TL;DR
World models may eventually make agents better at predicting what happens when they act, in physical or simulated environments. They are far behind large language models in product maturity and deployment evidence. Their development could end up more concentrated than text AI, because training and serving video- and simulation-heavy systems needs more compute, storage, power and privileged operational data than text does.
The practical consequence is that businesses and public bodies will be offered “physics-aware” systems, and asked to buy them, before anyone can show they work in an uncontrolled setting. The uncertainty runs in both directions: nobody has demonstrated a general-purpose, dependable world model for messy real-world conditions either.
In plain English
A world model is software that holds some representation of an environment and predicts how it changes, particularly in response to an action. A language model predicts the next piece of text. A world model is meant to predict what an environment will look like after something happens in it.
The three functions matter because each is judged against a different standard. A renderer is judged mostly on whether the output looks convincing. A simulator has to hold state and predict transitions correctly. A planner has to choose actions that still make sense given those predicted transitions. Moving along that chain, evaluation gets more expensive and more safety-sensitive at every step.
The recurring term for what goes wrong is the simulation gap, sometimes called the sim-to-real gap: the difference between how a system behaves in a generated environment and how it behaves around actual staff, machinery, weather, lighting and people doing unexpected things.
Fei-Fei Li and the World Labs team have published a functional taxonomy of world models, and they open by conceding the problem: this is “one of the most important and most overloaded terms in AI today.” Computer vision, robotics, reinforcement learning and generative AI all claim to be building world models, and each means something different by it.
The essay breaks it into three functions:
- rendering, which produces pixels for people to look at;
- simulation, which tries to capture structure, geometry and physical dynamics for machines to use;
- and planning, which selects an action toward a goal, like telling a robot to pick up a cup and put it somewhere else.
World Labs’ taxonomy is useful because it keeps “looks like a world” and “understands how the world will respond” as separate claims. A system can generate a convincing video, an explorable virtual scene or a game-like environment without being a safe simulator of what a warehouse arm, a delivery robot or a vehicle should do next.
Google DeepMind’s Genie 3 announcement says Genie 3 generates navigable worlds from a text prompt “at 24 frames per second, retaining consistency for a few minutes at a resolution of 720p.” That is a genuinely impressive interactive-generation result. iInteraction duration is capped at a few minutes, the model cannot reproduce real-world locations with geographic accuracy, and modelling several independent agents in a shared environment remains unsolved.
Stanford HAI’s 2025 AI Index reports that nearly 90% of notable AI models in 2024 came from industry, up from 60% in 2023, and that training compute doubles roughly every five months. Epoch AI puts frontier training compute growth at 4-5x per year over 2010 to May 2024, with the frontier specifically slowing to about 4.2x annually after 2018, and attributes much of it to larger training clusters running more hardware in parallel.
World-model development can push on all of those pressures at once, for reasons that have nothing to do with hype:
| Requirement | Typical LLM path | World-model path | Why it affects access |
|---|---|---|---|
| Training data | Text, code, documents, structured records | Video, multi-camera observation, depth, trajectories, sensor feeds, action logs | Video and sensor data cost far more to store and move, and the valuable action data usually sits with private operators |
| Ground truth | Human preferences, answers, code tests, text benchmarks | Object persistence, geometry, contact, motion, control outcomes, safety constraints | Evaluation needs richer instrumentation and often a physical test environment |
| Inference | Text generation, retrieval, limited tool calls | Repeated visual prediction, scene updates, control loops, sometimes real-time constraints | Latency, GPU memory and uptime become operational constraints rather than background cloud costs |
| Scarce input | High-quality data and compute | Compute, specialised data, power, facilities, hardware fleets, domain expertise | A startup or a public institution can be shut out even when the model weights are freely available |
Access to useful inference has genuinely got cheaper. The inference cost for a system performing at GPT-3.5 level “dropped over 280-fold between November 2022 and October 2024.” and it is why plenty of businesses can now use capable language systems without owning a data centre.
Compute concentration is an electricity and planning question as much as an AI-lab one. The IEA estimates that data centres used around 415 TWh of electricity in 2024, about 1.5% of global consumption, having grown at 12% per year over the preceding five years, and projects that to roughly double to around 945 TWh by 2030 in its base case. But world models don’t automatically need more compute than every language model, and I would be sceptical of anyone claiming a tidy multiplier.
The direction is still clear enough though. Video, three-dimensional representations, sensor streams and interactive rollouts move far more data than a text prompt. A system forecasting many frames, weighing several possible actions and responding fast enough to control a machine has a harder serving problem than a chatbot summarising a document. Electricity, networking, cooling, storage, redundancy, safety systems and local grid capacity are all a part of that cost.
Which changes the policy question. “How can every country train its own frontier model?” is too crude to be useful. The better version: which capabilities have to stay contestable, publicly auditable and locally available, and which should be pooled because duplicating them everywhere is waste?
Nobody needs to pretend that every university, council, hospital or SME should own a world-model training cluster, but the thing worth avoiding is a situation where the only credible answer to a public-interest, industrial or safety question is “ask the cloud provider that owns the model.”
The OECD’s framework for national AI compute capacity is still the right shape: build for capacity (access to compute), effectiveness (people, policy and allocation mechanisms that decide whether that compute helps anyone useful) and resilience (security, sustainability and sovereignty).
Europe has started building pieces of it. The European Commission says its EuroHPC AI Factories provide computing time and support services to European industry, research, academia and public authorities, and it has set out a €20 billion mobilisation target for AI Gigafactories intended to give EU innovators, researchers, SMEs and public-sector users large-scale compute and data access.
Those are infrastructure commitments but aren’t yet evidence of fair access, and the detail that decides it is allocation. Who will actually receive time on the machines? Can a university or a public-interest lab get capacity without a long procurement cycle? Are SMEs offered technical support, data-handling guidance and realistic quotas? Do public bodies keep the right to inspect how a system was built and tested, its security controls and its failure reports when systems built on that compute enter public services? Can European users move models, data and workflows without becoming dependent on one vendor’s proprietary runtime?
This can go wrong without anyone noticing. A national compute programme that funds large pre-training runs but cannot support a robot-test facility, a transport safety simulation, a public benchmark suite or an independent audit lab will reproduce exactly the same concentration in a different building.
For businesses and operators, the useful preparation is narrower and more immediate than “adopt world models.” Most don’t need one. They need a way to evaluate the proposals that will arrive under that label, and I would put these into procurement and governance now.
The usual argument about unequal AI access is about who can use a model. That still matters, particularly for schools, small firms and public services. World models add a second division underneath it: who determines what gets represented, what gets measured, whose environments supply the training data, and whose definition of failure becomes the standard.
A warehouse operator with proprietary video and robot logs can tune a model around its own workflows. A city with access only to a vendor dashboard may be asked to trust an opaque prediction about its roads, public spaces and infrastructure. A school system may be handed a polished educational tool with no say in the data, the evaluation or the behavioural assumptions inside it. Those are differences in institutional power. Changing subscription tier does not address them.
The field is early, and Li’s own essay says so. She describes World Labs’ Marble as “only the first chapter of a much longer arc being written across the field,” and states plainly that reconciling the tensions between rendering, simulation and planning inside a single architecture is, in her words, “the defining open problem in world model research today.” That’s not a reason for panic or for complacency. Terms can still be set, and doing it now costs far less than doing it once the systems are in place.