Author｜Zhang Qiming

Editor｜Chen Xiaoran

Two Chinese-American Scientists Join Forces, Propelling Silicon Valley's AI Competition into the 3D Realm

Despite being established for just two years, boasting a team of around 70 members, and keeping its revenue scale under wraps, World Labs has been appraised at a staggering $8.2 billion.

What AMD has acquired is Fei-Fei Li's team and a gateway to the era of physical AI.

A Deal of Monumental Significance

In the wee hours of Beijing time on September 29, AMD unveiled that it had inked a definitive acquisition agreement with World Labs, a spatial intelligence firm founded by Fei-Fei Li.

The deal, valued at roughly $8.2 billion (equivalent to approximately RMB 55 billion), will be entirely financed through AMD stock and is slated to be finalized by the end of 2026, pending regulatory approval and the fulfillment of other customary closing conditions.

Once the transaction is completed, Fei-Fei Li will join AMD as Executive Vice President and Chief Scientist, reporting directly to Chairman and CEO Lisa Su.

World Labs co-founders Justin Johnson and Ben Mildenhall will continue to helm the original team alongside Fei-Fei Li, driving forward research in world models and spatial intelligence.

This is a deal of monumental significance.

World Labs was established in early 2024 by Fei-Fei Li, Ben Mildenhall, Justin Johnson, and Christoph Lassner. By September of the same year, the company made its official debut with over $230 million in funding.

At that juncture, media reports indicated that its valuation had already surpassed $1 billion, with investors including a16z, NEA, Radical Ventures, and industrial capital from the likes of NVIDIA, AMD, and Adobe.

In February 2026, World Labs secured another $1 billion in funding, with participation from AMD, NVIDIA, Autodesk, Fidelity, Emerson Collective, and Sea, among others. Autodesk invested $200 million.

However, World Labs did not reveal its official valuation post this funding round. The previously reported figure of approximately $5 billion was merely a reference during the funding negotiation phase.

From a valuation exceeding $1 billion in 2024 to the current transaction price of $8.2 billion, World Labs' book value has ballooned approximately eightfold in two years. Nevertheless, this multiple does not equate to the actual returns for early investors.

World Labs has undergone multiple funding rounds, diluting the ownership stakes of original shareholders. Moreover, since this is an all-stock transaction, the final returns will also hinge on holding costs, transaction terms, and fluctuations in AMD's stock price.

Based on the approximately 70 employees reported by foreign media, the per-employee transaction value at World Labs exceeds $100 million. However, this does not imply that each employee has "created" $100 million in cash. What AMD is purchasing is the technical team, model capabilities, intellectual property, product access, and the opportunity to shape the next generation of AI infrastructure.

For a research-oriented company with undisclosed revenue and a mere two-year existence to be valued at $8.2 billion, it remains one of the most remarkable outcomes in AI entrepreneurship history.

Delivering on the Roadmap

Behind the lofty valuation lies World Labs' strategy of steering clear of the most crowded track of large language models and focusing its main resources on spatial intelligence.

Fei-Fei Li succinctly summarizes this approach: the universe is not just composed of words but of tangible things.

While language models can predict the next token based on context, they inherently lack an understanding of distance, depth, occlusion, gravity, motion trajectories, or the consequences of actions.

For AI to venture into robotics, autonomous driving, industrial simulation, spatial computing, and scientific research, it must comprehend 3D environments, infer how the world evolves, and predict potential outcomes after actions.

Based on this insight, World Labs has gradually unfolded its product roadmap over the past two years.

The first product to materialize was Marble. This multimodal world model can generate spatially coherent 3D worlds that can be continuously explored and edited based on text, images, videos, and rough 3D layouts. Unlike generating a single image or video, Marble endeavors to create a space where perspectives can be altered, navigation can continue, and subsequent creations can be made.

Subsequently, World Labs ventured into robotics simulation. In July 2026, World Labs acquired robotics company SceniX, bridging world models, learning-based simulation, and real-world robot training. Robots can first undergo rigorous training and testing in virtual environments before venturing into the real world, mitigating the costs and safety risks associated with real-world trial and error.

Atlas, unveiled on September 1, represents the most comprehensive realization of this roadmap to date.

According to World Labs, Atlas is a multimodal autoregressive diffusion Transformer trained from scratch, capable of uniformly processing text, images, videos, and 3D information. It can predict new observation perspectives based on existing images, reconstruct real-world scenes from a limited number of images, and output explicit 3D results.

While large language models predict the next token along a text sequence, Atlas attempts to predict the "next perspective" based on existing spatial information. Fei-Fei Li stated that this approach, merging generative models with multi-view geometry, has to a considerable extent resolved the long-standing sparse reconstruction problem in computer vision.

This capability can be applied to robotics reinforcement learning environments, architectural and real estate reconstruction, film and entertainment, medical scene generation, and industrial digital twins.

However, Atlas is still in the early access stage for a select group of partners, and World Labs has not disclosed its revenue or customer scale.

The journey from technological breakthrough to scalable revenue remains a protracted one. Generating a "plausible-looking" 3D space is not synonymous with creating a physically accurate simulation environment where robots can train safely.

What truly renders the $8.2 billion valuation steep is the possibility that World Labs could emerge as the gateway to the next generation of AI.

Buying into the Future

For AMD, the significance of this acquisition extends far beyond adding another AI product to its portfolio; it is about preemptively entering the stage of defining next-generation models and computational workloads.

Over the past few years, AMD has been systematically completing its AI infrastructure portfolio through acquisitions. In 2022, the company finalized its acquisition of Xilinx for a transaction value of $49 billion and acquired DPU vendor Pensando for approximately $1.9 billion. In 2024, it acquired European private AI lab Silo AI for $665 million, and in 2025, it completed the acquisition of ZT Systems, with a final confirmed transaction price of approximately $4.4 billion.

World Labs propels AMD's acquisition portfolio further into the model and application layers. Based on the currently announced transaction value, this will become AMD's second-largest acquisition in history, following Xilinx, and the company's most audacious bet in the AI model space.

AMD is no longer content with merely selling chips to model companies. Lisa Su's judgment is that the deeper the understanding of models, software, and end-to-end workloads, the better the ability to design superior computing systems.

This strategy already has a tangible foundation. In the second quarter of 2026, AMD's data center business revenue reached $6.7 billion, a year-on-year increase of 107%.

OpenAI and Meta have each planned to deploy up to 6 gigawatts of AMD GPUs, and Oracle previously announced the construction of its first batch of 50,000 MI450 GPU clusters.

However, with NVIDIA still wielding a significant platform advantage, AMD cannot rely solely on catching up in GPU performance. The computational structures forged by robotics simulation, spatial reconstruction, and world models, as well as the new demands for memory, networking, low-latency inference, and hardware-software synergy, will all influence next-generation chip design.

World Labs thus resembles a "next-generation workload radar" set up within AMD, developing models while also shaping hardware in reverse. If physical AI becomes the new growth trajectory after language models, AMD will receive demand signals earlier and even participate in setting technical roadmaps.

For Fei-Fei Li, joining AMD is not merely because "compute is expensive." Cutting-edge models necessitate not just compute but also the co-iteration of chips, systems, software, data, and applications. Continuing to raise funds independently means constantly balancing research investment, cloud computing costs, and commercialization. Joining AMD provides access to a more comprehensive engineering system, customer network, and the opportunity to influence hardware roadmaps.

Risks persist. The transaction has not yet been finalized, and regulatory and closing uncertainties loom. Spatial intelligence offers vast imaginative space but has not yet crystallized into a mature market comparable to language models. Cutting-edge research emphasizes free exploration, while public companies prioritize roadmaps, delivery cycles, and return on investment. Whether the two sides can maintain equilibrium remains to be seen.

An all-stock transaction also implies that World Labs shareholders are deeply entwined with AMD's stock price. The $8.2 billion represents the transaction value at the time of signing the agreement, not cash already in hand. Any fluctuations in AMD's stock price and the AI capital expenditure cycle will impact actual returns.

What AMD has procured for $8.2 billion is an option to step into the era of physical AI. The past AI competition revolved around who could train larger language models; the next phase will pivot to who can enable models to comprehend space, infer actions, and infiltrate the real world.

What AMD has acquired with World Labs is the possibility that the main battlefield will shift.