Weather forecasting has a new champion, and it does not solve a single physics equation. Google DeepMind’s GenCast, described in Nature in December 2024, is an AI model that learned from four decades of weather records and now predicts global weather more accurately than the ensemble system run by the European Centre for Medium-Range Weather Forecasts, long regarded as the world leader. In the authors’ evaluation, GenCast had greater skill than the European system on 97.2 percent of the 1,320 targets measured, and it was particularly strong at predicting extreme weather and tropical cyclone tracks. A full 15-day forecast takes about eight minutes on a single machine-learning chip, against hours on a supercomputer for a traditional model. The paper, published in Nature, marked the moment AI forecasting stopped being a laboratory curiosity and became a serious operational tool.
What the Evidence Shows
GenCast works differently from its predecessor GraphCast, which produced a single best guess. It is a diffusion model that generates fifty or more possible versions of the coming weather, giving forecasters a probability for each outcome rather than one confident answer. That matters for decisions such as whether to evacuate a coastline or open flood defences, where knowing the odds is more useful than a single prediction. The technology is now moving into everyday use: the European centre’s own AI system went operational in February 2025, NOAA runs AI models alongside its physics systems, and Google’s newer WeatherNext models power weather in Search, Gemini and Maps. Forecasters stress that AI still leans on physics-based systems for the snapshot of today’s atmosphere, so the two approaches are partners rather than rivals. Reporting and analysis from sources such as IEEE Spectrum track the same developments, and specialists following AI weather forecasting GenCast say the direction is now consistent across independent lines of evidence.
The Wider Context
Adoption of AI weather forecasting GenCast follows a familiar pattern with an unfamiliar speed. Early users tolerate rough edges in exchange for advantage, mainstream users arrive when the tool disappears into existing workflows, and the market consolidates around whoever made that transition feel effortless. Vendors are racing less on raw performance now than on integration, support and trust.
What Happens Next
The realistic forecast for AI weather forecasting GenCast is steady integration rather than sudden replacement: embedded in products people already use, improving in increments, and regulated in proportion to its reach. The companies and institutions that plan on that basis, training staff and redesigning processes instead of chasing launches, are the ones the adoption data tends to vindicate. You can follow related developments in our coverage of technology.
It is also worth remembering how quickly the picture in this area has changed before. Positions that looked settled a few years ago have been revised as new data arrived, and the same humility is appropriate now. The strongest coverage of this subject, and the most useful advice, tends to come from sources that update their views openly when the facts move, rather than defending an earlier position long after it has stopped fitting the evidence available.
