AMD buys Fei-Fei Li's World Labs to supercharge spatial AI

Language alone is no longer enough to get AI to the next level—it needs to understand how the physical world actually works. But generating high-fidelity 3D environments requires massive compute, and the biggest bottleneck has always been hardware.

AMD agreed to acquire spatial AI startup World Labs in an all-stock deal valued at roughly $8.2 billion. World Labs cofounder Fei-Fei Li will join AMD as executive vice president and chief scientist, reporting directly to CEO Lisa Su. The deal is expected to close before the end of 2026.

Beyond words to spatial intelligence

Fei-Fei Li cofounded World Labs in early 2024 alongside researchers Ben Mildenhall and Justin Johnson. The thesis was straightforward: the universe is made of physical objects, not just text tokens, and true frontier AI must reason over 3D structures.

The startup moved quickly. It launched its first commercial tool, Marble, on Nov. 12, 2025, allowing users to transform text, images, and basic layouts into editable 3D environments. After acquiring SceniX to bolster its robotics simulation stack, the lab unveiled its omni-model architecture, Atlas, on Sept. 1, 2026.

Much like an LLM predicts the next word, Atlas predicts unseen camera perspectives from 2D images, effectively tackling the sparse reconstruction problem by fusing generative models with multiview geometry.

Buying a seat at the frontier table

The acquisition deepens a relationship that started in 2025, when both teams began optimizing World Labs models for AMD GPUs. AMD was an early backer—Lisa Su invested early on, and Li made an appearance at AMD's CES presentation earlier this year. Interestingly, both AMD and Nvidia participated in World Labs' $1 billion round in February 2026, but AMD ended up taking the whole company.

As Li noted, scaling spatial intelligence requires getting much closer to the metal. Without tight hardware-software co-design, physical AI models become prohibitively expensive to run.

Bringing the team inside gives AMD an immediate answer to Nvidia, which has been pushing its own open-weight world models like Cosmos. World Labs’ workloads will now directly inform AMD's chip design and software roadmap.

Why it matters

For builders and creators, combining foundation models directly with silicon could dramatically lower the computing costs of generating 3D worlds and synthetic training environments for robots. However, Atlas invents plausible details to fill in unseen angles, meaning robotics teams will need to carefully test whether training on that hallucinated data is genuinely safe in the real world.

For the AI chip war, this is an aggressive and smart move by AMD. Instead of just chasing Nvidia’s software moat with generic driver updates, AMD is embedding one of the world's premier vision teams straight into its silicon design loop. The risk is scientific neutrality: while Li has promised to keep providing open models and platforms to the ecosystem, tethering a frontier lab to a single hardware vendor inevitably shifts research priorities toward AMD's own silicon roadmap.