AMD Acquires World Labs for $8.2 Billion in All-Stock Deal, Expected to Complete by End of 2026, Subject to Regulatory Approval
AMD announced a final agreement to acquire World Labs, a space intelligence laboratory led by Fei-Fei Li, for approximately $8.2 billion in an all-stock transaction. Following the acquisition, Li will serve as AMD's Executive Vice President and Chief Scientist, reporting to Lisa Su. The deal is expected to be completed by the end of 2026, pending regulatory approval.
World Labs, headquartered in San Francisco, develops models for generating, reconstructing, and simulating interactive 3D environments from text, images, and videos, as well as for robotic learning and simulation. Its first commercial product, Marble, is set to launch in 2025, generating walkable 3D worlds from prompts, targeting entertainment and robotic training. In September 2026, it will also launch the Atlas omnidirectional world model, along with the World API. Co-founders Justin Johnson and Ben Mildenhall will continue to lead the team, forming a cutting-edge research organization after the integration into AMD. The company previously acquired SceniX, which is related to simulation. Li's work on ImageNet and visual competitions has advanced computer vision; she advocates that general intelligence cannot rely solely on text but must understand space and physics.
Earlier this year, World Labs raised approximately $1 billion, with publicly disclosed investors including AMD, Autodesk, Emerson Collective, Fidelity, NVIDIA, and Sea. Its valuation reached around $1 billion within months of its establishment. Last year, both parties collaborated on training and inference optimization on AMD graphics cards, and Li appeared as a guest at CES this year. Li stated that the team has trained on AMD GPUs and aims to scale up, expand coverage, and get closer to hardware. Su emphasized that the next-generation computing platform must understand how models evolve, and World Labs' workloads will enter the chip roadmap. Both parties discussed an open ecosystem: hardware, software, platforms, and widely accessible open-weight models.
NVIDIA, in addition to selling chips, has engaged developers with open-world models like Cosmos. AMD previously disclosed models primarily focused on text and video, and this acquisition addresses the gap in spatial simulation, making it more likely for AMD card buyers to access world models optimized for their hardware. Robots lack real operational data, and synthetic scenes are written as training fuel for general humanoids and autonomous driving, such as Tesla and Figure. The transaction is stock-based; AMD's market capitalization was nearly $1 trillion on the day of the announcement, making dilution lighter compared to a cash acquisition, but it represents a bet on AMD's stock price for the sellers.
From a market mechanism perspective, this is a chip maker exchanging equity for models and talent, not a spot semiconductor order. The buyer, AMD, seeks workload definition rights and a chief scientist title; the seller's shareholders convert the next jump after approximately $1 billion into AMD stock. Funds will flow from AMD's equity into research staffing and GPU clusters. Beneficiaries will be robots and content pipelines needing synthetic 3D data; those under pressure will be independent laboratories still treating world models as software subscriptions, as well as NVIDIA's ecosystem, which must explain why it is both an investor in World Labs and a target. Regulatory approval is the only hard gate.
Additional data: World Labs claims that breakthroughs since its founding in 2024 have clarified spatial and physical issues. The team will continue model research and will not announce an immediate halt to Marble sales. No specific open-source timeline has been provided for the open model commitment.
Source: Public Information
ABAB AI Insight
Lisa Su is not just acquiring another chat model; she is transforming the input for chip design from benchmark testing to physical simulation workloads. Fei-Fei Li, coming from Stanford and ImageNet, stepping into the role of head of research at the company signifies AMD's public acknowledgment that the software stack can no longer just serve as a CUDA alternative. Last year's training collaboration was a trial period, appearing together at CES was a proposal, and the $8.2 billion stock deal is the marriage certificate. NVIDIA's presence on the funding list indicates that the world model is so hot that even competitors have written checks; after the acquisition, that check will turn into a conflict of interest that needs explanation. Marble will first create walkable 3D environments, while Atlas will focus on spatial intelligence, with the roadmap being that content entry supports data, which in turn feeds robots.
The capital path is entirely stock-based, binding the seller to AMD's stock price. The transition from $1 billion in financing to an $8.2 billion valuation reflects the pricing of spatial intelligence narratives by chip makers. Autodesk seeks 3D workflows, Fidelity wants exports, Sea aims for application distribution, and AMD wants to weld these demands onto its own silicon. Once synthetic data becomes standard for robot training, GPU hours will shift from text pre-training to continuous simulation, necessitating changes in caching, interconnects, and video memory in the roadmap. The open ecosystem commitment is aimed at countering CUDA lock-in, and whether it is fulfilled will determine if this premium is for platform or talent leasing.
In comparison, NVIDIA integrates simulation into its stack using Omniverse and Cosmos, while Google used research to feed back into infrastructure after acquiring DeepMind. The difference is that AMD, as a challenger, is acquiring a lab of "AI mother" level after the data bottleneck in robotics has been widely recognized. The industry phase is shifting from large language model arms races to spatial intelligence and embodied data. Whoever can provide both a generated world and train robots will define what the next generation of accelerators should look like.
Structural judgment indicates a reconstruction of the industry chain. The mechanism is that once model research, systems, and computing power must operate in the same building, the endgame for independent world model companies is to be absorbed by chip makers. Pricing power extends from selling cards to selling "worlds that can run on cards." Synthetic data reduces the cost of real data collection while moving the data factory into GPU clusters. Until regulatory approval, this remains a stock contract; after approval, developers will have an additional model rationale when selecting cards.
ABAB News · Cognitive Laws
Chip makers buy models to shape the next generation of silicon
Robots lack real data, making world models data factories
Competitors appearing on the shareholder list means the acquisition rewrites ecological boundaries