World Labs Acquired

World Labs: Fei-Fei Li's Spatial-Intelligence Unicorn, Acquired by AMD for $8.2B

Fei-Fei Li's spatial-intelligence company and a leader in Large World Models (Marble, RTFM, Atlas); agreed to be acquired by AMD for ~$8.2B in September 2026.

Executive Summary

World Labs is a spatial intelligence company founded in April 2024 in San Francisco by Fei-Fei Li, Stanford professor of computer science and member of the US National Academy of Engineering, along with Justin Johnson and Ben Mildenhall, a core inventor of NeRF. The company is small, with Revelio Labs estimating fewer than a hundred employees as of March 2026 and most reporting citing about 70, yet it is among the most watched companies in the world model field.

On September 28, 2026 AMD announced an all-stock acquisition of World Labs at a valuation of about 8.2 billion dollars, roughly 55 billion renminbi. The merger agreement was signed September 26, AMD acquires all equity, and closing is expected by the end of 2026 subject to regulatory approvals in multiple countries. It is AMD's second largest acquisition after Xilinx at about 50 billion dollars in 2022. After closing, Fei-Fei Li becomes AMD executive vice president and chief scientist reporting to chair and CEO Lisa Su, while the other two co-founders continue leading the research team and the lab retains its San Francisco office. AMD shares fell about 4 percent on the announcement.

The product timeline runs as follows. In its founding year the company released a large world model generating interactive 3D worlds from a single image. In November 2025 it launched Marble, the first commercial product, generating spatially consistent, editable and exportable 3D scenes from text, images, video or 360 panoramas on a free plus paid subscription. In July 2026 it acquired the robotics simulation startup SceniX, extending from generating 3D worlds to robot training environments. On September 1, 2026 it released Atlas, which the company calls the first natively multimodal world model pretrained from scratch on text, images, video and 3D data, supporting pixel-level camera control, 3D reconstruction, video camera switching and rapid conversion of real scenes into robot training environments.

On funding, September 2024 brought 230 million dollars at a valuation near one billion. February 2026 brought one billion dollars from investors including NVIDIA, AMD, Autodesk, a16z and Radical Ventures, with industry estimates around five billion. Seven months later the acquisition at 8.2 billion represented an increase above sixty percent.

In physical AI terms World Labs builds world models rather than simulation solvers or robots, operating at the model layer that lets AI understand, generate and simulate three-dimensional physical worlds. The value proposition is that robot and autonomous driving training needs large volumes of diverse interactive 3D environments while real-world collection is expensive and limited, and world models generate those environments on demand and support training policies inside generated worlds.

One framing point matters. This is a company with the highest technical reputation, zero commercial data and strategic-value pricing. It never disclosed revenue, has fewer than a hundred employees, and runs Marble on a freemium model with unknown users and revenue. The 8.2 billion valuation reflects the judgment that world models are next-generation AI infrastructure plus the scarcity of Li's team rather than current financials. Assessment must separate two layers: commercial value as an independent company, which is unknown, and strategic value to AMD, which is clear in using model research to define chips and systems.

Company Overview

Origins and Team

World Labs was founded in April 2024 and is headquartered in San Francisco. The founding team is a top-tier computer vision combination: Fei-Fei Li, driver of ImageNet, co-director of Stanford HAI and former head of Google Cloud AI; Justin Johnson, a researcher in computer vision and visual reasoning; and Ben Mildenhall, a core inventor of NeRF, the neural radiance field technique that changed 3D reconstruction.

The technical meaning is specific. ImageNet represents the large data plus deep learning paradigm, NeRF represents implicit neural representation of 3D scenes, and Johnson represents visual reasoning and generation. Together they map onto the three capabilities a world model needs: data scale, 3D representation and the chain from perception to reasoning.

From LWM to Marble to Atlas

In its founding year the company released a large world model generating interactive 3D virtual worlds from a single image, validating generative 3D while leaving limits in spatial consistency and editability.

Marble in November 2025 was the commercial product with multimodal input from text, images, video and 360 panoramas, precise control from 3D layouts, interactive editing of elements or whole scenes, expansion and combination of worlds, and export to multiple 2D and 3D formats for existing pipelines, priced as free plus paid subscription.

Atlas in September 2026 was the technical leap, pretrained from scratch and natively handling text, images, video and 3D data, with pixel-level camera control, 3D reconstruction, video camera switching and rapid conversion of real scenes into robot training environments. If native multimodality holds, Atlas addresses the alignment error inherent in generate-then-assemble approaches.

The SceniX Acquisition

The July 2026 acquisition of SceniX moved the company from generating visually appealing 3D worlds to generating environments usable for robot training, which requires physical attributes such as friction, mass and collision, task semantics such as graspability and traversability, and simulation interfaces to frameworks such as Isaac Sim and MuJoCo.

The acquisition shows awareness that visually realistic worlds alone cannot support robot training, which needs physically interactive and task-meaningful environments. That is precisely the boundary between world models and simulators, and where the company both competes and complements the NVIDIA Omniverse and Isaac ecosystem.

Technology

What World Models Solve

Understanding the technology requires distinguishing world models from generative models. Generative models learn data distributions and output pixels or tokens, while world models learn state transitions, predicting the next state given the current state and an action. That definition implies three technical elements: state representation, dynamics and controllability.

Three routes compete. Pixel-level video generation such as Sora-class models treats world modeling as video sequence generation with abundant internet video data and photorealistic output but no explicit 3D structure, poor physical consistency and no interactivity. 3D-native generation such as Marble and Atlas explicitly produces 3D representations through Gaussian splatting, neural radiance fields or meshes with interactivity, editability and exportability but scarce 3D training data and limits on quality and speed. Physical simulation such as Isaac Sim and MuJoCo guarantees correct dynamics with limited visual realism and scene diversity.

World Labs takes the second route and moves toward the third through SceniX. The bet is that robot training ultimately needs environments that are visually realistic, physically interactive and generatable on demand, and only the 3D-native route satisfies all three.

Choosing a 3D Representation

The core of a 3D-native world model is the scene representation, with four main options: voxels, point clouds, neural radiance fields and 3D Gaussian splatting.

NeRF represents density and color implicitly through an MLP with high rendering quality but slow training and rendering. 3D Gaussians use explicit Gaussian primitives with real-time rendering and easy editing by moving or deleting Gaussians, becoming the mainstream since 2023. Voxels and point clouds are more structured but limited in visual quality.

Mildenhall invented NeRF, and the company's apparent choice of 3D Gaussians, implied by Marble's real-time editing and export, shows a preference for efficiency over offline quality, since world models serve interaction and training where latency matters more than final render fidelity.

What Native Multimodality Means for Atlas

Atlas is described as pretrained from scratch and natively handling text, images, video and 3D data. Technically that means modalities share a representation space during pretraining rather than being separate encoders aligned afterward.

The assemble approach suffers alignment error between text descriptions and 3D geometry, and between video frames and 3D structure, with errors accumulating through alignment modules. Native multimodality processes tokens from different modalities in one Transformer so the model learns cross-modal correspondence itself. It is harder technically, requiring unified tokenization and large multimodal data, but delivers better cross-modal consistency.

Pixel-level camera control is another technical point, letting users set camera pose and motion precisely, which requires geometric understanding of 3D space rather than generating a clip that looks plausible from a text prompt.

Why AMD Is Buying

This is central to the deal. On the surface a chip company buying a model lab seems illogical, but AMD's logic is that next-generation compute requirements are set by model architecture, and understanding how models evolve is necessary to design the right chips.

World models impose different hardware demands than language models. Language models need large batches, long sequences and high memory bandwidth, while world models need real-time generation, 3D representation and heavy spatial queries, with higher demands on memory bandwidth, rendering pipelines and low-latency inference. Dedicated optimization in that direction, analogous to NVIDIA's software-hardware co-design in Omniverse and Isaac, requires a team that understands the models.

A second logic is ecosystem. NVIDIA's moat includes CUDA, Omniverse and Isaac, and AMD needs comparable software and model capability to challenge it. Buying World Labs acquires a voice on the model side.

Technical Shortfalls

First, physical accuracy. Generated worlds look realistic, but whether friction, mass and collision responses are accurate is unverified publicly, and policies trained in physically inaccurate environments learn wrong causal relations.

Second, scale and cost. Generating a high-quality 3D world costs far more than generating an image, and unless cost falls the economics of generating large training environments do not hold.

Third, missing evaluation standards. There is no accepted metric for how good a generated world is, with visual quality, geometric fidelity, physical consistency and task usability each needing different methods.

Fourth, zero commercial data. Marble has free and paid tiers without disclosed paying users, revenue or retention.

Product and Commercial Model

Marble Product Form

Marble is a 3D world generation tool for creators and developers taking text, images, video or 360 panoramas as input and outputting interactive, editable, exportable scenes, with precise control from 3D layouts, interactive editing, expansion and combination, and multi-format export.

Three user groups exist: creative professionals in games, film and advertising; developers needing large numbers of 3D environments; and researchers needing robot or autonomous driving training environments. Their willingness to pay and usage differ, with creatives paying per project, developers per API call and researchers by compute or subscription.

Integration With SceniX

SceniX brings robot simulation, and the integrated logic is to generate scenes with Marble and Atlas, make them interactive with SceniX physics and output to robot training frameworks. If that chain works the product upgrades from a 3D content tool to a robot training environment platform.

The latter carries higher commercial value, since robot companies will pay for high-quality, diverse and scalable training environments because real data collection is extremely expensive in robots, scenes and teleoperation hours. The threshold is also higher, requiring physical accuracy, task semantic labels and deep integration with training frameworks.

Business Model and Pricing

Marble uses freemium to acquire a creator community and monetize professionals. For a compute-intensive product like a world model, free tiers carry heavy cost because each generation consumes substantial GPU time.

An alternative is per-generation billing by scene count or compute, better matching the cost structure but less friendly to creators. Pricing details are undisclosed.

Product Direction After Acquisition

After closing, World Labs keeps its San Francisco office and the other two co-founders lead research while Li moves to AMD. Research therefore continues because model capability is what AMD needs, while commercialization priority may decline since AMD is buying research capability rather than subscription revenue.

For existing customers the roadmap may shift from standalone SaaS to solutions tied to AMD hardware. That is unwelcome for creators who do not need AMD chips, and welcome for robotics companies where training needs compute and integrated hardware and software lowers cost.

Competitive Landscape

Competitors in World Models

Three groups compete. Video generation models such as OpenAI Sora, Google Veo and Runway deliver high visual quality from abundant data but lack explicit 3D structure and physical interaction.

3D generation companies such as Luma AI, Tripo, Meshy and VAST generate individual objects or small scenes and commercialize as creator tools and game assets. They differ in scale, since World Labs builds explorable large scenes rather than single objects.

Simulation and robot training platforms such as NVIDIA Isaac Sim and Omniverse, MuJoCo and Genesis are physically accurate with mature ecosystems but weak at scene generation, requiring manual modeling or scanning. World Labs is complementary, generating scenes for import into simulation.

Relationship With NVIDIA

NVIDIA was an investor in the February 2026 round and is both the most important ecosystem partner and a potential competitor. Omniverse builds physically accurate virtual worlds and collaboration platforms while Isaac handles robot training, and World Labs generation could supply scene content to Omniverse while competing with Isaac if it builds its own training platform.

After the AMD acquisition the relationship becomes more delicate, since World Labs becomes an AMD asset and AMD is NVIDIA's direct competitor. NVIDIA gains a financial return as its stake is acquired but loses an ecosystem partner.

Comparison With Chinese Companies

Chinese world model and 3D generation companies include VAST with Tripo, Yingmu Technology and in-house 3D generation teams at large platforms. They commercialize more toward asset production for games, e-commerce and film, smaller in scale with clearer paths. World Labs positions more broadly as spatial intelligence infrastructure with earlier commercialization.

The Trend of Chip Companies Acquiring Upstream

AMD's purchase is not isolated. Chip and cloud companies have been extending upstream toward models, with Google owning DeepMind, Microsoft partnering with and building models, and Amazon investing in Anthropic. The logic is that models define hardware requirements, so owning models means defining next-generation chips.

The effect on independent model companies is two-sided: acquisition offers exit and resources while independent survival narrows to being acquired or finding gaps between giants.

Financing and Capital Structure

Funding Trajectory

September 2024 brought 230 million dollars at a valuation near one billion. February 2026 brought one billion dollars from NVIDIA, AMD, Autodesk, a16z and Radical Ventures with industry estimates near five billion. September 2026 brought an all-stock acquisition at 8.2 billion.

The trajectory is notable for rapid revaluation, eight times in two years, supported by three events: Marble commercialization in November 2025, Atlas in September 2026 and the emerging consensus that world models are next-generation AI infrastructure.

Deal Structure

The all-stock transaction at about 8.2 billion was signed September 26, announced September 28 and is expected to close by the end of 2026 subject to multi-country regulatory review.

All-stock means World Labs shareholders including NVIDIA, AMD itself, Autodesk and a16z receive AMD shares rather than cash. That conserves AMD cash and aligns shareholders, while shareholders keep exposure to an AI compute company with upside if AMD appreciates. The risk is that AMD share price volatility directly affects deal value, and the 4 percent decline on announcement reduced value by about 330 million dollars.

Valuation Methodology Problems

Eight billion dollars for fewer than a hundred people with no disclosed revenue cannot be explained by price to sales, earnings multiples or discounted cash flow. Three explanations are plausible: strategic pricing of team and technology including Li, a NeRF inventor and a 3D generation stack; option pricing if world models become robot training infrastructure; and defensive acquisition to keep the asset from NVIDIA or Google.

All three hold, and none is a financial investment logic but a strategic one, which explains why the buyer is AMD, which needs model capability to define chips, rather than a financial investor.

DISCREPANCIES

1. Two employee figures. Most reporting cites about 70 while Revelio Labs estimates fewer than a hundred as of March 2026. The two are consistent, with 70 inside the range, but differ in precision. This report uses fewer than a hundred, around 70. Source tier: MEDIA and third-party labor data, both estimates.

2. Three valuation points. About one billion in September 2024 from the funding disclosure, about five billion estimated by industry in February 2026 without official confirmation, and about 8.2 billion in September 2026 as acquisition consideration. The three reflect different dates and natures, funding valuation versus acquisition price, and are not directly comparable. This report lists them chronologically. Source tier: OFFICIAL (acquisition price and 2024 round) and MEDIA (February 2026 estimate).

3. Revenue never disclosed. The company has never published revenue, paying users or subscription conversion, and Marble runs freemium with unknown commercial scale. Some interpretations infer revenue scale from the existence of paid tiers. This report labels revenue unknown. Source tier: information absent.

4. Founding date wording. Official and most reporting cite April 2024 while some sources say early 2024. This report uses April 2024. Source tier: OFFICIAL crossed with MEDIA.

5. Exclusivity of the first multimodal world model claim. The company calls Atlas the first multimodal world model, which depends on how multimodal and world model are defined, whether video generation counts as a world model and whether 3D generation counts as multimodal, and cannot be independently verified. This report retains the claim labeled as company framing. Source tier: company marketing, UNVERIFIED.

6. Uncertainty on closing. The transaction still requires multi-country regulatory approval with closing expected by the end of 2026, while some coverage describes the acquisition as completed. This report uses agreement signed, closing pending. Source tier: OFFICIAL (AMD announcement).

7. Wording of Li's role. After closing she becomes AMD executive vice president and chief scientist, AMD's first chief scientist, reporting to Lisa Su, while some coverage simplifies it to joining AMD. This report uses the full title. Source tier: OFFICIAL (AMD announcement).

8. NVIDIA's dual role as investor. NVIDIA invested in the February 2026 round while competing with AMD, and after the acquisition its stake converts into AMD shares or is handled per deal terms, with some interpretations saying NVIDIA invested in a competitor. This report notes the dual role without inferring intent. Source tier: OFFICIAL (investment fact) distinguished from inference.

Risks and Diligence Gaps

First, unverified physical accuracy. Generated worlds look realistic but friction, mass and collision accuracy is unverified publicly, and policies trained in inaccurate environments learn wrong causality.

Second, zero commercial data. No revenue, paying users or retention, with the 8.2 billion valuation resting entirely on strategic value.

Third, generation cost. High-quality 3D generation costs far more than image generation, and the economics of scaling for robot training are unverified.

Fourth, integration risk. Keeping the research team and moving Li to AMD executive ranks has unknown long-term effects, and research culture may not blend with chip company engineering culture.

Fifth, regulatory approval. Multi-country review of a large technology acquisition carries uncertainty in the current environment.

Sixth, relationship with the NVIDIA ecosystem. As an AMD asset, World Labs shifts from partner to competitor against Omniverse and Isaac and may lose ecosystem support.

Seventh, technology route risk. If video generation overtakes 3D-native approaches on quality and cost, World Labs' advantage erodes.

Composite Assessment

Technology scores 8.5. The combination of Li, Johnson and Mildenhall is top-tier in computer vision and 3D generation, Marble's multimodal input, precise control, interactive editing and multi-format export reach commercial quality, Atlas's native multimodality and pixel-level camera control are a leap, and SceniX adds physical simulation. Deductions for unverified physical accuracy and unknown generation cost and scale.

Business scores 5.0. Marble has a clear product form and freemium model with well-defined user groups in creators, developers and robotics, and SceniX integration points at the high-value robot training scenario. Deductions for completely undisclosed revenue and paying users, generation cost potentially capping margins and possibly lower commercialization priority after acquisition.

Finance scores 7.0. The funding trajectory is strong from 230 million to one billion to an 8.2 billion acquisition, investors include NVIDIA, AMD, Autodesk and a16z, and the exit path is clear through an all-stock AMD acquisition. Deductions for zero revenue, strategic-based valuation and all-stock exposure to AMD share price.

Product scores 7.5. Marble's completeness in multimodality, editing and export leads peers, Atlas's full-modality capability is differentiated and SceniX completes the training environment. Deductions for undisclosed physical attributes, task semantics and depth of training framework integration, plus missing pricing detail.

Market scores 8.0. World models are seen as next-generation AI infrastructure, demand for synthetic environments in robot training and autonomous driving is real, and the AMD acquisition validates strategic value. Deductions for an early market, competition from video generation and NVIDIA's ecosystem lead.

Overall 7.2. This is the highest technical reputation with the least commercial data in this series: it went from zero to an 8.2 billion valuation in two years on a top team, a clear technical route in 3D-native world models and a correct judgment that robot training needs generatable, interactive 3D environments. Yet it never disclosed revenue, physical accuracy is unverified and generation cost is unknown. The judgment on physical AI is that world models are indeed a key layer connecting perception and action and providing training grounds for robots and autonomous driving, and that AMD's acquisition shows compute vendors recognize that strategic value. Moving from technical demonstration to production-grade training infrastructure still requires proof of physical accuracy, cost control and scalability, and none of that evidence exists yet.

References

  1. AMD acquisition announcement, September 28, 2026: all-stock acquisition of World Labs at about 8.2 billion dollars, merger agreement signed September 26, AMD acquiring all equity, closing expected by the end of 2026 subject to multi-country regulatory approval, AMD's second largest acquisition after Xilinx at about 50 billion dollars in 2022, Fei-Fei Li becoming AMD executive vice president and chief scientist reporting to Lisa Su, the other two co-founders leading the research team and the San Francisco office retained
  2. World Labs website, accessed 2026: positioning as a spatial intelligence company building frontier models that perceive, generate, reason and interact with virtual and physical worlds; Marble as the first product generating spatially consistent, high-fidelity, persistent 3D worlds with text, image, video and 360 panorama inputs, 3D layout control, interactive editing, expansion and combination and multi-format export; research page listing the September 28, 2026 joining of AMD and a July 28, 2026 discussion of spatial intelligence including the SceniX acquisition
  3. Product timeline from public reporting: a large world model in the founding year generating interactive 3D virtual worlds from a single image; Marble launched in November 2025 as the first commercial world model supporting editable and exportable 3D scenes from text, images or video on free plus paid subscription; SceniX acquisition in July 2026; Atlas released September 1, 2026
  4. Atlas capabilities from public reporting: pretrained from scratch and natively handling text, images, video and 3D data as a full-modality world model; pixel-level camera control, 3D reconstruction and video camera switching; rapid conversion of real scenes into robot training environments; applications in robot training, autonomous driving, game development, film effects and scientific research
  5. Company facts from public reporting: founded April 2024 in San Francisco; headcount about 70 with Revelio Labs estimating fewer than a hundred as of March 2026; founders including Fei-Fei Li as Stanford professor of computer science, member of the US National Academy of Engineering, co-director of Stanford HAI and head of Google Cloud AI from 2017 to 2018, along with Justin Johnson and Ben Mildenhall as a core NeRF inventor
  6. Financing data from public reporting: 230 million dollars in September 2024 at a valuation near one billion; one billion dollars in February 2026 from investors including NVIDIA, AMD, Autodesk, a16z and Radical Ventures with industry estimates near five billion; acquisition at about 8.2 billion in September 2026, an increase above sixty percent from February 2026
  7. Deal details from public reporting: AMD's current report on Form 8-K disclosing the merger agreement signed September 26; AMD shares falling about 4 percent on announcement; Li becoming AMD's first chief scientist
  8. Lisa Su statement in the AMD announcement: building computing platforms for next-generation AI requires deep understanding of how models evolve; Li and the World Labs team bring research leadership and model expertise; the combination can develop hardware, software and systems driving next-generation AI and strengthen the open AI ecosystem
  9. Fei-Fei Li statement in the AMD announcement: advancing next-generation AI requires close collaboration across model research, systems and compute; joining AMD provides resources and engineering depth to accelerate research and help define the infrastructure needed for a new AI era
  10. Technical background from public sources: world models learn state transitions rather than data distributions with three elements of state representation, dynamics and controllability; three routes are pixel-level video generation with abundant data but poor physical consistency, 3D-native generation that is interactive and editable but data scarce, and physical simulation that is accurate but limited in realism; 3D representations include voxels, point clouds, NeRF and 3D Gaussian splatting, the last becoming mainstream after 2023 for fast rendering and easy editing
  11. Industry background from public sources: NeRF represents density and color implicitly with an MLP for high quality but slow training and rendering; 3D Gaussians use explicit primitives for real-time rendering and easy editing; world models for robot training need physical attributes such as friction, mass and collision plus task semantics such as graspability and traversability, and must interface with training frameworks such as Isaac Sim and MuJoCo
  12. Competitive landscape from public sources: video generation includes OpenAI Sora, Google Veo and Runway; 3D generation companies include Luma AI, Tripo, Meshy and VAST; simulation and robot training platforms include NVIDIA Isaac Sim and Omniverse, MuJoCo and Genesis; chip and cloud companies extending upstream include Google with DeepMind, Microsoft with OpenAI and Amazon with Anthropic
This report is compiled from public sources for informational purposes only and is not investment advice. The company is private and pending acquisition, revenue and user data were never disclosed, valuation and deal terms follow AMD announcements and public reporting, and technical claims are company-reported without third-party verification.

Related reports66

This report is compiled from public sources. It is for reference only and does not constitute investment advice. Data is current as of the publication date; please refer to the company’s official filings.