Since 25 February 2025, ECMWF has run AIFS in production: the first data-driven global model operational at an international weather centre, producing a 10-day forecast in about two and a half minutes on a single A100 GPU. What gets left out is that it does not start from scratch — it inherits its initial conditions from the physics-based IFS data assimilation cycle, and runs at 28 km against the physical model’s 9 km. That resolution is not where local warnings are decided, and the extreme-event limits are documented by the model’s own authors.
Every six hours, at the European Centre for Medium-Range Weather Forecasts’ data centre in Bologna, a data assimilation system ingests roughly 60 million quality-controlled observations: polar-orbiting and geostationary satellites, radiosondes, commercial aircraft, ships, buoys, surface stations. The output is not a forecast. It is a three-dimensional snapshot of the atmosphere at time zero — the initial condition everything else departs from. It is the most expensive and least photogenic step in the chain, and it runs on supercomputers.
From there, since 25 February 2025, two paths diverge. One is the IFS, the physics-based model that numerically solves fluid dynamics equations on a 9-kilometre grid. The other is AIFS, a machine learning model — a graph neural network trained on the ERA5 reanalysis and on operational analyses — that solves no equations at all: it has learned, from decades of consecutive atmospheric states, to produce the next state given the current one, and iterates that step out to ten days. A full inference takes about two minutes and thirty seconds on a single A100 GPU, input and output included. It is the first data-driven global model put into operational service by an international weather centre.
The misreading: the GPU did not replace the supercomputer
The line going around — “Europe’s weather model now runs on one GPU in minutes” — is simultaneously true and misleading, because it cuts away half the process. AIFS does not generate its own initial conditions. It takes the ones produced by the IFS physics-based assimilation, the same six-hourly cycle over 60 million observations that requires HPC. Those two and a half minutes are the forward propagation time, not the time of the forecasting system.
This is not accounting pedantry. Assimilation is where operational meteorology meets the physical world: it is the step at which a radiosonde launch in the Arctic or a buoy in the Tyrrhenian Sea actually enters the model. A purely learned system that also produced its own initial state would be a different animal with a different relationship to observations. That is not what is running today, and the reason AIFS performs as well as it does is precisely that it inherits the most carefully curated object in the whole chain.
ECMWF itself is explicit on this: the two systems are presented as complementary rather than substitutive — the position publicly stated by Florian Pappenberger — while Director-General Florence Rabier has called the launch a milestone. Neither formulation is “AI won”.
Twenty-eight kilometres is not nine
Here is the second misunderstanding, and it bears directly on the warnings question. AIFS Single runs on a reduced Gaussian N320 grid, roughly 0.25 degrees, about 28 kilometres. AIFS ENS, operational since 1 July 2025 with 51 members, sits at 31 kilometres. The deterministic IFS and its physical ensemble work at 9 kilometres. ECMWF states that the coarser resolution “reduces spatial representativeness”.
Twenty-eight kilometres is a scale at which a back-building thunderstorm over a river plain, a convective cell confined to an Alpine valley, or a storm surge in a narrow gulf simply do not exist as distinct objects. Those are the scales that decide whether a road closes. Warning systems everywhere rest on high-resolution limited-area models and on radar nowcasting — tools built for exactly those scales, themselves fed by boundary conditions from global models. A 28-kilometre global model, learned or physical, contributes upstream: it points, it does not sign.
This is worth stating plainly because in public coverage AIFS is often framed as if it were already landing on the desk of whoever issues the bulletin. There is no documented use of it as the basis for issuing official warnings, and the geometry of the problem explains why the opposite would be surprising.
What it actually gains
None of which makes the gains cosmetic. ECMWF reports improvements of up to 20% on specific measures — tropical cyclone tracks, and surface temperature for the ensemble — and an energy reduction of roughly a thousandfold per individual forecast. On that second figure caution is warranted: it is a centre-stated number, and the perimeter it covers is not publicly documented — whether it includes data assimilation and training is unclear. AIFS was trained on Leonardo, the EuroHPC supercomputer hosted at Cineca in Bologna, in about three days.
The version history reads like a system under active maintenance rather than a trophy: pre-operational phase four times a day from October 2023, version 1.0.0 in production on 25 February 2025, version 1.1.0 on 27 August 2025 — released to fix a precipitation problem, with a 4–6% improvement on upper-air and surface variables. A precipitation bug patched in a minor release is exactly the kind of detail triumphalist coverage omits and that tells you how an operational system really behaves.
The structural limit: the mean is smooth
The authors themselves document the least comfortable point. Models trained by minimising mean squared error tend to smooth their fields at longer lead times: when uncertainty grows, the MSE-optimal move is to blur — to produce something closer to the climatological mean than to a crisp scenario. The consequences are measurable and known: under-forecasting of the heaviest precipitation events, underestimation of tropical cyclone intensity.
A study published in Science Advances found that on record-breaking events physics-based models remain ahead of AI-based ones, particularly for heatwaves, whose frequency and intensity the learned models tend to underestimate. The explanation is intuitive once stated: a system that learned the historical distribution struggles to extrapolate outside its training domain, and a record event is by definition outside that domain. A physics-based model has no distribution to fall back to — it has equations, and equations do not know a value is anomalous.
This inverts the common intuition that AI will spot extremes before humans do, and the available evidence does not support it. Combined with the 28-kilometre grid, it maps out quite precisely where AIFS should not be placed today.
The part that matters for whoever builds models
There is an aspect of this story cited far less than the two and a half minutes, and probably more consequential. ECMWF has published the model weights on HuggingFace; the training framework is open source — it is called Anemoi, co-developed with European national meteorological services, and it received the EMS Technology Achievement Award 2025 — and since 1 October 2025 the real-time data catalogue is licensed under CC-BY-4.0. AIFS Single and AIFS ENS together produce around 6 terabytes of data per day.
In practice this means a national weather service, or a research group, can train its own variant on Anemoi rather than depend on a private company’s closed weights. That is a difference in industrial structure, not in score. Meanwhile ECMWF itself runs third-party data-driven models daily for comparison and verification: DeepMind’s GraphCast, Huawei’s Pangu-Weather, Microsoft’s Aurora, NVIDIA’s FourCastNet. And the United States’ NOAA moved AI models into operations in January 2026.
One final figure, pointing the other way from the technical enthusiasm: a UK Met Office study from August 2026 found public trust in AI-based forecasts to be lower than trust in existing methods. For a warning system, where the only metric that ultimately counts is how many people do what the bulletin asks, that number weighs as much as any verification score.