Google DeepMind's Cyclone-Forecasting Model Already Helped the NHC Call Hurricane Melissa's Landfall in Jamaica

Google DeepMind’s WeatherNext AI system, published in Nature and open-sourced with code and weights on GitHub, now produces three-day tropical-cyclone forecasts as accurate as what earlier systems could only manage two days out — roughly a full day of additional lead time on storm track, intensity, and wind structure. DeepMind calls it a decade’s worth of typical meteorological progress compressed into a single release. The model generates 1,000 possible storm-evolution scenarios per cyclone, up from 50 previously, and produces a full 15-day forecast in under a minute on a single TPU, using input data at roughly 100 times coarser resolution than traditional intensity models require. It isn’t theoretical: the US National Hurricane Center used WeatherNext during the 2025 Atlantic season to help forecast Hurricane Melissa’s rapid intensification ahead of its landfall in Jamaica, giving emergency planners more runway.

For any hospitality operation on a hurricane-exposed coastline — the Caribbean, the Gulf, Florida — an extra day of accurate lead time is not an abstraction. It’s the difference between an orderly pre-storm checkout and rebooking cycle and a scramble, between insurance and evacuation decisions made with margin and ones made against the clock. DeepMind isn’t alone in pushing this into the open, either: Europe’s ECMWF open-sourced its own AI forecasting system, AIFS, under a permissive license in July, using roughly 1,000 times less energy than traditional physics-based models — and Hugging Face has since shipped a compatibility patch that lets it run on any GPU or CPU rather than requiring expensive data-center hardware, aimed at putting state-of-the-art forecasting within reach of smaller operators, not just national weather agencies.

The pattern across both releases is the same: the best cyclone-forecasting models are moving from proprietary agency infrastructure toward open weights that a revenue-management or emergency-planning team could, in principle, run themselves. For coastal hospitality operators who currently rely on whatever lead time a paid weather-data vendor provides, that’s worth watching — the gap between the forecast a subscription buys and the forecast anyone can run is narrowing.