Google DeepMind Launches WeatherNext 3: AI Weather Forecasting Takes Aim at Grid Operators

Google DeepMind and Google Research announced on September 3, 2026 the deployment of WeatherNext 3, the third generation of their AI weather forecasting model. The technical jump is substantial: global resolution improved to 5 kilometres (from 25 km in the prior version), hourly updates (versus every six hours), and data latency cut to 3-4 hours from roughly 7. The model now specifically predicts wind speed at 100 metres — turbine hub height — along with cloud cover and surface sunlight.

On accuracy, Google claims improvements of up to 60% over NASA's IMERG satellite product, 30% over MRMS radar, and 50% on day-plus precipitation forecasts. The model plugs directly into Google Search, the Gemini app, Google Maps, and cloud building blocks BigQuery and Earth Engine — distribution reaching well beyond professional meteorologists.

The timing is not incidental. The US added more than 90 GW of new generation capacity in 2026, including 51.2 GW of solar and 25.7 GW of storage — resources whose output depends directly on weather conditions. Peak demand is projected to grow roughly 26% by 2035, driven in large part by data centres, whose electricity consumption could reach 176 GW by then — five times the 2024 level.

On this front Google is up against established specialists — Vaisala, Solcast, DNV's WindGEMINI, IBM's HyperWatch, and Swiss start-up Jua with its EPT-2 model — but with a distribution advantage those vendors lack: native integration into cloud infrastructure many energy operators already run on. Cardan-AI's companion analysis looks at why forecast accuracy is becoming mechanically more valuable as the grid decarbonizes and electrifies.