Skip to content
AI & SustainabilityAnalysis

The energy cost of a forecast: what AI weather models save, and what the figure leaves out

A research analysis of what the published figures say about the energy AI weather models use. Europe's forecasting centre reports that its AI model cuts the energy of making a forecast by about 1,000 times, and the US national weather agency says its AI model uses about 0.3% of the computing of its standard model. Those figures are the agencies' own, they count the forecast run and not the training or the physics systems the AI still depends on, and they say nothing about accuracy at a beach. Below: the numbers, what they cover and what they leave out.

AI that saves energy

Most of what is written about AI and energy is about AI using more of it. Weather forecasting is a case that runs the other way. A traditional forecast solves the physics of the atmosphere on a supercomputer with thousands of processors. An AI model, once trained, produces a comparable forecast in minutes on far less hardware: ECMWF's AI model can produce a 10-day global forecast in about 2 minutes 30 seconds on a single GPU.

The savings the forecasting agencies report are large. ECMWF says its AI model brings a reduction of approximately 1,000 times in the energy used to make a forecast, and that its AI ensemble runs over 10 times faster than the physics-based system. The US national weather agency, which launched its AI models on 17 December 2025, says its main AI model uses about 0.3% of the computing resources of its standard model for a 16-day forecast and finishes in approximately 40 minutes, and that its AI ensemble of 31 members uses 9% of the computing of the physics ensemble.

The published figures

Each figure is the publishing body's own statement, and each measures something slightly different:

  • About 1,000 times less energy
    Who states it
    ECMWF, for its AI model (2025)
    What it covers
    Making one forecast, against its physics model
    What it leaves out
    Training; the physics system that supplies the starting conditions
  • About 0.3% of the computing
    Who states it
    US national weather agency (Dec 2025)
    What it covers
    One 16-day forecast, against its standard model
    What it leaves out
    It is computing, not measured energy; training
  • 9% of the computing
    Who states it
    US national weather agency (Dec 2025)
    What it covers
    A 31-member AI ensemble, against the physics ensemble
    What it leaves out
    Same as above
  • 21 to 1,273 times less energy
    Who states it
    A 2026 life-cycle study of 7 AI models
    What it covers
    Training plus a year of forecasts, against ECMWF's physics model
    What it leaves out
    Low-resolution forecasts only; development runs before the final training

What the figures leave out

Training is the first gap. The headline savings count the forecast run, and an AI model has to be trained before it can make one. The 2026 life-cycle study, which compared 7 AI weather models with ECMWF's physics model, found that training costs considerably more than a single physics-based forecast but is rapidly repaid: over a year of use, the AI models were estimated to use 21 to 1,273 times less energy. So the saving survives the accounting, but the range is wide, and the low end is a long way from 1,000.

Development is the second. A September 2026 review by a Spanish supercomputing centre found that of 12 leading AI weather and climate models, 3 report no computing figure at all and the rest report only the final training run. It argues that a final run can understate the development effort behind it by roughly 4 to 10 times, an estimate drawn from one AI company's spending, not from weather models. The same review notes that the same experiment can produce up to 30 times more carbon emissions depending on the electricity grid it runs on.

The third gap is that AI forecasts do not replace the physics. ECMWF says its AI ensemble relies on physics-based data assimilation for its starting conditions, and that its physics ensemble remains indispensable for high-resolution fields. The AI models were also trained on decades of past weather reconstructed with physics models. The supercomputers stay on.

Cheaper is not the same as better at your beach

The energy saving is real, but it is bought partly with resolution. ECMWF's AI model has a grid spacing of 28 km and its AI ensemble 31 km, against 9 km for the physics models. The agencies report gains on large-scale weather, of up to 20% on some measures, and the US agency reports that the first version of its AI model is worse at tropical cyclone intensity. None of these figures is about a sea breeze, which is smaller than a 28 km grid can resolve.

Cheap forecasts also invite more forecasts. The US agency now runs a hybrid ensemble of 62 members, half physics and half AI, and it outperforms either alone on most measures. That is a better forecast, not a smaller energy bill. Whether total energy falls depends on what agencies do with the saving.

What it means for riders, and for AYON Ecosystem

A forecast is one of the lowest-energy things AI does, and it is getting cheaper. That makes weather a useful counter-example in the debate about AI's footprint, alongside the costs described in our analysis of a single prompt. But a low energy cost does not tell a kitesurfer whether the wind number is right.

AYON Ecosystem claims no energy saving of its own. AYON Wind and AYON Wave read a blend of public physics models and state no AI forecasting, as our audit of forecast apps recorded. Our next study tests an AI model and a physics model against real wind stations at kite beaches, and will report accuracy and, where it is published, the energy behind each forecast.

What would make this analysis wrong

The per-forecast figures are the agencies' own statements, not independent measurements, and the US figures describe computing resources, not metered energy. We read the life-cycle study's summary, not its full text, so its method and exclusions are reported only as its abstract states them. No published figure we found gives the energy of a forecast in absolute terms for both kinds of model. AI weather models change quickly, so each figure is dated by its source and this analysis will be updated when better measurements appear.

Evidence discipline

This is research analysis of published sources, not an audit. Company figures are the companies’ own disclosures; lab results describe their test conditions, not every product; projections are not measurements. Figures in this field change fast, so each one is dated by its source.

References & data

  1. 1.ECMWF — ECMWF's AI forecasts become operational (2025) — AI model operational: approximately 1,000 times less energy for making a forecast; grid spacing 28 km against 9 km for the physics model; gains of up to 20% on some measures; same initial conditions as the physics model.
  2. 2.ECMWF — ECMWF's ensemble AI forecasts become operational (Jul 2025) — AI ensemble operational, 1 July 2025: over 10 times faster; energy reduced by approximately 1,000 times; 31 km against 9 km; relies on physics-based data assimilation; physics ensemble remains indispensable for high-resolution fields.
  3. 3.HPCwire — NOAA deploys new generation of AI-driven global weather models (Dec 2025) — US AI models announced 17 December 2025: about 0.3% of the computing resources for a 16-day forecast, in approximately 40 minutes; 31-member AI ensemble at 9% of the computing; 62-member hybrid ensemble outperforms both; degraded tropical cyclone intensity in the first version.
  4. 4.Rieutord et al. — Energy and carbon footprint considerations for data-driven weather forecasting models (2026) — Life-cycle comparison of 7 data-driven models with ECMWF's physics model: training costs more than a physics forecast but is rapidly compensated; 21 to 1,273 times less energy over a year of use; low-resolution forecasts.
  5. 5.A checklist to assess the energy and carbon impacts of ML/AI applications in Earth system modeling (arXiv 2609.00847, Sep 2026) — Review from a Spanish supercomputing centre: of 12 frontier models, 3 report no compute figure and the rest only the final run; final runs may understate development compute by roughly 4 to 10 times; up to 30 times more emissions depending on the grid; a 10-day forecast in about 2 minutes 30 seconds on a single GPU.
  6. 6.AYON Ecosystem — Which forecast is your app using? — AYON Ecosystem's audit of forecast apps: AYON Wind and AYON Wave name their source, state no AI forecasting and publish no wind accuracy figures yet.
  7. 7.AYON Ecosystem — The hidden cost of a prompt — AYON Ecosystem's analysis of the energy cost of a single AI prompt and what multiplies it.