Technology
Google Launches WeatherNext 3, Bringing AI Forecasting to Energy Grid Operators
Google has released a new AI weather forecasting model, WeatherNext 3, that predicts wind speed at 100 metres above ground level—approximately the height of a modern wind turbine—along with cloud cover and surface sunlight levels, updating every hour. The launch positions Google to compete directly with established providers of weather data to energy traders, grid operators, and renewable energy developers.
Google DeepMind and Google Research released the model on September 3. It generates a global forecast every hour at resolutions of up to five kilometres for surface variables including temperature and moisture. This marks a significant improvement over the previous version, WeatherNext 2, which operated on a 25-kilometre grid and updated only every six hours.
According to Google, the new energy-specific variables are designed to help grid operators and renewable energy developers forecast power generation from wind and solar assets and align it with demand.
While much attention has focused on the consumer-facing rollout—WeatherNext 3 now powers weather information in Google Search, the Gemini app, Google Maps, and the Google Maps Platform Weather API—the model's enterprise capabilities carry greater commercial significance. The forecast data can be accessed through BigQuery and Earth Engine, or downloaded in bulk from Google Cloud Storage, without requiring customers to configure their own models.
Growing demand for AI-driven forecasting
Grid operators face increasing unpredictability on both sides of the energy equation. On the supply side, renewables now represent the bulk of new generation capacity. S&P Global Market Intelligence's US Grid Outlook 2026 forecasts that solar and energy storage will lead new capacity additions this year, at 51.2GW and 25.7GW respectively, out of more than 90GW in planned additions. Because solar and wind output depends on weather conditions rather than demand patterns, accurate short-term forecasting becomes more critical as renewable capacity expands.
On the demand side, the expansion of AI infrastructure is driving unprecedented load growth. S&P Global identifies the proliferation of data centres across North America as a key factor pushing utilities to revise electricity demand forecasts upward. Deloitte's 2026 Power and Utilities Industry Outlook projects that peak demand could increase by approximately 26% by 2035, with data centre demand potentially reaching 176GW—five times 2024 levels.
Forecasting errors carry real costs. If grid operators underestimate wind power generation, they must purchase backup electricity on short notice, typically from gas plants held in expensive reserve. Overestimating output can force wind and solar farms to curtail production because the grid cannot absorb the excess supply. Both scenarios are costly and stem from forecasting inaccuracies.
A competitive market
Google enters a market with established players. Companies including Vaisala, Solcast, DNV's WindGEMINI, and IBM's HyperWatch already provide weather forecasting services to the energy sector. Swiss firm Jua also competes in this space, claiming its EPT-2 model outperforms Microsoft Aurora and DeepMind's earlier GraphCast model in accuracy while updating 24 times daily—compared to what it describes as a typical four daily updates among competitors.
Google's competitive advantage lies in its reach. The same forecast data appears across BigQuery, Earth Engine, the Google Maps Platform API, and as the default weather result in Google Search. No specialist competitor offers this breadth of integration, and the hourly update frequency narrows a gap that rival vendors have used to differentiate their offerings.
However, incumbents maintain one technical argument. Jua contends that physics-based models, such as the European Centre for Medium-Range Weather Forecasts' HRES, still outperform purely data-driven AI models during extreme, record-breaking weather events, because physics-based systems incorporate established rules governing atmospheric behavior, whereas AI models rely on patterns learned from historical data. Jua, which sells a physics-constrained product, has a vested interest in this argument, though it does highlight legitimate concerns for grid operators regarding storms that fall outside a model's training data.
Evaluating the technical claims
A central claim behind WeatherNext 3 is that it learns from live observations rather than relying solely on simulations. Most AI weather models, including WeatherNext 2, are trained using output from numerical weather prediction systems—physics-based simulations run on supercomputers that carry a data lag of roughly six hours. This delay can introduce inaccuracies in fast-changing variables like precipitation and surface temperature.
WeatherNext 3 incorporates live geostationary satellite imagery and trains directly on data from individual weather stations. However, the shift is more limited than some reporting has suggested. Google's own system architecture shows the model still uses one-hour satellite mosaics alongside traditional historical analysis. Ilan Price, a senior research scientist at DeepMind, told Bloomberg that performance gains stem from using the most recent available information rather than waiting for the next data analysis cycle. Reports indicate the remaining data lag has been reduced to three to four hours, down from approximately seven—meaning dependence on numerical weather prediction has been reduced but not eliminated.
Google's accuracy claims also warrant careful interpretation. The company reports improvements of up to 60% compared to NASA's IMERG satellite product, 30% compared to MRMS radar data, and 10% compared to rain gauge readings at early forecast lead times, using a standard probabilistic scoring method. These figures represent three distinct baselines and should not be combined. The widely cited claim of 50% improved precipitation forecasting specifically applies to forecasts made a day or more in advance. Each reported figure includes an