AMD to Acquire Fei-Fei Li's World Labs for $8.2 Billion, Naming Her Chief Scientist
AMD will buy World Labs, the spatial intelligence startup co-founded by Fei-Fei Li, for $8.2 billion in stock. Li becomes executive vice president and chief scientist reporting to Lisa Su.
After the transaction closes, Li will join AMD as executive vice president and chief scientist, reporting directly to chair and chief executive Lisa Su. World Labs co-founders Justin Johnson and Ben Mildenhall will continue to lead the company's team inside AMD, forming a frontier AI research organization. The two sides said they plan to build an end-to-end AI ecosystem spanning hardware, software, platforms and widely accessible open models. The acquisition is expected to close before the end of 2026, subject to regulatory approvals and customary closing conditions.
The purchase followed a year and a half of deepening ties. AMD's investment arm took part in World Labs' B round in September 2024, when the company was valued at $1.3 billion post-money; angel investors in that round included Geoffrey Hinton, Andrej Karpathy and Jeff Dean. In 2025 the two companies began joint engineering work on training and inference optimization for AMD GPUs, covering distributed training, communication, mixed precision and memory management on the training side, and operator fusion, batching, quantization and compiler-level tuning for inference. For World Labs' world models, the work also involved video processing, three-dimensional representations and camera-controlled inputs.
The collaboration moved into public view at CES 2026, where Su and Li presented World Labs' Marble product together. The demonstration used footage of AMD's Silicon Valley office, with a small number of photographs converted into navigable 3D environments that preserved real geometry and scale while changing design style. On February 18, 2026, World Labs closed a $1 billion C round at a $5.4 billion post-money valuation, with AMD Ventures again participating alongside NVIDIA and Autodesk, which separately disclosed a $200 million investment. A month later, on March 19, AMD announced it would expand the partnership on AMD Instinct GPUs, covering workload optimization, cloud coordination and market support.
World Labs works on spatial intelligence, building world models that perceive, generate, reconstruct and simulate 3D environments. In early September 2026, about a month before the acquisition announcement, it released Atlas, a multimodal world model pretrained from scratch. Atlas natively handles text, images, video and 3D data using a multimodal autoregressive diffusion transformer architecture that organizes inputs into a shared spatial context and generates new content consistent with existing content in 3D space. World Labs calls the core capability novel view prediction: given one or more images and camera spatial relationships, the model generates views from positions, angles and paths it has not seen, analogous to a language model predicting the next token with visual content in 3D space as the output. Atlas can reconstruct real scenes from one to dozens of images and output new viewpoints and explicit 3D results, generate video up to one minute at 1440p with precise camera control, interpret spatial and temporal evolution from video, and produce images and 360-degree panoramas from text, with outputs including point clouds and 3D Gaussian splats. World Labs has also acquired SceniX, extending into robot simulation with reinforcement learning environment generation and Real-to-Sim workflows. Its user-facing Marble product lets users create, edit and share persistent high-fidelity 3D worlds from text, images, video or 3D structures, aimed at creative design, robot training and scientific discovery; Atlas will drive future versions of Marble.
AMD said these models represent a new class of AI workload whose computing characteristics differ notably from today's large language models, including large-scale vision and video training, multi-view spatial context processing, diffusion and autoregressive generation, 3D geometry and depth prediction, long high-resolution video output, robot simulation and low-latency real-time interactive inference. In its announcement, AMD said that as AI extends into inference, robotics, simulation and physical AI, infrastructure requirements will become more varied, and that World Labs' model development experience will help it shape future roadmaps by letting chip and system design teams engage directly with these workloads instead of adapting after models are fixed.
The deal also bears on AMD's competition with NVIDIA, which already offers open world models such as Cosmos while AMD had no comparable lineup. NVIDIA participated in World Labs' $1 billion round in February, and World Labs public materials have referenced Atlas being trained or demonstrated on NVIDIA Blackwell GPUs. Whether World Labs will keep supporting NVIDIA hardware or maintain a cross-platform deployment strategy after the acquisition has not been disclosed.
World Labs raised about $1.2 billion across its A, B and C rounds, beginning in April 2024, before agreeing to the $8.2 billion sale, roughly 1.5 times the $5.4 billion post-money valuation of its last round. Venture capital analyst Paul Bonnet estimated that the four co-founders — Li, Justin Johnson, Christoph Lassner and Ben Mildenhall — together with the team still hold about 56 percent of the shares, worth about $4.6 billion, and that after accounting for an option pool of about 15 percent, the four founders hold about $3.4 billion combined. The largest multiple belongs to A round investors a16z and Radical Ventures, which entered in April 2024 at a valuation of about $300 million; at the $8.2 billion price that is roughly a 19-fold return, or about $1.2 billion in gains, in under two and a half years. B round angels came in at a $1.3 billion valuation for about a 5-fold return and an annualized internal rate of return near 125 percent. C round backers committed $1 billion in February 2026 at a $5.4 billion valuation and exit about seven months later at 1.5 times. Bonnet put the blended return at about 2.9 times, with investors contributing roughly $1.2 billion and sharing about $2.3 billion in profit, because most of the capital arrived in the final round.