AI Is Learning to Rehearse Reality
Inside the $4.5 billion race to build AI that can simulate what happens next—and act before reality delivers the answer.
A basketball leaves a player’s hand.
For a fraction of a second, the future is still open. The ball can fall cleanly through the net, glance off the rim, or bounce toward the baseline. A person who has watched enough basketball begins predicting the outcome before impact. The eyes catch the angle. The body anticipates the rebound. A player moves before the ball arrives.
A modern video model can generate this scene beautifully. It can reproduce the arena lights, the flex of the rim, the ripple of the net, even the disappointment on a fan’s face. Yet beauty answers the easiest question. The harder question arrives before impact: where will the ball go, and what should the player do now?
That gap between drawing the future and preparing for it has become one of the largest bets in artificial intelligence.
Language models became powerful by predicting symbols. World models are chasing a more unforgiving target: the next state of a scene, the consequence of an action, and the uncertainty between the two. From November 2025 through June 2026, six prominent US companies pursuing different versions of this idea raised $4.525 billion across their latest publicly announced or credibly reported rounds. NVIDIA, AMD, Alphabet, Amazon, Adobe, and Autodesk have entered through models, chips, cloud agreements, venture investments, and workflow partnerships.
The phrase “world model” makes the movement sound more unified than it is. Underneath it sit several competing beliefs about intelligence. One camp wants to generate the world in pixels. Another wants to ignore pixels and predict in a compressed internal space. A third teaches agents inside imagined futures. A fourth builds persistent 3D environments that software can inspect. A fifth begins with the robot and works backward from action.
They are climbing the same mountain from different sides. Capital is funding every trail because the prize is larger than better video. A machine that can rehearse consequences can create its own training experience. It can test a decision before taking it. Eventually, it can act in the physical world with less human supervision.
The race begins with a deceptively simple question: what does a machine need to know about reality in order to choose well?
Reality Has a Higher Standard
The idea reaches back long before generative AI. In 1943, psychologist Kenneth Craik proposed that the mind carries a “small-scale model” of external reality and uses it to try possibilities before acting. Reinforcement learning later turned that intuition into machinery. An agent observes an environment, takes an action, sees what changes, and builds an internal representation of those transitions.
The model never needs to contain the whole world. A driver approaching an intersection can ignore the motion of every leaf. The other car, the possible paths, the uncertain intent of its driver, and the consequences of braking belong in the decision. Intelligence comes partly from knowing what can be discarded.
In the strict control-theory sense, a world model represents how an environment changes, often after an action, so an agent can predict or plan. The commercial label has expanded far beyond that definition. Video generators predict observations. Meta’s Joint Embedding Predictive Architecture predicts compressed representations. World Labs constructs spatial state. Robot models from Physical Intelligence and Skild AI produce actions, sometimes without exposing a separately inspectable simulator.
This distinction changes how investors should read a demonstration. A renderer promises that the future will look plausible. A simulator promises that the future will respond plausibly when something changes. A planner promises a useful action. Each step raises the cost of being wrong.
An impossible reflection in a commercial video can be edited. A collision error in a factory simulation can waste days of engineering. A wrong command from a robot can damage equipment or hurt someone. The label may be shared; the product contracts are radically different.
For any company selling simulation or planning, one test cuts through the language: if the agent takes a different action, does the model predict the resulting state well enough to improve the decision?
Video generation has yet to clear that bar consistently. OpenAI’s original Sora research report argued that large video models showed a path toward general-purpose simulation. The same report documented failures in object state, glass shattering, and long-duration coherence. Since then, video has become much more convincing. The strongest published tests still find a gap between visual realism and dependable physical prediction.
WorldModelBench found that standard video-quality scores often miss violations of physical laws. Physics-IQ found no statistically significant relationship between visual realism and physical understanding among the models it tested. In a separate 2D collision-and-motion testbed, scaling improved familiar cases and new combinations of familiar elements while failing to produce reliable out-of-distribution extrapolation. The models behaved more like expert imitators than discoverers of universal laws.
This is the tension at the heart of the market. Video models are learning more about the world as they get better at generating it. The knowledge can still break precisely where a machine needs it most: after an unfamiliar intervention.
A beautiful world proves that the renderer works. Simulation begins when changing the cause changes the consequence correctly.
Five Beliefs About Tomorrow
The modern story starts in 2018, when David Ha and Jürgen Schmidhuber published World Models. Their agent compressed an environment into a learned internal space, then trained inside what the authors called its “dream.” The visual output was crude by current standards. The idea was radical: an agent could practice inside a model rather than spending every lesson in the real environment.
DeepMind pushed the same principle toward decision-making. MuZero mastered games without reconstructing every detail or receiving their rules. It learned the dynamics required for reward, value, and action. DreamerV3 imagined trajectories inside a compact learned model and trained behavior from those imagined outcomes. Using one configuration across more than 150 tasks, it became the first algorithm reported to collect diamonds in Minecraft from scratch without human data or a hand-designed curriculum.
MuZero and Dreamer carry a less cinematic business lesson: completeness can become a liability. The best internal model may be the smallest one that preserves the consequences needed for a decision. Every irrelevant detail consumes computation. Every omitted critical variable creates failure.
Yann LeCun has built his argument around this trade-off. His Joint Embedding Predictive Architecture avoids forecasting every pixel because the exact future is unknowable and much of it is irrelevant. A model can instead predict a useful representation of a missing region, another view, or a future state. It learns the shape of the world without spending all its capacity on the flicker of leaves or the texture of a wall.
Meta’s V-JEPA work has begun connecting that philosophy to physical control. V-JEPA 2 learned largely from video, then added an action-conditioned predictor trained on a relatively small quantity of robot trajectories. The 2026 V-JEPA 2.1 paper reported that one-step cup-grasp success rose from 60% to 70% versus V-JEPA 2; an eight-step planning configuration reached 80%. It also reported roughly ten times faster short-horizon navigation planning than its SD-VAE baseline. The manipulation evaluation used ten trials per skill, so these results are signals from a laboratory rather than deployment evidence. Even within that narrow frame, they reveal something important: a machine can learn representations from watching the world and later turn them into plans without generating a photorealistic movie of the future.
The video camp is making the opposite wager. Rather than discard visual detail, it wants scale to absorb the structure hiding inside it. Google DeepMind’s Genie 3 generates navigable environments as a user moves through them. Runway’s GWM-1 family extends real-time video generation into explorable worlds, avatars, and action-conditioned robot rollouts. Odyssey is building interactive simulations. Luma describes its ambition as models that can generate, understand, and operate in the physical world.
Their raw material is abundant. The internet contains an enormous visual record of falling objects, moving bodies, flowing water, human gestures, street scenes, tools, machines, and rooms. If language models extracted structure from text at scale, perhaps video models can extract intuitive physics from motion at scale.
Fei-Fei Li and World Labs approach the problem through space. Her thesis of spatial intelligence begins with a fact language models can easily obscure: people live in three dimensions. We navigate rooms, infer depth, understand occlusion, and manipulate objects. World Labs’ Marble turns text, images, video, or rough layouts into persistent 3D environments. Those environments can export Gaussian splats, visual meshes, and coarse collider meshes for existing tools.
Marble is a meaningful step beyond a fixed clip because the user can move through the result and take the geometry elsewhere. Its collider mesh does not transform it into a validated physics engine. World Labs has been unusually clear about that boundary. In its functional taxonomy, renderers create what people see, simulators preserve state and dynamics that software can use, and planners decide what to do. The long-term ambition is to bring those functions together. The current product begins with space.
NVIDIA is attempting a broader fusion. Cosmos 3 combines reasoning, world generation, and action generation across multiple forms of input. Omniverse supplies an environment for simulation. Isaac supplies robotics tools. GPUs and networking provide the computation. This route does not ask learned models to replace classical physics immediately. Neural systems can generate scenes, sensor data, rare events, and candidate futures while deterministic engines enforce the constraints engineers already trust.
That hybrid may dominate the next few years. It can create economic value before anyone builds a universal simulator.
Physical Intelligence and Skild AI start closer to the machine that must move. They are training robot foundation models across tasks and, in Skild’s case, across different bodies. Their systems sit beside the strict world-model category. A successful policy must exploit regularities in how actions change an environment, but an observation-to-action policy can encode those regularities without predicting future states in a separately inspectable simulator.
The difference may sound philosophical until money enters the picture. Investors have valued Physical Intelligence above $5 billion and Skild above $14 billion. Those prices already assume substantial progress beyond today’s bounded demonstrations. They are wagers that learning across tasks, environments, and robot bodies will eventually produce an economic machine brain.
One phrase now contains at least five beliefs: dream inside compact dynamics, predict in latent space, scale video into simulation, build spatial worlds, or learn action directly. The eventual winner may combine all five. The market has begun paying long before that convergence is proven.
The Money Arrived Before the Proof
The financing did not build slowly. It arrived in waves.
In November 2025, Luma AI announced a $900 million Series C at a reported valuation above $4 billion. Around the same time, Bloomberg reported that Physical Intelligence had raised $600 million at a $5.6 billion valuation. In January 2026, Skild AI announced a $1.4 billion Series C at a valuation above $14 billion. Runway followed in February with $315 million at $5.3 billion. Eight days later, World Labs announced $1 billion in new funding. By June, Odyssey had raised $310 million at a $1.45 billion valuation.
Together, those six latest rounds total $4.525 billion. The sum is deliberately narrow; it covers selected US-headquartered companies rather than the whole market. Its composition says more than its size.
Capital is backing the forgiving end of the market and the unforgiving end at the same time. Runway and Luma can sell creative generation now while financing a larger simulation ambition. World Labs can enter design and media workflows through 3D creation. Odyssey can turn interactive environments into a product before solving universal physics. Physical Intelligence and Skild are priced against a future labor market whose revenue remains gated by reliability in the real world.
The public companies circling these startups are also buying different forms of optionality.
NVIDIA has the broadest position. It supplies the dominant accelerated-computing platform, builds Cosmos, Omniverse, and Isaac, and has invested in World Labs, Runway, Skild, Luma, and Odyssey. Its advantage comes from participating across several technical routes. Its risk comes from assuming too much of the future compute pool while AMD and hyperscalers build alternatives.
AMD Ventures has backed World Labs, Runway, Luma, and Odyssey. These stakes create strategic access to major multimodal-model developers. They are not evidence of Instinct demand until those companies disclose AMD-based training or inference workloads. The distinction matters for public investors because a venture relationship can validate a category without moving product revenue.
Alphabet owns perhaps the most unusual collection of assets. Google DeepMind has Genie, Dreamer, and years of model-based reinforcement-learning research. Project Genie has been grounded in Street View imagery. Waymo gives Alphabet a separate embodied-AI business and a potential source of future learning loops, although public evidence does not show its driving logs training Genie. CapitalG led Physical Intelligence’s 2025 financing, while GV backed Odyssey. Alphabet has research, distribution, venture exposure, and an autonomous-driving business; the world-model economics may still remain buried inside a company of its size.
Amazon’s participation in Odyssey reveals the cloud strategy more cleanly. Odyssey named AWS its preferred cloud provider and said it would optimize for Trainium. Amazon can fund the application and compete for the workload it creates.
Autodesk has the clearest disclosed bridge into an existing workflow. Its $200 million investment in World Labs includes a strategic advisory role and plans to explore integration with Autodesk design software. If that work becomes a product, world generation gains access to architects, designers, engineers, and media creators who already pay for spatial tools. The model does not need to become general intelligence to save them time.
This is where the capital story becomes more interesting than the funding total. Venture investors are betting on standalone model companies. Strategic investors are positioning chips, clouds, and software around those models. If the frontier commoditizes, much of the durable value can still flow to the companies that own computation, distribution, workflows, and deployed machines.
The deepest bet is therefore larger than any startup valuation. World models could turn video inference, simulation, and robot learning into persistent computational workloads. Every generated world can be explored. Every exploration produces more states. Every policy can be tested across thousands of variations. The appetite for computation grows with the number of futures a machine can afford to rehearse.
Rehearsal Becomes a Business
The world-model market will earn revenue in the reverse order of its ambition.
Creative work comes first because creative errors are cheap. A studio can use generated video for previsualization. An advertising team can explore concepts before a shoot. A game developer can sketch a world before artists build it. A designer can generate spatial alternatives while a person chooses what survives. Runway and Luma can fund long-term research with products whose current value is visible in a single session.
The next market begins when generated environments become rehearsal spaces.
Robotics companies need rare failures that are dangerous or expensive to collect physically. Autonomous-vehicle systems need edge cases that appear only after millions of miles. Factories need to test layouts before moving equipment. Architects need environments that remain coherent while they are changed. A useful simulator can reduce real-world training hours, expose policy failures earlier, and widen the range of conditions an agent sees before deployment.
Perfection is unnecessary. Economic usefulness arrives when the simulator improves a measurable outcome. If synthetic experience increases robot success, lowers testing costs, or catches a design error, the model has earned its place in the workflow.
This is why hybrid systems have the strongest near-term position. Learned models can supply variation; CAD systems, physics engines, and digital twins can supply constraints. The neural model does not need to rediscover every engineering law from video before customers benefit. It needs to generate the uncertainty and diversity traditional tools handle poorly.
By 2029, the market should begin judging world models by intervention accuracy and sim-to-real performance. Video quality will remain commercially useful, but the strategic benchmark will shift: when an action changes, does the predicted outcome change correctly, and does training on that prediction improve behavior outside the simulator?
That transition favors companies with proprietary action-and-outcome data. Internet video shows what happened. Robot trajectories, driving logs, simulation rollouts, and human interaction traces show what an agent did and what changed afterward. A deployed fleet or a high-volume workflow can turn each failure into another training case.
Distribution then becomes part of the intelligence. A model inside Autodesk can learn from design workflows. A robot model inside a working fleet can collect new trajectories. A simulation platform tied to NVIDIA’s tools can sit beside training and deployment. A benchmark leader without a workflow receives less feedback and becomes easier to replace.
Through 2029, the strongest businesses will sell useful incompleteness: creative rendering with paying users, hybrid simulation connected to engineering or robotics, and compute platforms that earn revenue across several architectures. Universal reality engines can wait. Customers will pay for smaller worlds that solve expensive problems.
Physical action comes last because its errors are unforgiving. A household robot can succeed nineteen times out of twenty and still be unusable if the twentieth failure breaks a glass beside a child. Long tasks compound small mistakes. Hardware wears out. Safety certification takes time. Customer environments resist standardization.
Industrial work will move sooner. Warehouses, factories, laboratories, and logistics sites can constrain the environment and define success clearly. “General” robot models may first earn money through bounded jobs such as sorting, inspection, material handling, or repetitive manipulation. Their intelligence can generalize more than traditional automation while the workplace limits the chaos they must survive.
Revenue therefore moves from forgiving outputs to unforgiving ones: pixels first, simulation next, physical action last.
The Scarce Asset Is Consequence
The long-term thesis begins where the near-term product ends.
David Silver and Richard Sutton describe an approaching “era of experience”, in which agents improve through sustained interaction rather than depending mainly on static human-created data. A world model gives that agent somewhere to practice between encounters with reality.
Imagine a warehouse robot that fails to grasp a new package. The failure creates a trajectory: the camera view, the attempted movement, the contact, the slip, and the final state. A world model identifies where its prediction diverged. Simulation generates variations around that failure—different angles, lighting, packaging, friction, and timing. The policy practices against those variations. The improved robot returns to the warehouse and produces new evidence.
Deployment feeds simulation. Simulation improves the policy. The policy returns to deployment.
This loop is the real long-term moat. The LLM boom competed for human-created text. Physical AI will compete for trajectories recording what an agent did, what changed, and where its prediction failed. Companies with machines in the world can collect data unavailable in passive video. Companies with simulation can multiply each physical lesson. Companies that own both can compound faster.
The architecture may eventually blend today’s camps. A system could render observations for people, maintain a compressed state for reasoning, simulate counterfactual futures, and generate actions for machines. NVIDIA’s Cosmos 3 is moving toward a combination of reasoning, generation, and action. World Labs expects the borders among renderers, simulators, and planners to narrow. Meta is showing how latent prediction can feed robot planning. Runway is extending video generation into action-conditioned rollouts.
The winner will be the company that closes the experience loop. Until deployment data improve the simulator and the simulator measurably improves the deployed policy, “unified world model” remains option value rather than operating proof.
Three failures could break the investment thesis. Video scaling may remain trapped in plausible imitation, forcing customers to gather near-exhaustive deployment data. Generated environments may fail to improve real policies, leaving classical simulators in control of high-value engineering work. Robotics may remain constrained by hardware, maintenance, safety, and customer integration after model quality improves.
If those conditions persist, the category can still expand while most value accrues to chips, clouds, existing simulation software, and deployed fleets. Independent model labs would face the same pressure language-model companies already feel: immense training costs, rapid imitation, and uncertain differentiation.
For public-market investors, NVIDIA currently offers the broadest exposure, Alphabet the widest collection of internal options, Autodesk the clearest workflow bridge, Amazon a cloud-and-silicon route through Odyssey, and AMD a portfolio of strategic relationships whose hardware payoff remains unproven. None deserves a valuation change solely because “world models” appears in an investor deck. The signal will come from workload growth, product integration, customer adoption, and evidence that simulated experience changes real behavior.
Private investors need an even sharper question. A vivid demonstration can hide a weak business. The useful test is whether intervention predictions hold over the required horizon, whether simulation improves a customer workflow or real policy, how much deployment data remain necessary, and who owns the feedback when the model fails. Revenue from a current product is far more valuable than a financing round sustaining a product customers may want later.
The defining metric will be decision usefulness per dollar.
Before the Robot Moves
Return to the basketball suspended above the rim.
The renderer creates the arena. It makes the shot look real.
The simulator carries the ball forward. It predicts the miss and the direction of the rebound.
The planner chooses a path. The robot moves before the ball arrives.
Capital is funding all three moments. Over the next few years, rendering should produce the clearest revenue and hybrid simulation the strongest enterprise pull. Robot-planning companies may continue attracting the richest private valuations, with operating proof lagging until reliability improves. Over the longer term, these functions can converge inside systems that learn continuously through action and consequence.
The decisive breakthrough will come from a machine that can imagine the wrong move, understand what follows, and choose differently before reality makes the mistake expensive.
That is the deeper wager behind the world-model boom.
AI learned to speak by predicting the next word.
Now it is learning to move by predicting the next world.