Summary
Google DeepMind has announced WeatherNext 3, which it describes as the first global operational AI weather model capable of producing a new forecast every hour with up to 5 km spatial resolution and hourly time steps. Traditional numerical weather prediction models refresh every six hours at most because they simulate atmospheric physics step by step — a process that is both slow and computationally expensive at high resolution. WeatherNext 3 sidesteps this by learning directly from historical atmospheric observations and, crucially, ingesting raw satellite data rather than pre-processed analysis products, which is what enables the hourly update cadence.
The model delivers three native resolutions in a single pass: 25 km for broad atmospheric patterns, 9 km for surface variables like wind and pressure, and 5 km for temperature and humidity. For renewable energy applications, the team added wind speed and direction at 100 meters — the typical operating height of a wind turbine — plus cloud cover and solar radiation estimates for solar farm management. The video frames high local resolution as particularly consequential for coastal and mountainous regions where weather can shift substantially over a few kilometers.
WeatherNext 3 will roll out across Google Search, Gemini, and Maps, giving the model direct reach to billions of users. The team cites precision agriculture, flood evacuation timing, and grid-scale renewable energy management as the highest-impact downstream applications. DeepMind positions the model as closing the historical tradeoff between global coverage and local resolution that constrained both traditional and earlier AI forecasting approaches.
📺 Source: Google DeepMind · Published September 03, 2026
🏷️ Format: Keynote Launch
