World-model labs become acquisition targets, and Decart sets the price

Text-to-scenario simulation data for robots is crowded. NVIDIA Cosmos and Isaac, DeepMind Genie, World Labs, Decart, Odyssey, Moving Atoms and Veeda all target it, many with lab-scale funding. Big acquirers are also absorbing the field: World Labs bought SceniX, and Decart has been in sale talks. A new entrant would struggle to win on the model itself. The realistic opening is a narrow vertical layer on top of these models: scenario libraries for particular warehouse or drone tasks, matched to specific sensors, with policy-evaluation tooling and proof that results carry over from simulation to real robots. That business sells outcomes rather than a world model. The benchmark idea has more room, though it is not empty. Physics-IQ Verified, PAI-Bench-G and many academic benchmarks (PhyGround, WorldBench, Morpheus, WMBench) exist, but they are academic, judged by an LLM, or built by model vendors. Company claims, such as Moving Atoms and Odyssey both saying they top Physics-IQ, are mostly self-reported. I found no neutral commercial auditor that acquirers or robot buyers could pay to verify physics claims, test robot-policy outcomes and check private models under NDA. That diligence-style service is the clearest gap. Its risks are that the buyer base is small and deal-driven, and that open academic benchmarks keep improving and set a free baseline.

• Decart (Oasis 3)

  An interactive world model for physical AI that takes text prompts and generates rare scenarios such as blizzards or blocked cameras. Reports from August 2026 describe acquisition talks with Anthropic, Nvidia and others at about $6–7B. One later report says Anthropic walked away. None of this is confirmed.

• Moving Atoms (Atom 1)

  Launched in August 2026. Lets users test a robot policy across a thousand simulated environments and says this costs up to 50x less than testing on real robots. It claims to top Physics IQ ahead of Cosmos, but that claim is self-reported.

• World Labs (Marble, Atlas) + SceniX

  Fei-Fei Li's lab. It generates 3D worlds from text and images and bought the robotics-simulation startup SceniX in July 2026. One search result title says AMD acquired World Labs (https://www.noze.it/en/insights/amd-acquires-world-labs/); I did not check that report.

• NVIDIA Cosmos / Isaac Sim / Omniverse Replicator

  The established open stack. Cosmos Predict generates world-model video from text, and Isaac Sim and Replicator produce synthetic data for warehouse logistics. Its main advantage is distribution.

• Google DeepMind Genie 3 (Waymo world model)

  Waymo's driving world model is built on Genie 3 and generates edge-case scenarios from text prompts.

• Odyssey (Odyssey-3)

  A single world model that drives cars and controls humanoids and drones. It reports the top score on the Physics-IQ Verified leaderboard.

• Veeda AI

  Founded in Toronto in June 2026 by Sanja Fidler, formerly of NVIDIA. Builds foundation world models that generate simulated environments for robots.

• One Robot (YC W26)

  Trains world models for each customer on that customer's wrist-camera and joint data, then simulates action outcomes. Small, with about $500K raised.

• Physics-IQ / Physical AI Bench (PAI-Bench-G)

  Existing public leaderboards for physical plausibility in world models. PAI-Bench-G uses an MLLM as judge, and its maintainers say it can miss subtle physics violations.

• Academic physics benchmarks (PhyGround, WorldBench, Morpheus, PhysicsMind, WMBench)

  Many academic benchmarks from 2025–26 test physics laws, conservation and robot-policy evaluation. Their methods are open, but none of them is a commercial diligence service.

• Reka PhysicalRealismBench

  Uses synthetic data with deliberately planted physics violations, so the correct answer is known exactly. Built by a model company, not a neutral auditor.

Checked 2026-10-12