Fei-Fei Li has successfully secured a highly sought-after stable position, sending a wave of panic through entrepreneurs in the embodied intelligence sector.

World Labs, the world model company founded by Li Feifei, has been acquired. Some believe this is a strong alliance between a computing power giant and a star startup; others call it a "betrayal" of the independent exploration of world models by the "Godmother of AI"; even more people claim it is a transfer of benefits.

On September 28, chip giant AMD officially announced that it has signed a definitive agreement with World Labs to acquire the AI unicorn, which was established only two years ago, in an all-stock transaction valued at approximately $8.2 billion (about 55 billion yuan). The transaction is expected to be completed by the end of 2026, but is still subject to regulatory approval and satisfaction of other customary closing conditions.

This is the second-largest acquisition in AMD's history, second only to the $50 billion acquisition of Xilinx in 2022.

On the surface, this is a win-win situation where both sides get what they need: World Labs obtains computing power support and continuous capital from a large manufacturer, while AMD gains a cutting-edge technology R&D team to quickly build core capabilities in the physical AI field.

But the wave of controversy arrived almost simultaneously with the official announcement. At present, the technical routes of world models have not converged, nor have they been proven commercially viable, let alone formed a mature business model. AMD spent a high price of $8.2 billion to acquire a company whose technical route is not necessarily correct and has not generated scaled revenue. An investor in the embodied intelligence industry said bluntly that Li Feifei's company has only completed demos in more than two years since its establishment, and has not figured out how to build a world model at all. It looks more like a cash-out exit.

After the acquisition news came out, many investors and entrepreneurs in the embodied intelligence industry began to worry about the prospects of world models. How much imagination space is left for the independent entrepreneurship story in the world model track?

01. Is it a strong alliance, or a cash-out by shareholders?

Prior to this acquisition, World Labs' latest financing round took place in February this year. At that time, the company, which had been established for less than two years, completed a $1 billion financing led by Autodesk, with participation from institutions including AMD, NVIDIA, Fidelity, and Emerson Collective, with a post-money valuation of about $5 billion. Calculated based on the $8.2 billion acquisition consideration, World Labs' valuation has increased by more than 60% in about 7 months.

Tracing further back, World Labs was jointly founded in early 2024 by Li Feifei together with Justin Johnson, Ben Mildenhall, Christoph Lassner and others, headquartered in San Francisco, with a technical focus on spatial intelligence and large-scale world models. Only half a year after its establishment, the company completed its first large-scale financing round of $230 million in September 2024, with top-tier institutions including a16z, NEA, and Radical Ventures participating, directly crossing the $1 billion unicorn threshold in valuation.

Regarding this acquisition, Lisa Su, Chair and CEO of AMD, explained: "Building computing platforms for the next generation of AI requires a deep understanding of how models evolve. Dr. Li Feifei and the World Labs team possess outstanding scientific research leadership and model expertise. We will work together to leverage these insights to develop the hardware, software and systems that power the next generation of AI, and grow the open AI ecosystem."

But in the eyes of some practitioners and investors, this transaction is far from as peaceful as the official statement suggests. There are two main points of controversy: first, why is a company that basically only has demos worth $8.2 billion; second, the technical route of world models has not yet been finalized, why did AMD make a hasty move?

"A listed company acquired a company with no revenue, no business model, and technology that is still in a very early stage at a high price, and the acquirer is also an early investor of this company. If this happened to an A-share listed company, many people would call it a joint cash-out," an analyst from a securities firm that has long focused on the embodied intelligence industry told AIX Finance.

The claim that "the acquirer is an early investor" is not unfounded. AMD Ventures, the investment arm of AMD, participated in World Labs' first financing round in September 2024, and continued to participate in the $1 billion financing round in February this year, listing on the shareholder register alongside NVIDIA; Lisa Su herself is also an early investor and staunch supporter of World Labs. This means that in this transaction, AMD is both the buyer and a shareholder of the acquired company.

However, judging from the immediate reaction of the capital market, this second-largest acquisition in AMD's history does not seem to have caused much of a stir. The news broke out after US stock market hours, and AMD's stock price barely fluctuated in after-hours trading.

What really makes practitioners skeptical is the technical strength of World Labs itself. A world model practitioner told AIX Finance that there are currently a huge number of technical routes for world models, no route has been verified as the correct answer, there are many domestic startups that take the same 3D route as World Labs, and Li Feifei's company has not built particularly strong competitive barriers.

The current technical schools of world models can be roughly divided into several categories: the 3D spatial intelligence school represented by World Labs, the non-generative latent space school represented by AMI Labs founded by Yann LeCun, and the platform infrastructure school represented by NVIDIA, among others. The underlying logic and implementation methods of each route are significantly different, and it is far too early to tell which one is better.

"At present, World Labs still faces the risk of being 'eliminated' by general large models. The 3D capabilities of general large language models are getting stronger and stronger. From a business perspective, selling World Labs at this point in time and at this price is probably the best window of opportunity," the aforementioned practitioner said.

At the same time, after Li Feifei sold the company, some world model practitioners are worried that the relevant technologies in the industry will be difficult to get through in a short period of time, or will eventually become a game only for large technology companies.

But at least for now, this acquisition will not have a direct impact on World Labs' continued R&D of world models. After the transaction is completed, Li Feifei will officially join AMD as Executive Vice President and Chief Scientist, reporting directly to Lisa Su, Chair and CEO of AMD. Two other co-founders of World Labs, Justin Johnson and Ben Mildenhall, will continue to lead the team in charge of model R&D.

Another point of controversy is how the $8.2 billion price tag is calculated. World Labs currently only has two products: one is Marble, a 3D scene generation platform for developers released at the end of 2025, which adopts a subscription model and generates a certain amount of revenue, but the revenue scale and profit level have never been disclosed to the public; the other is the underlying world model Atlas, which is currently still in the demo stage.

What is clear is that what AMD pays is not the commercial value of the company, but the technical premium for securing a seat at the table in the physical AI field in the future, as well as the value of the cutting-edge R&D team.

02. Li Feifei gets sufficient resources, AMD gets the admission ticket, and the world model track loses an independent startup

If we extend the timeline, this acquisition is not entirely unexpected.

At the end of 2016, Li Feifei, who was a tenured professor in the Department of Computer Science at Stanford University and director of the Stanford Artificial Intelligence Laboratory and Vision Laboratory, joined Google on academic leave. Two years later, she left Google and returned to Stanford. At that time, many voices argued that independent academics were not suitable for working in large technology companies.

It was not until around 2024, with the rapid advancement of large language model technology, that academia and industry began to explore the combination path of large language models and world models, hoping to leverage the capabilities of LLMs to make up for the shortcomings of world models in spatial cognition and temporal understanding, and the concept of world models became popular again.

In January of that year, Li Feifei started to found World Labs.

At the end of 2025, Turing Award winner, one of the three giants of deep learning, and former chief AI scientist at Meta, Yann LeCun, announced that he was leaving Meta to establish Advanced Machine Intelligence Labs (AMI Labs), with the goal of building systems that can understand the physical world, have persistent memory, and can reason and plan complex action sequences.

Two iconic figures in the AI field have successively left academia to start businesses, directly driving the world model track to boom. In China, about 10 world model-related startups were established throughout 2025, and the cycle from their establishment to the completion of large-scale financing was only 3 to 6 months. Entering 2026, according to statistics from ITjuzi, more than 20 new world model startups were established in the first 7 months alone, with massive capital pouring into the track.

The overseas market is also dominated by both giants and startups. In addition to World Labs and AMI Labs, there are NVIDIA, Meta, Google, Runway and other players.

Against this backdrop, the acquisition of a star world model company by a traditional chip giant has become a notable landmark event for this track.

For the acquired World Labs, selling at a good price at a time when the product is not yet mature and the industry is still in an unknown stage, brings at least two major benefits.

The first is capital. Essentially, world models, like large language models, are an industry that requires massive continuous capital investment. World Labs has raised a total of about $1.2 billion in the two years since its establishment, which sounds like a lot, but it is not generous in the cutting-edge model competition where expenses often reach billions of dollars. Backed by AMD, its capital pressure can be largely relieved.

The second is computing power. In the general-purpose GPU field, AMD is the world's second largest computing power chip supplier after NVIDIA. IDC estimates that NVIDIA accounts for about 81% of the AI accelerator market, and AMD accounts for about 13%. For World Labs, this means a stable, sufficient and deeply adapted computing power supply.

In fact, the cooperation between the two sides has been laid out long ago. World Labs has carried out in-depth technical cooperation with AMD since last year, starting from model training and inference optimization on AMD GPUs. At the CES in January this year, Lisa Su invited Li Feifei to the keynote speech stage, where Li Feifei said she couldn't wait to see its performance on the next-generation MI450. Nine months later, Lisa Su simply bought the entire company.

For AMD, the benefits of this transaction are equally clear.

The first is that it gains more cards to compete with its old rival NVIDIA. NVIDIA has already laid out Cosmos, a general training platform, in the world model field, and many startups have trained their own world models on Cosmos, such as XPeng and Figure AI. NVIDIA is trying to replicate its success in computing chips, binding hardware with software ecosystems, and becoming the "shovel seller" in the world model era.

Not long ago, NVIDIA also announced the acquisition of open source model platform Hugging Face for about $13 billion, firmly holding the developer hub in the large language model era. When its rival has secured the text large model ecosystem, AMD chose to open up a second battlefield, betting on spatial intelligence and physical AI. After acquiring World Labs, the 3D generation route it bets on is currently one of the mainstream routes of world models, and also one of the potential paths to AGI.

A deeper change lies in the inversion of the R&D model. Traditional chip manufacturers mostly adapt hardware after models are formed, passively letting "hardware chase after models". After AMD incorporates the model R&D force into its own system, it can align hardware design in advance with the computing power needs of next-generation spatial intelligence and robot simulation.

The second is brand label and narrative. Although AMD's market value once exceeded one trillion US dollars this year, and its share price has risen by more than 180% cumulatively within the year, it is still not in the absolute center of the AI boom narrative. The combination of labels such as "Godmother of AI", cutting-edge scientific research team, world model and physical AI provides AMD with a new valuation anchor.

For the entire world model industry, this transaction means that it has lost an independent star startup. At least for now, it is still too early to tell which technical route will make world models a success, or whether world models will only remain a technical concept at the demo stage, and it is difficult to become an independent business in the short term.

Under the capital boom, this acquisition makes it clearer that the industry is still in a chaotic and early stage. No one can guarantee that their own route is definitely correct, and no one can be sure that they can reach the finish line, even if she is the "Godmother of AI".

This article is from WeChat Official Account "AIX Finance", author: Wang Hanxing, editor: Wei Jia, published with authorization from 36Kr.