Google DeepMind's WeatherNext 3 delivers hourly global forecasts from raw satellite data

Google DeepMind has released WeatherNext 3, which it bills as the first global weather AI model capable of producing forecasts every hour rather than at coarser daily intervals. The key architectural shift is that the model ingests raw satellite imagery directly, rather than relying solely on the numerical physics-based predictions that have driven forecasting for decades. That approach lets it track fast-moving conditions like rain and snow and, according to Google, produce more accurate results even for locations absent from its training data.

The model is an ensemble that predicts surface variables at relatively high resolution — temperature and humidity down to 5km, wind and other variables to 10km. Beyond consumer use, it targets industry: forecasting radiation and cloud cover for wind and solar farm operators managing renewable output. Google is folding WeatherNext 3 into Search, Maps, and Gemini, while also offering it to enterprises through BigQuery, Earth Engine, the Google Maps Platform, and Cloud Storage, along with an interactive platform for exploring live forecast layers and tracking tropical cyclones.

The significance is competitive as much as scientific: hourly, satellite-driven AI forecasting undercuts the cost and complexity of traditional numerical weather models and pushes forecasting deeper into Google’s product stack and cloud offerings. Google frames it as an experimental research platform, cautioning that official warnings should still come from national meteorological agencies.