For fifty years, forecasting the weather required supercomputers solving the equations of the atmosphere for hours. In two years, machine learning models have reshuffled the deck: GenCast beats Europe's reference ensemble on 97.2% of targets and produces a global 15-day forecast in 8 minutes, ECMWF has been running its own AI model operationally since February 2025 with roughly 1,000 times less energy, and the National Hurricane Center ranked an AI model first for cyclone track and intensity in 2025. For companies whose demand, production or risks depend on the sky, this is a concrete change. Here is what it means, seen from Belgium.
The scientific tipping point came in December 2024: Google DeepMind published in Nature the GenCast model, trained on 40 years of ERA5 reanalysis data. Across 1,320 combinations of variables and lead times, GenCast beats ENS, the reference ensemble of the European Centre for Medium-Range Weather Forecasts (ECMWF), in 97.2% of cases, and in 99.8% of cases beyond 36 hours. Where a supercomputer grinds for hours, GenCast produces an ensemble of 50 or more scenarios over 15 days, at 0.25-degree resolution, in 8 minutes on a single TPU v5 chip. Its predecessor GraphCast, published in Science in late 2023, already beat the deterministic reference model on more than 90% of 1,380 targets.
The institutional shift followed within months. ECMWF, the European intergovernmental organisation of which Belgium is a member, took its own machine learning model into operations: AIFS Single on 25 February 2025, the first fully operational AI model at a major weather centre, with gains of up to 20% on tropical cyclone tracks and energy use per forecast cut by a factor of roughly 1,000; then AIFS ENS on 1 July 2025, an ensemble of 51 forecasts that outperforms physics-based models on many measures, including surface temperature, while running more than 10 times faster. Both systems run side by side with the physics-based IFS model, and their outputs are made available as open data.
The software giants are widening the field. Aurora, Microsoft's foundation model published in Nature in May 2025 and trained on more than one million hours of geophysical data, outperforms operational forecasts on air quality, ocean waves and cyclone tracks, at roughly 5,000 times lower computational cost; it beat the National Hurricane Center (NHC) on 5-day cyclone tracks for the 2022-2023 seasons. Google followed through: for the 2025 season, the NHC's verification report ranked WeatherNext the top-performing individual model for track and intensity; the model anticipated Hurricane Melissa's intensification to Category 5 five days ahead of its historic landfall in Jamaica. Since 18 November 2025, WeatherNext 2, 8 times faster and with hourly resolution, has powered forecasts in Google Search, Gemini and Pixel.
Belgium knows the price of a forecast taken seriously: the July 2021 floods killed 39 people and generated nearly 74,000 claims for €2.3 billion, the costliest natural disaster in the country's history (Assuralia). The World Meteorological Organization notes that 24 hours' advance warning can cut storm damage by up to 30%. As an ECMWF member state, Belgium benefits directly from AIFS, and its weather-exposed sectors (North Sea offshore wind, agriculture, construction, port logistics) are among the first potential users of forecasts that are more accurate, more frequent and cheaper.
Weather is not a niche topic: Munich Re puts natural catastrophe losses at $224 billion in 2025, 92% of them weather-related (and 97% of insured losses). Below the catastrophes, ordinary weather variability weighs every week on demand, lead times and costs. Better, cheaper forecasts translate directly into recurring decisions.
Three factors combine. Speed: an ensemble forecast that required hours of supercomputer time now comes out of a single chip in minutes, which allows scenarios to be multiplied (WeatherNext 2 generates hundreds) and refreshed more often. Cost: roughly 1,000 times less energy per forecast for AIFS; at constant budget, a national weather service or a company can produce far more. Access: AIFS outputs are open data, Aurora is published as open source, GenCast's code and weights have been released; state-of-the-art forecasting is no longer reserved for those who own a supercomputer. The humanitarian stake follows the same slope: only half of the world's countries have multi-hazard early warning systems, and the UN aims for universal coverage by the end of 2027 with its Early Warnings for All initiative; the World Bank estimates universal access to early warning would prevent at least $13 billion in asset losses every year.
AI models do not start from nothing: they depend on the data assimilation of physics-based centres for their initial conditions, and therefore on the satellites, radars and stations that feed it. Their resolution remains coarser (31 km for AIFS ENS versus 9 km for ECMWF's physics-based ensemble), which limits the very local: isolated thunderstorms, valley gusts, hail. And unprecedented events remain the hard spot for systems trained on the past. Weather centres speak of complementarity and hybrid systems, not replacement.
Which decisions depend on the weather: demand, production, sites, transport, safety? What did a bad forecast cost over 12 months? This quantification frames the project.
ECMWF open data (AIFS), services from the national weather institute, commercial APIs that integrate AI models: start with existing feeds before considering an in-house model.
Demand forecasting, staff planning, maintenance windows, internal alerts: one case, before/after indicators, one test season.
Use ensembles rather than a single scenario, compare AI and physics-based output, keep human validation for safety decisions, then extend to the other weather-sensitive decisions.
No, they complement them. AI models such as ECMWF's AIFS depend on physics-based data assimilation for their initial conditions, and their resolution (31 km for the AIFS ensemble) remains below that of the physics-based model (9 km). Weather centres run both side by side and are exploring hybrid systems that combine the strengths of both approaches.
GenCast (Google DeepMind, Nature, December 2024) beats ECMWF's reference ensemble on 97.2% of 1,320 verification targets, and on 99.8% beyond 36 hours. ECMWF's AIFS improves tropical cyclone tracks by up to 20%. For the 2025 cyclone season, the National Hurricane Center's verification report ranked Google's WeatherNext the top individual model for track and intensity.
In three steps: identify the weather exposure of the business (demand, production, logistics, sites, risks), access the forecasts (ECMWF's AIFS outputs are open data, Aurora is open source, commercial providers integrate them), then plug this data into recurring decisions: demand forecasting, staff and site planning, energy hedging, internal alerts. Measure the before/after effect on a limited perimeter, then extend.
They are progressing fast: WeatherNext anticipated Hurricane Melissa's intensification to Category 5 five days ahead in October 2025, and GenCast outperforms the European reference on extremes. But rare events remain the hard spot: models learn from the past, and an unprecedented phenomenon is by definition under-represented in it. Good practice remains the use of ensembles (50 or more scenarios), comparison with physics-based models, and human validation for safety decisions.
Molderez Consult helps Belgian companies quantify their weather exposure, plug AI forecasts (ECMWF open data, specialised APIs) into their demand, planning and risk decisions, and measure the gain, from the first indicator to the internal alert system.
Discuss my project