AMD will buy World Labs for $8.2 billion to deepen AI research

What World Labs is building

Founded in San Francisco in early 2024 by Li and researchers Ben Mildenhall and Justin Johnson, World Labs has about 70 employees. It develops “world models”: AI programs that create, reconstruct and simulate three-dimensional spaces from text, images and video.

In early September, the startup introduced Atlas, its first world model. It can create, reconstruct and simulate full spatial scenes from just a few images.

Li’s argument is that language alone is not enough to build capable AI. In a blog post, she wrote that the universe is made of real things, not words. In a longer essay, she argued that today’s multimodal language models struggle with spatial reasoning: when estimating distance, direction and size, they perform only slightly better than chance. They also cannot mentally rotate objects or predict simple physical events.

World Labs sees physical-world reasoning as relevant to tasks across science, entertainment and robotics. The idea has attracted attention beyond the startup. Google DeepMind co-founder Demis Hassabis has described world models as an important step toward artificial general intelligence. A research group led by Peking University proposed OpenWorldLib, a standardized definition requiring world models to perceive their surroundings, interact with them and remember past interactions.

AMD is buying research as well as a team

Su told Bloomberg Television that understanding the full process behind AI workloads helps AMD build better systems. The acquisition, Li wrote, reflects the need to connect research more closely with hardware: without deliberate work on the hardware side, AI efficiency is limited.

That logic also puts the deal in the context of AMD’s competition with Nvidia. AMD’s market capitalization is about $1 trillion, compared with Nvidia’s roughly $5 trillion. Nvidia has spent years pairing AI chips with models and services, including open-weight models and offerings built to run on its hardware. Its recent $13 billion purchase of AI platform Hugging Face extended that ecosystem strategy.

World Labs had raised $230 million in a 2024 Series A round, with Andreessen Horowitz, Nvidia and AMD among its investors. In early 2025, it raised another $1 billion from investors including Autodesk. Bloomberg valued the startup at about $5 billion at the time. The $8.2 billion price represents a several-fold increase in just a few months.

The price is for potential, not proof

I think the most revealing part of the announcement is what it emphasizes: AMD is buying expertise and a research team, not a proven commercial product. Atlas is an early model, and the source does not describe a mature business or a product pipeline that could justify the price on its own.

The acquisition may also suggest that even World Labs’ substantial fundraising was not enough to compete independently with leading AI labs and technology companies spending at enormous scale. That is an inference, not a stated reason for the sale. But it highlights the tension in AMD’s strategy: the company is paying a premium to learn how future AI workloads might work, while the hardware that could benefit from that knowledge has yet to be designed.