Companies & labspartial
Evidence standard for this page: product claims cite the companies’ own announcements; funding figures are attributed press reports (TechCrunch, Bloomberg, Forbes, CNBC), not independently verified — valuations especially are often company-leaked and unconfirmed. Dates are as of 2026-07-18.
The money
Forbes (Josipa Majic Predin, June 30, 2026): $3B+ of VC went into world-model startups in the first half of 2026 alone — AMI Labs $1.03B, World Labs $1B, Odyssey $310M, Decart $300M — alongside Wayve’s $1.2B Series D at a reported $8.6B (CNBC, Feb 2026). NVIDIA has backed World Labs, AMI, Decart, and Odyssey. AMI’s own CEO offers the caution: “In six months, every company will call itself a world model to raise funding.”
World Labs (Fei-Fei Li) — the spatial-intelligence flagship
Out of stealth Sep 2024 with $230M (a16z, NEA, Radical Ventures; TechCrunch); $1B Series B Feb 2026 anchored by Autodesk’s $200M, at a reported ~$5B+ valuation (Bloomberg; company declined to confirm). Total raised ≈ $1.23B. The shipping cadence is the real story:
- “Generating Worlds” (Dec 2024) — navigable 3D worlds from one image.
- RTFM, “Real-Time Frame Model” (Oct 2025) — an autoregressive diffusion transformer with no explicit 3D representation: a learned renderer using posed frames as spatial memory, at interactive frame rates on a single H100. A notable counterpoint to their own splat-based Marble: persistence without geometry.
- Marble GA (Nov 2025) — see 3D & spatial.
- World API (Jan 2026) — text/image/pano/video → navigable 3D worlds, credit-priced.
- Marble 1.1 / 1.1 Plus (Apr 2026) — quality bump; Plus auto-expands coverage.
AMI Labs (Yann LeCun) — the latent bet, capitalized
LeCun left Meta and raised a $1.03B seed at a reported $3.5B valuation (Mar 2026; TechCrunch + 6 corroborating outlets — briefly Europe’s largest seed until Ineffable Intelligence’s $1.1B a month later). LeCun is Executive Chairman; CEO is Alexandre LeBrun. The thesis is JEPA made commercial: “LLMs are limited to the discrete world of text… they lack a model of the world” (LeCun to MIT Tech Review). Training on video/audio/sensor/lidar for robotics, industrial control, healthcare. No product yet — year one is research.
Decart — real-time as the moat
Oasis (Oct 2024, with Etched) → MirageLSD (Jul 2025) — real-time video-to-video world transformation at <40ms/frame → Oasis 3 (Jun 2026), a pivot to physical AI: photorealistic multi-camera driving simulation at $0.02/sec, with admitted consistency and physics caveats (TechCrunch). Funding: $100M Series B at $3.1B (Aug 2025, Fortune), then $300M at ~$4B (May 2026) from Radical, NVIDIA, Adobe, Toyota.
Odyssey (Oliver Cameron) — interactive video, Amazon-aligned
Explorer (Dec 2024): text/image → 3D worlds as Gaussian splats, exportable to Unreal/Blender. Then interactive streamed video (May 2025, ~$1–2/user-hour on H100s). $310M Series B at $1.45B (Jun 2026, TechCrunch) with Amazon equity + AWS/Trainium alignment. Cameron: “the GPT-3 moment for world models.”
Runway — from video gen to world models
Coined “General World Models” in Dec 2023; GWM-1 (Dec 2025) made it real in three variants — Worlds, Robotics, Avatars — beside the Gen-4 video line and the Game Worlds beta (Aug 2025). $315M Series E at $5.3B (Feb 2026, TechCrunch); ~$300M ARR reported.
The robotics cluster
- 1X — naming care: Redwood is the 160M VLA policy on the NEO humanoid; the world model is 1XWM (Jan 2026, 14B): imagine a video rollout from a prompt, then an inverse-dynamics model converts it to joint actions. TechCrunch notes the “any prompt” framing is aspirational.
- Physical Intelligence — π0 (open-sourced Feb 2025) → π0.5 (Apr 2025, cleans unseen homes); $600M at $5.6B (Nov 2025, Bloomberg); reported talks at $11B+ (not closed).
- Skild AI — omni-bodied “Skild Brain”; $1.4B Series C at $14B+ (Jan 2026, SoftBank-led) — the largest valuation in the cluster.
- NVIDIA — GR00T VLAs are robot foundation models; NVIDIA’s world-model brand is Cosmos, and they connect via GR00T-Dreams (Cosmos generates synthetic training data).
Niantic Spatial — the geospatial angle
Sold Pokémon GO and the games division to Scopely for $3.5B (Mar 2025); the geospatial AI arm spun out as Niantic Spatial with $250M. Its Large Geospatial Model — trained on 30B+ posed images — remains announced-but-unreleased as of mid-2026.
Google — the gravitational center
DeepMind holds more of this survey’s threads than anyone: Genie 2/3 (video-generative), Veo 3 — whose “Video models are zero-shot learners and reasoners” paper (arXiv:2509.20328) introduced “chain-of-frames” reasoning and 78% zero-shot maze-solving — SIMA 2 (Nov 2025): a Gemini-powered agent whose task completion roughly doubled over SIMA 1 (~31% → ~62%) and which trains inside Genie 3’s generated worlds, plus Physics-IQ (evaluation) and the Waymo World Model (driving). The agent-learns-inside-the-world-model loop that Ha & Schmidhuber sketched in 2018 is now running at industrial scale, entirely inside one company.
Reading the landscape
Three observations fall out of the sourced record: (1) every camp got funded — pixels (Runway, Decart), latents (AMI), splats (World Labs, Odyssey), robots (the cluster) — so the debate will be settled by products, not papers; (2) real-time is the emerging moat — RTFM, MirageLSD, Oasis 3, Odyssey’s streaming, Genie 3 all compete on latency now, not just fidelity; (3) the driving pivot is universal — Decart, Wayve, Waymo, and NVIDIA all route world models toward autonomy, where simulation has a paying customer.