Inside the labs building the next generation of AI, a phrase has been gaining weight: world models. A large language model like ChatGPT, Claude, or Gemini predicts what comes next in text. World models, in contrast, learn dynamics from observation, then simulate forward to test what happens next. They model the world itself, rather than just descriptions.
Yann LeCun, who left Meta in late 2025 to launch Advanced Machine Intelligence Labs, has built his research program around it. Demis Hassabis, who runs Google DeepMind, has made world models central to its push toward more general AI. Sam Altman has called OpenAI’s Sora a world simulator, a claim that is contested. Fei-Fei Li raised a billion dollars for her company World Labs to pursue what she calls “spatial intelligence.” Jensen Huang, meanwhile, is building the simulation platforms and compute behind the next wave of AI, as NVIDIA did for large language models.
The term is used loosely; not everything marketed as a “world model” qualifies in the strict architectural sense.
The bet, and the reason LeCun rejects video-generators like Sora, is that a model trained on how a system behaves rather than how it looks, an architecture he calls JEPA, will generalize better to the physical world. It is not a product category but an architecture that could take AI from fluent at language but with no real model of the physical world, to a grounded understanding of how that world behaves.
If they are right, this is more than another commercial AI cycle. It is the period in which the substrate of the next AI gets built, and what gets built now, by whom and on what data, will shape what AI can do for years. The potential for solving problems in climate, oceans, the biosphere, and the biology of disease is vast.
What AI has already accomplished for the Earth #
I started my career as a climate scientist at NASA running ocean-atmosphere simulations on supercomputers. I later co-wrote, with the World Economic Forum and Microsoft’s chief environmental officer, two of the earliest reports on AI and the Earth system. A lot of what we predicted has happened.
AI now spots wildfires and methane leaks from orbit. Today’s weather forecasts are unrecognizably better. Google’s flood forecasting runs in over 150 countries. Neural weather models, from DeepMind’s GraphCast to systems now run by the public forecasting agencies themselves, have matched or beaten the best physics-based forecasts at a fraction of the compute cost, though they still trail on extremes and tail risk.
These are real gains. But the hardest problems have barely moved: what a hurricane will do at landfall, when the next drought breaks, how ocean circulation behaves as the ice melts.
Sub-seasonal weather forecasts—the window that drives water, energy, and agricultural planning—remain weak. The forests, soils, and vegetation that absorb roughly a third of our emissions carry the largest uncertainty in the entire carbon budget. Unlike fossil fuel emissions or atmospheric carbon dioxide, this land carbon sink cannot be measured directly at the global scale; it has to be inferred. And tropical convection, the storm systems that deliver rainfall for billions, unfolds at scales too small for global models to capture, so they fall back on rough approximations that scientists have tried to improve for decades.
Why haven’t these gaps closed? Not for lack of computing power, now trillions of times more powerful. Not for lack of data, now planetary in scale. Not for lack of physics, well established for the parts we understand. The bottleneck is representation: finding a way to model systems we can’t describe exactly because the physics is only partly understood and the measurements are sparse. Simply scaling up today’s language models doesn’t solve that.
The Earth system is modeled in pieces (atmosphere, ocean, ice, land), which simplifies away the signals that live in the coupling between them. The variables that matter most are largely unobserved: root-zone soil moisture, the deep ocean, the cavities under ice shelves. Higher resolution helps—the newest models capture storms directly—but a finer grid running the same approximations is still running approximations.
The hardest problems in science sit in the gap: too poorly understood to write down in equations, too sparsely observed to learn from data alone.
This is where world models could matter, not by replacing physics, but by learning the dynamics from the joint Earth-system record, with the physics we trust enforced as hard constraints. For chaotic systems, a world model can learn the unwritten dynamics while respecting the written ones.
Some of this is already visible. AlphaFold, NVIDIA’s Earth-2, and GraphCast are in operational use across biology and weather forecasting. They work where physics is partly understood, and observations are rich. What none yet does is learn the dynamics of the open systems whose uncertainty has barely narrowed: sea level, the carbon cycle, the coupled behavior of a warming planet.
World models will not deliver certainty. They will not collapse the sea-level range to a point estimate. Their value is narrower but real: tighter, more honest ranges of plausible futures: the band coastal planners and banks actually need. For trillion-dollar adaptation choices, that matters.
There is also a hard limit. A world model cannot forecast a regime the Earth has never entered, like the world after an Atlantic circulation collapse, because there is no data from the far side to learn from. No method can. But the relevant question is how close we are to it, which these systems can begin to help answer.
The opportunity #
Three forces have converged. The architecture has matured: these systems can now be trained at scale. The Earth is now instrumented at a level that makes learning dynamics across the atmosphere, ocean, land, and ice possible. And the capital that built the large language model era has begun to reallocate.
Enterprise applications have customers, contracts, and quarterly results. In 2026, the investment following world models is overwhelmingly driven by these metrics. But where the output is a public good, such as a narrower sea-level range, a better carbon-cycle model, or earlier warning of how the next drug-resistant pathogen will spread, the commercial model breaks down.
And what these models train on shapes what they become. A world model trained predominantly on warehouse logistics, driving footage, and engineering data—the territory of physical-AI ventures like Prometheus, the Jeff Bezos startup building an “artificial general engineer,” now valued at $41 billion—learns a particular kind of physics. One trained also on planetary observation, cell-biology, and grid dynamics learns something else.
The same architectures can work for the systems we live inside, given different data and different choices. The question is whether the labs commit to these problems as ambitiously as to enterprise, and whether the institutions holding the world’s most valuable scientific data make it available.
Beyond the enterprise #
The language era taught AI to describe the world. World models aim to learn how it behaves. If their strongest advocates are right, they will become the substrate on which the next generation of AI is built: a scientific tool for understanding systems we cannot fully specify, and the capability AI needs to reason about and act in the physical world.
But getting better at the easy problems is not the same as reaching the hard ones. Whether world models reach the open systems depends on which data gets stitched together, which questions the leading labs take on, and whose problems get attention.
The AI build-out is only beginning, and public acceptance will help determine its pace. The stronger the case that these systems help solve humanity's hardest problems, the stronger the case for building the infrastructure they require.
The technical work will continue regardless. The question is no longer whether we can build world models. It is what we choose to model—and whether the hardest problems shape these systems from the start, or inherit them.