AMD said on Sept. 28 that it would acquire World Labs, the spatial-intelligence startup founded by AI researcher Fei-Fei Li, in an all-stock deal valued at about $8.2 billion. The transaction, expected to close by the end of 2026, would make Li an executive vice president and chief scientist at AMD, reporting directly to Chair and CEO Lisa Su.
The deal would be AMD’s second-largest acquisition, behind its roughly $50 billion purchase of Xilinx in 2022.
The relationship between the two companies predates the acquisition. AMD was an early investor in World Labs, and the companies began working together last year on model training and inference optimization. When World Labs raised $1 billion in February, both AMD and Nvidia were among its investors.
The announcement quickly drew attention across the AI and investment communities. It also renewed interest in a less obvious beneficiary of advances in world models: XR.
In an interview with Pimax founder Robin Weng at the company’s Shanghai office, Weng said he saw the acquisition as part of a much larger shift in how computing will evolve as AI moves into three-dimensional environments.
There was also a curious coincidence. AMD China and Pimax share the same address: 799 Naxian Road in Shanghai’s Zhangjiang technology district.
Weng has spent the past decade building Pimax around spatial computing and high-end VR. The company recently completed a corporate restructuring as it prepares for an initial public offering.
For Weng, the deal was at once expected and unexpected.
The expected part was AMD’s interest in World Labs. Weng believes a future 3D internet, in which people interact with persistent three-dimensional environments through XR devices, will require far more computing power than today’s largely two-dimensional digital world. From that perspective, he said, the strategic logic was easy to see.
The surprise was the price.
Roughly $8.2 billion for a company founded only about two years ago was enough to make him smile. “You can’t help but be a little envious,” he joked.
Still, true to his engineering instincts, Weng was more interested in the technology behind the deal than in the valuation.
A 3D world has a much bigger compute bill
World models are expensive.
In a technical presentation, Pimax estimated that generating and maintaining interactive 3D environments could require 100 to 1,000 times as much computing power as video-generation systems such as Sora.
World Labs has made a similar point in its own technical writing. It has argued that simply extending today’s video-model architectures to interactive 3D would quickly become impractical.
A 4K interactive video stream running at 60 frames per second, the company has said, could require the generation of more than 100,000 tokens every second. Keeping a world coherent over more than an hour of interaction could involve a context exceeding 100 million tokens, a workload World Labs has described as both technically and economically unrealistic on today’s infrastructure.
XR adds another layer of demand.
A VR headset must render separate views for both eyes while responding almost instantly to the user’s movements. The industry generally targets refresh rates of around 90 hertz and motion-to-photon latency below roughly 20 milliseconds to reduce discomfort.
Pimax’s Crystal Super illustrates the scale involved. The headset has a resolution of 3,840 by 3,840 pixels per eye, or roughly 29 million pixels across both displays, about 3.5 times the pixel count of a 4K screen. Running demanding 3D applications at the headset’s native resolution typically requires a high-end GPU such as Nvidia’s RTX 4090 or 5090.
The result is a computing problem at both ends: enormous workloads in the cloud to generate and maintain 3D worlds, and equally demanding hardware at the edge to render and interact with them in real time.
What AMD may be buying with Atlas
World Labs’ latest model, Atlas, offers a glimpse of where that demand could go.
The company describes Atlas as a multimodal world model that works across text, images, video and 3D. From a single image, it can generate scenes in which the camera position can be controlled, allowing a user to move through the generated environment from different viewpoints.
World Labs has not disclosed how much computing power was required to train Atlas. But it has said the model emerged from a series of training runs that increased progressively in scale, with additional capabilities appearing as more compute was added.
AMD has already been moving closer to that workload. In March, its venture arm said World Labs was expanding its use of AMD Instinct GPUs for world-model computing. Li, in announcing her move to AMD, also argued that AI systems would be constrained in both efficiency and scale without dedicated advances in hardware.
That linkage between model development and chip design is central to how Weng reads the acquisition.
AMD has said World Labs’ frontier-model research could give the chipmaker insight, years in advance, into what future AI systems will require from hardware.
Weng sees a more strategic ambition: to put AMD’s chips closer to the center of the world-model stack as the technology evolves, potentially making its hardware a native computing platform for the next generation of 3D AI.
In that sense, AMD is buying more than a model company. It is also making a bet on the amount and type of computing power that world models may eventually demand.
Two markets that could move with world models
Weng expects world models to develop rapidly as more computing power becomes available, drawing a comparison with the acceleration of large language models after 2017.
Two industries, he said, stand to benefit directly: Physical AI and XR.
The first is already a major target for the semiconductor industry.
Robots and autonomous vehicles need exposure to vast numbers of situations before they can operate reliably in the physical world. Collecting all of that experience in real environments is expensive and slow, while rare or dangerous scenarios are particularly difficult to reproduce.
World models offer another route. Interactive simulated environments can give machines places to train repeatedly, encounter variation and fail safely before being deployed in the real world. That was one of the ideas behind Li’s decision to create World Labs.
AMD has similarly pointed to robotics, simulation and Physical AI as areas that will diversify demand for AI infrastructure. Nvidia is pursuing the same opportunity through Cosmos, its open family of world models, making AMD’s acquisition of World Labs part of a broader race to supply the computing layer behind Physical AI.
Simulation, however, is only part of the equation. Training machines to operate in the real world also requires data captured from actual human activity.
Pimax recently joined the Hive Data Co-creation Initiative launched by Maniformer and is exploring multimodal, first-person data-capture systems. That work is aimed at gathering the kinds of real-world interaction data that simulated environments alone cannot provide.
The second market is XR.
For years, one of XR’s biggest constraints has been content. Building high-quality 3D environments by hand is expensive and time-consuming, limiting how quickly immersive experiences can be created.
World models could reduce that bottleneck.
World Labs’ Marble system can generate explorable 3D environments from text or images. Those environments can be exported to engines such as Unity and Unreal, or rendered directly in VR.
If creating 3D worlds becomes dramatically easier, the amount of content available for XR could expand with it. That, in turn, would increase demand for higher-resolution displays, faster rendering and more powerful edge computing.
For Weng, Physical AI and XR meet at the same underlying idea: spatial intelligence.
One asks machines to understand and act in physical space. The other gives people a way to perceive and interact with digital space. Both depend on systems that can represent, generate and compute in three dimensions.
That is where Weng believes Pimax’s decade of work could take on a different significance.
“We’ve spent the past 10 years building for gamers. And somehow, we’ve run straight into Physical AI,” Weng said.
He sees a parallel with Nvidia.
“Nvidia started out building graphics cards for gamers. About a decade ago, AI suddenly gave that technology an entirely new role. Today, Nvidia has become part of the foundation of AI.”
“Pimax may be at a similar turning point,” Weng said. “We want to become a general-purpose computing platform for the spatial computing era.”