On September 28, 2026, AMD entered into a definitive agreement to acquire World Labs in an all-stock transaction valued at approximately $8.2 billion. The companies expect the transaction to close by the end of 2026, subject to regulatory approvals and customary closing conditions.

If the transaction closes, World Labs is expected to continue its AI model research within AMD, and co-founder and CEO Fei-Fei Li will become AMD executive vice president and chief scientist, reporting to AMD CEO Lisa Su. The agreement follows AMD’s participation in World Labs’ early 2026 financing round and a technical partnership between the companies around model training and inference optimization on AMD GPUs.

World Labs last raised $1 billion in new funding on February 18, 2026 from investors including AMD, Autodesk, Emerson Collective, Fidelity Management & Research Company, NVIDIA, and Sea. Autodesk said that its strategic investment accounted for $200 million of the round and gave it an advisory role.

World Labs had previously announced $230 million of funding when it emerged from stealth in September 2024, bringing total disclosed funding to approximately $1.23 billion.

Media reports in January 2026 indicated that World Labs was discussing the round at a valuation of about $5 billion.

World Labs builds Marble, a web application that converts incomplete descriptions of a place into persistent, explorable 3D environments. Users can provide a text prompt, photograph, image set, video, 360-degree panorama, or rough 3D blockout. Marble then infers depth, geometry, lighting, and unseen portions of the scene to produce a navigable world.

Unlike video-generation systems that produce fixed camera sequences, Marble builds the spatial stage behind the images. It represents scenes primarily as 3D Gaussian splats, collections of spatially positioned, colored, semitransparent points that render a photorealistic environment. Users can stop, turn around, revisit locations, or choose new camera paths without regenerating the output.

Workflows can start with a text description for ideation, uploaded concept art, or a phone panorama of a physical space. For more control, Marble's Chisel mode lets users construct rough blockouts from boxes, planes, and imported assets, separating spatial structure from visual style. For example, a user can lay out a rectangular room with a central staircase and prompt Marble to render it as a weathered Art Deco hotel lobby, using the blockout to constrain arrangement and the prompt to control appearance.

Users can edit generated worlds, expand them beyond their original boundaries, and combine them into larger navigable environments. Outputs include Gaussian splats (.PLY), collider meshes (.GLB), high-resolution textured meshes, 360-degree panoramas, and rendered videos for Unreal Engine, Blender, Three.js web applications, NVIDIA Isaac Sim, MuJoCo, and other production and simulation tools. World Labs' open-source Spark rendering library enables browser-based splat experiences across desktop, mobile, and VR.

Launched in January 2026, the World API exposes Marble as embeddable infrastructure. Developers can submit inputs programmatically and receive navigable 3D environments for game-environment generation, architectural walkthroughs, interactive film experiences, and synthetic robotics environments. Early integrations include Preview for filmmaking, Fenestra and xFigura for architecture, Interior AI for room visualization, and Escape.ai for converting video into navigable spaces.

Atlas entered early access on September 1, 2026, as World Labs' next-generation multimodal world model. It processes text, images, camera poses, depth maps, and video in a shared spatial context, with camera-controlled video up to one minute at 1440p, multi-image scene reconstruction, and explicit 3D output. Atlas accepts explicit camera geometry rather than interpreting text commands like most video models, giving filmmakers and simulation systems more deterministic control. It is designed for future Marble versions and additional products but remains in selective access as of September 2026.

The acquisition would connect World Labs’ model development more directly with AMD’s compute roadmap. AMD said World Labs’ work across reasoning, robotics, simulation, and physical AI will help shape future hardware, software, and systems for emerging AI workloads, while World Labs described tighter hardware integration as a way to improve the efficiency and scale of model training and inference.

World Labs sells spatial-intelligence capabilities through three channels: self-service Marble subscriptions for individual creators and small teams, the World API for developers embedding world generation into their products, and custom enterprise arrangements for studios, design firms, and robotics organizations. The result is a B2C, B2B, and B2B2C model.

Monetization follows a credit-based subscription model. Each paid tier includes a monthly credit allowance for compute-intensive actions. World generation costs 1,500 credits, with additional charges for expansion, high-quality mesh export, and per-second video enhancement. The structure links revenue to underlying GPU costs, while top-up purchases monetize production bursts without requiring permanent upgrades. API credits are sold separately and generally expire one year after purchase.

The cost structure is compute-intensive, unlike software with nearly costless incremental usage. Each generation consumes accelerator time, and larger, higher-resolution, or more geometrically detailed outputs cost more to produce. Near-term gross margins are likely lower and more variable than those of mature seat-based software. Potential margin improvements include model distillation, reduced diffusion steps, KV caching, batch scheduling, and specialized serving infrastructure. Atlas's transformer architecture was explicitly designed to use serving techniques developed for large language models.

The go-to-market combines bottom-up creator adoption with ecosystem partnerships. The free tier drives discovery, paid tiers monetize production use, and the API turns third-party applications into distribution channels without directing their users to Marble. Case studies across gaming, VFX, architecture, and robotics provide technical sales collateral in an emerging category where interoperability can matter more than conventional feature advertising. Through the Autodesk partnership, embedding Marble or Atlas capabilities in Autodesk's architecture, manufacturing, and media products would give World Labs access to established professional workflows without replacing the incumbent application layer.

World Labs operates at the intersection of generative 3D creation, real-time interactive worlds, and physical-AI simulation. Its bet is that persistent, explicitly exportable 3D representations will be more valuable to production and simulation workflows than transient interactive video, while competitors are narrowing the distinction between video models and world models from both directions.

Runway is the most direct broad-spectrum rival. Its GWM Worlds 2, released September 3, 2026, generates continuous 720p video at 24 fps with synchronized audio. It supports first- and third-person navigation, text-directed events, independently controlled subjects and cameras, and multiplayer roles with no preset session duration. Runway also distributes creative tools through integrations with Adobe Premiere Pro, After Effects, and DaVinci Resolve, and has commercialized a robotics product offering policy inference, photorealistic evaluation, and synthetic-data augmentation.

Runway's worlds remain video-native. Its documentation acknowledges imperfect long-term memory and geometry degradation under difficult camera motion. Marble and Atlas instead produce reusable spatial representations that can be exported, edited, and integrated into engines and simulators. These capabilities matter for customers that need deterministic camera placement, collision meshes, or asset-level control.

Google DeepMind's Genie 3 produces 720p interactive environments at 24 fps and is available to U.S. Google AI Ultra subscribers. Google can subsidize world-model access within a broader AI subscription and integrate it with Gemini, Veo, YouTube, and robotics research. Odyssey competes on immediacy and audiovisual richness by streaming continuously generated interactive video with synchronized audio through Starchild-1. Decart combines proprietary low-latency inference infrastructure with Oasis for physical-AI simulation, vertically integrating the serving stack to optimize real-time deployment economics.

Meshy and Tripo AI compete in individual 3D asset generation rather than scene-scale environments. Meshy offers text- and image-to-3D with plugins for Unity, Blender, and Maya, while Tripo adds retopology, rigging, and animation in a vertically integrated asset pipeline. For many game and visualization teams, assembling inexpensive generated assets in Unreal or Unity may offer more control than adopting a generated-world representation.

Luma AI competes through its Genie 3D model and its work on controllable, scene-consistent world-model video. Stability AI spans image, video, audio, and 3D generation, including artist-directed environment tools developed with Electronic Arts. Adobe aggregates its own models with models from Google, OpenAI, Runway, Kling, Luma, and others in one creative environment. Its Scene to Image feature shows how incumbent suites can absorb spatial controls without building full world models.

Traditional game engines remain indirect substitutes. Unreal Engine's procedural content generation framework, World Partition, Nanite, and established asset pipelines support production-grade open environments with control, determinism, physics, and performance optimization that generated Gaussian-splat worlds cannot yet match.

NVIDIA is both an investor in World Labs and a potential platform competitor. Cosmos provides open world foundation models, reasoning models, data-processing tools, and deployment software for physical AI, with deep integration into Omniverse and Isaac Sim. NVIDIA controls the GPU, systems software, model, simulation, and deployment stack, while its open models encourage self-hosting. This creates a structural threat if world generation becomes a standard feature rather than a stand-alone platform.

Meta's V-JEPA 2 challenges the assumption that useful spatial intelligence requires high-fidelity generative worlds. Trained on over one million hours of video, V-JEPA operates in latent space for efficient prediction and planning, with zero-shot robot planning after action-conditioned post-training on just 62 hours of robot data. For planning tasks where visual fidelity is secondary, a cheaper latent model may outperform a photorealistic simulator on cost and latency.

Roblox represents a separate competitive vector, with native 3D and 4D creation tools built directly into a platform with hundreds of millions of users. Its AI-assisted world-building approach is coupled with distribution and monetization in ways that stand-alone generation tools cannot easily replicate.

World Labs' current commercial wedge is creative and professional 3D-world creation. Its spatial-intelligence platform could also extend into robotics simulation, enterprise workflows, and new spatial-computing form factors.

The July 2026 acquisition of SceniX extends World Labs from visual scene generation into the broader physical-AI stack. SceniX's real-to-sim-to-real engine reconstructs real robot tasks, creates controllable variations, trains policies, and tests whether simulated performance predicts real-world outcomes.

Potential recurring-revenue products include simulation software for robotics labs, synthetic training datasets, policy evaluation services, and libraries of reusable environments. Demand stems from the cost and scarcity of embodied-AI data: physical data collection consumes robot time, requires manual resets, and cannot systematically cover all relevant environments and failure modes. Atlas extends this use case by reconstructing spaces from ordinary videos and modeling interactions involving rigid, articulated, and deformable objects.

The Autodesk partnership provides a distribution path into professional design workflows. Autodesk has established tools across architecture, engineering, construction, manufacturing, and media entertainment. If World Labs' models become embedded in products such as Maya, 3ds Max, Fusion, or Revit, the company could gain large-scale distribution without replacing the incumbent application layer.

Other expansion paths include game studios through Unreal Engine and Unity integrations, film and virtual production through LED-stage workflows, and architecture through partnerships with firms such as Fenestra and xFigura. Each vertical has different accuracy, compliance, and procurement requirements, while sharing the same generation and API infrastructure. The $8B–$10B 3D asset market is the near-term addressable opportunity, followed by larger markets in games, VFX, architecture, product design, and digital twins.

Spark 2.0, World Labs' open-source web rendering library, streams large Gaussian-splat environments through browsers using WebGL2 across desktop, mobile, and VR devices. This could make World Labs a content supplier for spatial-computing platforms such as Quest and Apple Vision Pro without requiring specialized desktop software.

The World API makes generation programmable, allowing AI agents and applications to create environments on demand. As AI agents require environments in which to perceive, act, train, and test, world generation could become a recurring compute workload rather than a one-time content-creation task. Real estate, retail, insurance, tourism, and healthcare are adjacent markets that could adopt spatial generation if capture-to-reconstruction workflows become sufficiently streamlined and accurate.

Compute-intensive margins: World generation, mesh conversion, video enhancement, and simulation are GPU-intensive operations with nontrivial marginal costs, so World Labs' gross margins will structurally resemble those of a compute-heavy infrastructure business rather than nearly costless SaaS, and the company may struggle to achieve the margin profile implied by its $5B valuation if customers prove highly price-sensitive while generation remains expensive, especially as open models from NVIDIA Cosmos and Meta V-JEPA compress model-level pricing.

Visual plausibility versus physical accuracy: Marble and Atlas can produce scenes that look physically plausible without a causal understanding of forces, gravity, or structural physics, while current output is primarily static rather than fully dynamic and physically simulated, limiting the company's ability to serve engineering-grade digital twins, safety-critical robotics validation, and precision AEC applications unless it closes the gap between visual coherence and structured, semantic, interactive, and physically reliable worlds.

Platform capture by incumbents: NVIDIA can vertically integrate world-model functionality into Cosmos, Omniverse, and Isaac Sim while controlling the underlying GPU infrastructure; Autodesk can absorb spatial generation as a feature inside its professional tools; Adobe can aggregate competing models into a single creative workflow; and game engines can add AI-assisted procedural generation natively, potentially reducing World Labs from a platform supplier to a replaceable model endpoint if its explicit 3D outputs, editing controls, and workflow integrations do not remain materially differentiated.