AMD Acquires World Labs for $8.2 Billion to Advance AI and Spatial Intelligence

Author | Hua Hua

If we rewind to 2014, Lisa Su had just taken over AMD.

At that time, AMD did not yet resemble the AI star company it is today. The company was under pressure, its business was struggling, and some even doubted whether this veteran chipmaker would survive to see another opportunity.

Also during that phase, Li Fei-Fei was doing something that seemed even more tedious: preparing a large enough collection of images for the machine.

On September 28, AMD announced its acquisition of World Labs, founded by Fei-Fei Li, for approximately $8.2 billion in an all-stock transaction. Upon completion of the deal, Fei-Fei Li 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. Dig deeper, and it’s two paths that have been traveled for over a decade finally converging.

In 2014, Lisa Su became CEO of AMD.

At that time, AMD had been struggling for several consecutive years.

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

At the time, AMD’s challenges were easy to understand: it had limited resources, Intel dominated the CPU market, and Nvidia firmly controlled the growth opportunities in GPU computing.

However, after Lisa Su took the helm, AMD did not expand its front lines further.

She did the opposite, reallocating limited engineering resources toward high-performance computing.

In 2017, AMD launched Ryzen processors based on the Zen architecture, alongside EPYC processors designed for data centers. Zen became AMD’s pivotal product line for re-entering the high-performance computing market.

That year, AMD's revenue rebounded to $5.33 billion, and the company returned to annual profitability.

Subsequently, EPYC entered the server market, and the Instinct series GPUs began handling AI and high-performance computing tasks.

AMD's business focus has also gradually expanded from personal computer processors to data centers.

In 2025, AMD achieved annual revenue of approximately $34.6 billion, with data center revenue reaching about $16.6 billion, a 32% year-over-year increase.

Over the course of more than a decade, a company transformed from a chipmaker on the verge of being marginalized by the market into a central player in the AI infrastructure race.

Lisa Su's strategy is clear: capitalize on shifts in computing demand to bring chips into larger computing markets.

But chip companies have a natural problem.

Chips need to be designed many years in advance.

Waiting until a computational demand has become consensus before preparing hardware is often too late.

The real challenge for chip companies isn't catching up to already defined demands—it's anticipating the next computing trend.

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

Li Fei-Fei's relationship with AI can be traced back to a collection of photographs.

Around 2006, she began leading the ImageNet project.

At the time, much of the research in computer vision focused on enabling machines to identify objects within images. Machines could determine whether a photo contained a cat, but struggled with more complex visual information.

ImageNet eventually accumulated over 15 million images, covering approximately 22,000 categories.

Images were collected from the internet, and their classification and labeling required extensive human involvement. Li Fei-Fei’s team even used Amazon Mechanical Turk to break down the massive image labeling task among a large number of temporary workers worldwide.

Amazon Mechanical Turk is a crowdsourcing platform launched by Amazon, allowing businesses to distribute large volumes of small tasks to users worldwide. The labeling work for ImageNet was carried out using this platform.

This is an extremely tedious task.

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

In 2012, AlexNet achieved a breakthrough performance in the ImageNet image recognition challenge, bringing deep neural networks into the center of computer vision research on a large scale.

One of ImageNet's most important contributions was showing researchers that large-scale data itself can be a powerful driver of progress in machine learning.

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

Faces, objects, scenes, and text are gradually becoming visual information that can be processed by algorithms.

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

She began to continue exploring the spatial relationships behind the visuals.

A photo shows a table.

Computer vision systems can identify "tables".

But once entering the real world, machines must handle tasks far more complex than simply identifying an object.

It needs to determine where the table is, how far the tabletop is from the robot, where the cup is placed on the table, whether taking two steps forward would cause the robot to collide with the table corner, and whether other items on the table would be affected after picking up the cup.

All these questions involve space.

Li Fei-Fei later expanded her research focus 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 understand the three-dimensional world.

The models released by World Labs no longer merely generate two-dimensional images from text; they can now generate three-dimensional environments from text or images, and further enable these environments to be exploratory and editable.

This means the objects the model processes have changed.

Text models process relationships between words; image models process relationships between pixels. Spatial intelligence models go further by integrating objects, distances, positions, time, and actions into a single system.

In July 2026, World Labs acquired the spatial intelligence company SceniX, further advancing its focus on robotics and simulation.

This path is becoming increasingly clear, enabling the model to build an internal environment capable of continuously understanding, predicting, and simulating the real world.

This capability is especially important for bots.

When a robot faces a cup, identifying "this is a cup" is only the first step. It also needs to understand the cup’s spatial position, shape, weight, how to grasp it, and the potential outcomes of its own actions.

ImageNet addresses object recognition, while World Labs aims to solve the position, relationships, and changes of objects in three-dimensional space.

The primary computational units for large language models are tokens, whereas world models deal with far more complex inputs, requiring simultaneous processing of space, time, object relationships, and environmental changes, while also enabling the model to predict the outcomes of actions.

These models require large amounts of data, as well as more intensive training and operation.

More importantly, it naturally connects to fields such as simulation environments, robotics training, autonomous driving, game development, and digital twins.

Once the model can predict real-world changes within a virtual environment, the nature of computational tasks will change.

Past computers primarily processed rules defined by humans. World models aim to enable machines to build their own internal representation of the real world and use this representation to make predictions and decisions.

This will alter the underlying computations, the data the GPU must process, memory capacity, and the architecture for training and inference—all of which will adapt to the model's form.

This is exactly where AMD desperately needs it.

A chip company that designs products solely based on today’s AI demands will always be playing catch-up.

If you could anticipate what the next computational task will require, you have the opportunity to design hardware and software around that task from the very beginning.

World Labs provides precisely a window into future computing demands.

The relationship between AMD and World Labs did not happen suddenly.

In February this year, World Labs completed a $1 billion funding round, with AMD participating.

This funding round values World Labs at approximately $5 billion.

World Labs' spatial intelligence model runs on AMD Instinct GPUs, with both teams collaborating to optimize the model's training and inference.

For AMD, this collaboration allows direct observation of how new models perform in real-world computing environments.

Information about which stages consume the most computational power and where memory bottlenecks are likely to occur is far more valuable than simply reading a business plan.

Acquiring World Labs, AMD provided a clear rationale: World Labs' research will help AMD understand emerging AI workloads and influence the design of next-generation hardware, software, and systems.

AMD's intentions are now clear.

AMD isn't just buying a model—it aims to embed its understanding of next-generation AI workloads directly into its product development cycle.

The AI industry has established clear computing demands: training large models requires massive amounts of GPUs, running models demands high efficiency, and cloud providers must continuously build data centers.

These requirements can all be quantified and used by chip companies to plan their product roadmaps.

The world model is not yet this clear.

Will the world model first be applied to robotics, autonomous driving, or gaming? Can it generate scalable revenue? How should hardware be aligned? The industry is still exploring these questions.

Even the definition of world models themselves is still evolving, and this precisely creates an opportunity for chip companies.

If you wait until the world model becomes a mature industry to design corresponding hardware, competition may already be underway.

AMD chose to acquire World Labs, effectively bringing a team dedicated to researching next-generation computing needs directly into its internal organization.

This is completely different from traditionally buying an AI application.

Apps can bring users and revenue, while the value created by the research team may be reflected in chip architectures, software stacks, and computing systems years down the line.

AMD previously reclaimed the CPU market with Zen, and later entered the data center and AI computing sectors with EPYC and Instinct.

World Labs corresponds to what tasks will drive the next round of computational growth.

For AMD, instead of guessing, it’s better to bring the people researching the next generation of computing right beside you.

This acquisition also created an interesting combination.

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

The mountain hasn't formed yet, but some people have already started building steps for it.

[Outside the layout]:

In the past, humans defined problems first and then had machines calculate the solutions.

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

The world itself is becoming a calculation.

This article is from the WeChat official account "Beyond the Layout," authored by Hua Hua.