatial-intelligence workloads.

AMD has recently expanded rapidly: the $8.2 billion World Labs deal, an announced Taalas acquisition, a planned PsiQuantum investment and a data-center compute-leasing agreement with Core Scientific.

Nvidia, meanwhile, acquired Hugging Face for $12.9 billion in early September and has introduced Cosmos, Omniverse and GR00T for robotics and physical AI.

Facing Nvidia’s integrated hardware–software ecosystem, AMD repeatedly emphasized an open AI ecosystem from September signing through November closing.

It aims to offer developers an alternative combining open hardware, open software and internally developed frontier models.

Competition among chip giants can no longer be settled through one graphics card’s specifications.

This article was originally published by QbitAI. Author: Keleixi.

Related topics

After reading

Join the AINRK community

Read how to join the community and connect with people working in AI.

Community information (opens in a new tab)