Lisa Su looked up, Fei-Fei Li looked down

Author | Huahua

If we turn the clock back to 2014, Lisa Su had just taken over AMD.

Back then, AMD showed no sign of becoming the AI star company it is today. The company was under huge pressure, its business was mired in difficulties, and some even doubted that this veteran chip firm would not survive to seize the next round of opportunities.

It was also during that period that Li Fei-Fei was working on a seemingly more "foolish" task: preparing a sufficient number of images for machines.

On September 28, AMD announced that it would acquire World Labs founded by Li Fei-Fei in an all-stock transaction valued at approximately $8.2 billion. After the deal closes, Li Fei-Fei will join AMD as Executive Vice President and Chief Scientist, reporting directly to Lisa Su.

On the surface, this is a chip company acquiring an AI startup. Looking deeper, it marks the convergence of two paths that have been advancing for more than a decade.

1. Lisa Su Pulled AMD Back From the Brink

In 2014, Lisa Su took over as CEO of AMD.

By that time, AMD had been underperforming for years in a row.

In 2014, AMD's revenue was around $5.5 billion with a net loss of $403 million; in 2015, its revenue further dropped to $3.99 billion, and the net loss widened to $660 million.

AMD's predicament at that time was not hard to understand: it had limited resources, Intel held an absolute dominant position in the CPU market, and Nvidia firmly controlled the growth opportunities in GPU computing.

But after Lisa Su took the helm, she did not expand AMD's business lines to cover more areas.

Instead, she took the opposite approach, reallocating the limited engineering resources to high-performance computing.

In 2017, AMD released the Ryzen processors based on the Zen architecture, and launched the EPYC processors for data centers at the same time. Zen became the key product line for AMD to re-enter the high-performance computing market.

In that year, AMD's revenue rebounded to $5.33 billion, and the company returned to full-year profitability.

After that, EPYC entered the server market, and the Instinct series GPUs began to undertake AI and high-performance computing tasks.

AMD's business focus gradually extended from personal computer processors to data centers.

By 2025, AMD's full-year revenue reached approximately $34.6 billion, of which data center business revenue was around $16.6 billion, a year-on-year increase of 32%.

It took the company more than a decade to transform from a chip firm that was almost marginalized by the market to a core player in the competition for AI infrastructure.

Lisa Su's path is very clear: seize the changes in computing demand, and push chips into larger computing markets.

But chip companies have an inherent problem.

Chips need to be designed many years in advance.

By the time a computing demand has become a universal consensus, it is often too late to start preparing hardware.

The real difficulty for chip companies is not to catch up with clearly defined demands, but to figure out how to predict the next computing trend in advance.

AMD needs to know which new tasks will become important before a consensus is formed. This is one of the reasons why AMD acquired World Labs.

2. Li Fei-Fei Led Machines to Learn to "See" First

The story of Li Fei-Fei's connection with AI can start from a pile of photos.

Around 2006, she began to lead the ImageNet project.

At that time, a large part of computer vision research focused on enabling machines to identify objects in images. Machines could tell if there was a cat in a photo, but could hardly handle more complex visual information.

ImageNet eventually accumulated more than 15 million images, covering about 22,000 categories.

The images were collected from the Internet, and classification and annotation required a large amount of manual participation. Li Fei-Fei's team even used Amazon Mechanical Turk to break down the huge image annotation work and assign it to a large number of temporary workers around the world.

Amazon Mechanical Turk is a crowdsourcing platform launched by Amazon, where enterprises can split a large number of fragmented tasks to global online users. The annotation work of ImageNet was completed with its help.

This was an extremely trivial and tedious task.

But it laid an indispensable foundation for the subsequent breakthroughs in deep learning.

In 2012, AlexNet achieved breakthrough results in the ImageNet Image Recognition Challenge, and deep neural networks began to move into the center of computer vision research on a large scale.

One of the most important contributions of ImageNet is that it made researchers realize that large-scale data itself can become a key force driving the progress of machine learning.

Since then, the ability of machines to recognize images has improved rapidly.

Faces, objects, scenes, and texts have gradually become visual information that can be processed by algorithms.

Li Fei-Fei did not stop at the question of "what can machines see".

She began to further explore the spatial relationships behind visual perception.

3. From Image Recognition to Spatial Intelligence

There is a table in a photo.

A computer vision system can recognize the "table".

But when entering the real world, the tasks machines need to handle are far more complex than recognizing a single object.

It needs to judge where the table is, how far the tabletop is from the robot, where the cup is placed on the tabletop, whether the robot will hit the corner of the table if it walks two steps forward, and whether other items on the table will be affected after the cup is picked up.

All these problems are related to space.

Li Fei-Fei later extended her research direction further to spatial intelligence.

In 2024, she co-founded World Labs with Justin Johnson, Rob Fergus, Christoph Lassner and others, with the goal of enabling AI to build an understanding of the three-dimensional world.

The models released by World Labs no longer only generate 2D images based on text, but can generate 3D environments based on text or images, and further endow these environments with characteristics of being explorable and editable.

This means that the objects processed by the model have changed.

Language models process the relationships between words, image models process the relationships between pixels. Spatial intelligence models go a step further, processing objects, distances, positions, time and actions in the same system.

In July 2026, World Labs acquired spatial intelligence company SceniX, continuing to deepen its layout in the direction of robotics and simulation.

This development path is becoming increasingly clear: to enable the model to build an internal environment that can continuously understand, predict and simulate the real world.

This capability is especially important for robots.

When a robot faces a cup, recognizing "this is a cup" is only the first step. It also needs to know the spatial position, shape, weight and grasping method of the cup, as well as the possible results of its own actions.

ImageNet solved the problem of object recognition, while World Labs aims to solve the position, relationship and change of objects in the three-dimensional space.

4. World Models Have Changed the Objects of Computing

The main computing object of large language models is Token, while the computing objects faced by world models are far more complex: they need to process space, time, object relationships and environmental changes at the same time, and enable the model to predict the results after an action occurs.

This type of model requires a large amount of data, as well as more resource-intensive training and operation.

More critically, it is naturally connected to fields such as simulation environment, robot training, autonomous driving, game development, and digital twin.

Once the model can predict the changes of the real world in the virtual world, the tasks processed by computers will change accordingly.

Computers in the past mainly processed rules predefined by humans. World models try to let machines build their own internal representation of the real world, and then use this representation for prediction and decision-making.

This will change underlying computing: the data that GPUs need to process, memory capacity, training and operation architecture, and software stack design will all be adjusted according to the model form.

This is exactly what AMD urgently needs.

A chip company that only designs products based on today's AI demands will always be in a position of playing catch-up.

If it can figure out what the next type of computing task needs in advance, it will have the opportunity to design hardware and software around this task from the very beginning.

What World Labs provides is exactly a window to observe future computing demands.

5. From Investment to Acquisition, It Took Only Half a Year

The partnership between AMD and World Labs did not happen out of the blue.

In February this year, World Labs completed a $1 billion financing round, in which AMD participated.

This financing round valued World Labs at approximately $5 billion.

World Labs' spatial intelligence models run on AMD Instinct GPUs, and the teams of both sides are jointly optimizing the training and operation of the models.

For AMD, this kind of cooperation allows it to directly observe the performance of new types of models in real computing environments.

Information such as which links consume the most computing power and where bottlenecks easily occur in memory is far more valuable than simply reading a business plan.

AMD's reasons for acquiring World Labs are also very clear: World Labs' research can help AMD understand emerging AI workloads, and influence the design of next-generation hardware, software and systems.

AMD's intention is already obvious.

What AMD bought is not just a model. It hopes to integrate the understanding of next-generation AI workloads directly into its own product design cycle.

6. What AMD Bought Is an Undefined Market

The AI industry has formed clear computing demands: training large models requires a large number of GPUs, running models requires high efficiency, and cloud vendors need to continuously build data centers.

All these demands can be quantified, and can be used by chip companies to plan their product roadmaps.

World models do not have such clear definitions yet.

It is still being explored in the industry whether world models will first land in robotics, autonomous driving or gaming, whether they can generate revenue at scale, and how hardware should cooperate with them.

Even the definition of world model itself is still evolving, but this precisely leaves a window of opportunity for chip companies.

If you wait until the world model becomes a mature industry before designing the corresponding hardware, the competition will likely have already started.

AMD chose to acquire World Labs, which is equivalent to directly bringing a team that studies next-generation computing demands into its internal organization.

This is completely different from buying an AI application in the traditional sense.

Applications can bring users and revenue, while the value brought by the research team may be reflected in the chip architecture, software stack and computing systems several years later.

AMD previously recaptured the CPU market through Zen, and entered the data center and AI computing fields through EPYC and Instinct.

What World Labs corresponds to is what kind of tasks will drive the next round of computing growth.

For AMD, instead of guessing, it is better to bring the people who study next-generation computing right beside them.

This acquisition has thus created a very interesting combination.

Lisa Su looks up, searching for the next peak of computing. Li Fei-Fei looks down, seeing the real world at the foot of the mountain.

The mountain has not yet taken shape, but some people have already started building steps for it.

Words beyond the layout:

In the past, humans defined problems first, and then let machines do the calculation.

When machines begin to understand the real world, the order may be reversed.

The world itself is becoming a computing problem.

This article is from the WeChat official account "Beyond the Layout", author: Huahua, published with authorization from 36Kr.