World Labs builds 3D spaces where Sora rendered video

World Labs launched Atlas on September 1 to build 3D spaces rather than brute-force video frames. The release followed an abrupt retreat across the sector. According to geekpark, OpenAI shut down its video-generation model Sora in March. Atlas operates instead as a multimodal world model pre-trained from scratch. According to ifanr, World Labs is developing these spatial systems to reconstruct and simulate interactive 3D environments.

New visual perspectives emerge directly from sparse 2D image inputs. According to woshipm, Atlas calculates novel viewpoints and performs 3D reconstruction from a small set of photos. It does not guess frames.

The underlying architecture couples generative modeling with multi-view geometry. According to ifanr, Fei-Fei Li stated that Atlas outperformed certain specialized models and made progress on sparse reconstruction in computer vision. Treating viewpoints as predictable spatial coordinates shifts generative technology away from ephemeral clips toward stable, reproducible digital space.

Abandoning video diffusion for sparse multi-view tokens

Video diffusion treated physical reality as an unanchored sequence of hallucinated pixels. According to geekpark, OpenAI shut down its video-generation model Sora in March, highlighting the structural limits of generative rendering without an underlying geometric foundation. When diffusion models lack coordinate systems, moving camera trajectories routinely warp rigid structures or erase occluded items without warning. Scenes simply drift.

On September 1, World Labs launched Atlas, a multimodal world model pre-trained from scratch, according to geekpark. The system frames scene generation as a geometric task rather than a frame prediction problem.

Atlas anchors visual predictions directly to multi-view geometry. According to woshipm, the model predicts new visual perspectives and executes sparse 3D reconstruction from a small set of 2D images. As ifanr reported, Atlas combines generative modeling and multi-view geometry to predict novel viewpoints from 2D images. Fei-Fei Li stated that Atlas outperformed certain specialized models and made progress on sparse reconstruction in computer vision, according to ifanr. The geometry holds. The architecture predicts consistent spatial coordinates instead of guessing consecutive video frames.

This shift transforms interactive world building. According to woshipm, World Labs focuses on spatial intelligence and world models, using images and videos to reconstruct interactive 3D scenes that simulate physical rules. Furthermore, ifanr reported that World Labs develops spatial intelligence models that reconstruct and simulate interactive 3D environments from text prompts alongside image and video inputs. Viewpoints behave as predictable points in space rather than vanishing pixel streams. Rigid geometry replaces diffusion noise.

How Marble and SceniX plug into simulation engines

World Labs occupies an unusual position in spatial computing. As woshipm notes, the company has been founded for less than three years and lacks consumer-facing products. It remains in an early stage of commercialization. Yet geekpark reported that it raised $230 million at a $1 billion valuation when emerging from stealth mode in 2024. That capital supports an industrial focus rather than consumer video tools. Fei-Fei Li co-founded World Labs in early 2024 alongside Ben Mildenhall and Justin Johnson with a focus on spatial intelligence, according to ifanr.

Their approach avoids pixel generation. Instead, approximately 70 researchers across data, computer vision, large model training, system architecture, and computer graphics build environments that machines can parse and measure, as ifanr reported.

Simulation engines cannot operate on hallucinations. They require coordinate fields and collision boundaries. To supply that structure, World Labs acquired SceniX in July to integrate robotics simulation capabilities into its research, as ifanr reported. According to woshipm, SceniX specialized in robotics simulation in virtual 3D environments. The acquisition links spatial reconstruction directly to robotics training, letting autonomous agents calculate paths and test mechanics.

This pipeline materialized with Marble, which geekpark recorded as the company's first product release in November 2025. Marble converts text, images, videos, or panoramic photos into editable 3D scenes. Rather than outputting flat video files that lock camera angles into place, Marble creates explicit spaces ready for physics engines. Simulation platforms can ingest these structural assets directly, turning raw inputs into interactive terrain.

Balancing open weights against proprietary commercial APIs

Accessing spatial models requires balancing gated commercial interfaces against downloadable weights. When AMD reached an agreement to acquire Fei-Fei Li's startup World Labs for approximately $8.2 billion, the purchase highlighted how hardware vendors intend to manage that boundary. The agreement arrived 25 days after Nvidia moved to acquire Hugging Face. According to geekpark, AMD and Salesforce had both expressed acquisition interest in Hugging Face prior to that deal.

Developer access hinges directly on how vendors package these spatial architectures. Open releases cultivate baseline community experimentation, while proprietary hosted services protect enterprise margins.

Lisa Su clarified that AMD will offer both open models and proprietary models in the future. As reported by ifanr, Su explained that AMD acquired World Labs because developing a deeper end-to-end perspective allows the firm to construct superior systems. Hardware and software converge under one roof. This structure allows AMD to supply open research artifacts while reserving specialized commercial endpoints for private deployment.

That distribution strategy draws directly from Fei-Fei Li's historical work in visual intelligence. According to woshipm, Li built the ImageNet dataset, creating a foundational resource for deep learning and visual AI. The project reached its official release in 2009 after roughly two and a half years of assembly. Organizing it involved nearly 50,000 annotators across 167 countries, who filtered down 15 million labeled images from a pool of nearly 1 billion candidates. Broad adoption followed that public release. Translating spatial generation into common workflows depends on repeating that balance between open resources and commercial tools.

Evaluating Spatial World Models vs. Brute-Force Generative Video

• Generative video pipelines relying on brute-force frame generation following the shutdown of tools like OpenAI's Sora in March Shift development from unconstrained 2D video diffusion toward sparse 3D reconstruction architectures like World Labs' Atlas (launched on September 1), which combines generative modeling and multi-view geometry to infer novel viewpoints from a small set of 2D images.
• Pipelines requiring editable, interactive 3D environments derived from multi-format inputs (text, images, video, or panoramic photos) Deploy specialized spatial generation systems such as World Labs' Marble (released in November 2025), which is explicitly built to convert multimodal media into editable 3D scenes rather than flat video sequences.
• Embodied AI and robotics simulation workflows requiring virtual environments governed by physical constraints Adopt robotics-focused simulation platforms such as SceniX (acquired by World Labs in July) to generate interactive 3D environments that simulate physical rules rather than purely visual approximations.
• Long-term compute stack and accelerator infrastructure planning amid high Nvidia market concentration (IDC estimates Nvidia at approximately 81% and AMD at 13%) Track spatial model optimization on alternative accelerator architectures; AMD and World Labs completed deep training and inference optimization on AMD GPUs across 2025 ahead of AMD's planned $8.2 billion acquisition, alongside enterprise rollouts such as Anthropic deploying up to 2 gigawatts of AMD MI450 GPUs starting in the first half of 2027.

Decoupling spatial pipelines from CUDA ahead of the 2026 close

Nvidia remains deeply entrenched across machine learning pipelines. According to geekpark, IDC estimates that Nvidia controls approximately 81% of the AI accelerator market, leaving AMD with approximately 13%. Nvidia reinforced its position by acquiring assets from Groq for $20 billion over the past year and agreeing to acquire Hugging Face for approximately $13 billion on September 3. Chasing text generation offers diminishing returns. According to woshipm's author, the next phase of competition belongs to the physical world rather than virtual text and images.

Without dedicated silicon, spatial software hits hard boundaries. Fei-Fei Li stated that without dedicated hardware support, artificial intelligence remains limited in efficiency and scale, leaving it trapped in the digital world.

AMD's response bypasses text models to anchor spatial runtimes on its own silicon. The company signed an agreement to acquire World Labs in an all-stock transaction valued at approximately $8.2 billion. According to ifanr, the acquisition is expected to close before the end of 2026, subject to regulatory approvals. The deal marks AMD's second-largest transaction behind buying Xilinx for approximately $50 billion in 2022. Technical collaboration on AMD GPUs began in 2025. Upon closing, Fei-Fei Li will serve as Executive Vice President and Chief Scientist reporting directly to CEO Lisa Su, while co-founders Justin Johnson and Ben Mildenhall continue advancing model research.

As woshipm's author argues, acquiring World Labs allows researchers to participate in defining future chip architectures rather than forcing models to adapt to pre-existing hardware. AMD possesses the capital to fund that pivot. Its data center revenue rose 107% year-over-year to $6.7 billion in the second quarter, while its market capitalization surpassed $1 trillion on September 21 after rising approximately 185% year-to-date.

Strategic Divergence in AI Hardware: AMD vs. Nvidia Market Share, M&A, and Computing Architecture

World Labs focuses on spatial intelligence, using images and videos to reconstruct interactive 3D scenes that simulate physical rules. Founded less than three years ago, it lacks consumer-facing products. Yet after joining a $1 billion financing round in February, AMD committed to an $8.2 billion acquisition paid entirely in stock without a cash payout. The woshipm author argues that this allows researchers to help define future chip architectures rather than adapting spatial models to pre-existing hardware.

Fei-Fei Li, creator of the ImageNet dataset, will serve as Executive Vice President and Chief Scientist reporting directly to AMD CEO Lisa Su. The woshipm author argues that upcoming AI competition will unfold across the physical world.

For readers outside China

• Availability: World Labs released its multimodal world model Atlas on September 1 and its editable 3D scene product Marble in November 2025. However, the source material notes that World Labs remains in an early stage of commercialization and lacks consumer-facing products. AMD CEO Lisa Su stated that AMD will offer both open models and proprietary models in the future following the acquisition, which is expected to close before the end of 2026 subject to regulatory approvals. Specific international platform access and geographical availability are not disclosed in sources.
• Pricing: End-user and API pricing for World Labs products (including Atlas and Marble) are not disclosed in sources. Sourced financial milestones indicate World Labs raised $230 million at a $1 billion valuation in 2024, completed a $1 billion financing round in February, held a $5.4 billion valuation in its prior financing round, and entered an all-stock acquisition agreement with AMD valued at approximately $8.2 billion.
• Closest Western equivalents: OpenAI Sora; Hugging Face
• Data residency: Data residency, server hosting regions, and user data governance policies are not disclosed in sources. The operational integration of World Labs into AMD remains subject to regulatory review and closing conditions through late 2026.

Sources

• woshipm AMD拟82亿美元收购李飞飞的公司：世界模型到底值钱在哪？ https://woshipm.com/ai/6473081.html
• geekpark AMD 82 亿美元收购 World Labs，买的不只是世界模型 https://geekpark.net/news/372000
• woshipm AMD 82 亿美元收购World Labs，李飞飞写下：去探寻一个更新的世界 https://woshipm.com/share/6471705.html
• ifanr 苏姿丰550亿元买下李飞飞世界模型，AI 圈两位女王联手了 https://ifanr.com/1682291