AI video looks real but misses physics, studies find, as simulators still fetch billions

Whether a neural network trained on enough video can stand in for the physics, rendering and engineering software that industry already runs on, or whether AI will instead end up operating that software, is still open. The industrial announcements at Nvidia's GTC Berlin conference leaned towards the second option, with Synopsys and Siemens wiring AI models into existing simulation and product-data tools rather than replacing them. That is Polstob's reading of the releases, not a verdict. This analysis tests the view that AI makes sophisticated software more useful by driving it.

The sharpest evidence on the question comes from benchmarks. Physics-IQ, set out in a paper first posted in January 2025, tests video models on scenes involving fluids, light, solid objects, magnets and heat. Across the systems tested, including OpenAI's Sora, Runway, Pika, Google's Lumiere, Stable Video Diffusion and VideoPoet, the authors found the models' grasp of physics severely limited, and no better among those whose output looked most realistic. Their summary was blunt: "visual realism does not imply physical understanding". They added that some test cases could already be solved, and that they expect rapid progress.

Other research points the same way. A study led by Bingyi Kang trained video models on simulated clips of objects moving and colliding under classical mechanics, and found they handled cases close to their training data but broke down on unfamiliar ones, copying the nearest example they had seen instead of learning a rule. The authors concluded that making models bigger would not on its own be enough. Google DeepMind lists limits for its Genie 3 interactive world model too, including sessions that hold together for minutes rather than hours.

In the Sora technical report, published in February 2024, OpenAI wrote that "scaling video generation models is a promising path towards building general purpose simulators of the physical world". It said Sora could keep people and scene elements consistent as the camera moved, and could often, though not always, keep track of people and objects when they were hidden or left the frame, and that these abilities came from scale rather than from rules written in for 3D or objects. The same report admitted that Sora "does not accurately model the physics of many basic interactions", giving glass shattering as an example.

The Physics-IQ authors themselves left room for learned physics, writing that some principles may be learnable from observation alone. One newer entrant, Odyssey, said on 8 October 2026 that the Pro version of its Odyssey-3 world model set a new high on a verified version of the Physics-IQ test. That is the company's own claim, which Polstob has not checked. If such gains hold up under outside testing, the line between looking right and being right could narrow.

Nvidia's annual report describes Omniverse as software for real-time 3D design collaboration and for simulating virtual worlds, and says the company has sped up releases of open model platforms including Cosmos for physical AI. On its blog, Nvidia has shown a developer directing a frontier AI model to assemble a self-driving test environment modelled on San Francisco's Market Street, connecting assets, traffic, sensor simulation and a driving model in stages and checking each link. The agent did not replace the simulator; it built and wired one.

Tripo says its EdenLive renderer separates what happens in a scene from how it looks: a code agent, game engine or simulation engine decides how the scene evolves, and the model paints the visuals in real time. Siemens describes its Digital Twin Composer, built on Omniverse technology, as the layer that brings the physics of tools such as Tecnomatix Plant Simulation and Process Simulate into one managed 3D scene. In film, Digital Domain says its machine-learning muscle system turns a skin simulation driven by simulated bones, muscles and tissue into deformations animators can use directly in their rigs.

Synopsys completed its acquisition of Ansys on 17 July 2025 for aggregate consideration of about $34.9bn, made up of $17.6bn in cash and $17.1bn in Synopsys stock at fair value, with the balance in assumed equity awards and other items, according to its annual report. The same filing calls Synopsys the global leader in engineering simulation and analysis software. It also records that in December 2025 Nvidia bought about 4.8 million Synopsys shares at $414.79 each in a private placement, for net proceeds of $2.0bn.

Siemens completed its purchase of Altair, a simulation and analysis software provider, on 26 March 2025 at an enterprise value of about $10bn, according to its closing announcement. Siemens said the deal added mechanical and electromagnetic simulation, high-performance computing, data science and AI to its portfolio.

At GTC Berlin, Synopsys said it had joined Nvidia's Halos safety ecosystem with a workflow it describes as grounded in physics and traceable, spanning sensor simulation, digital twins and safety analysis. Siemens and Nvidia said they are combining Siemens's Teamcenter product-lifecycle software and Process Preparation X with Cosmos to prepare circuit-board inspection using a virtual reference board and synthetic data before the first physical board reaches the line. Both are company announcements, and the inspection project's published goals are targets, not results.

Not everything supports the orchestration view as a business. In the risk factors of its latest annual report, Nvidia lists Omniverse among its standalone software offerings and warns that it may incur significant costs and "may not achieve any significant revenue" from them. The filing gives no separate revenue figure for Omniverse.

The visual-effects industry offers mixed evidence. Digital Domain has said that, for the six months to 30 June 2026, its loss and net current liabilities may cast significant doubt on its ability to continue as a going concern, though its directors consider the going-concern basis of accounting appropriate. Separately, Digital Domain says it suspended its traditional short-form work in Shanghai after easy-to-use AI tools changed client budgets and practices. Technicolor's UK studio business went into administration; its administrators pointed to rising labour costs and the lingering effects of the writers' strike, and did not mention AI.

Wētā FX describes a narrower role for AI: its chief technology officer, Kimball Thurston, has said its assistive AI is not allowed to change creative intent and only fixes pixels.