Google has unveiled the third iteration of its AI-based weather forecasting model — WeatherNext 3. Unlike its predecessors, this product is tailored to the specific needs of the energy sector, where the accuracy of meteorological data directly translates into financial results.
New Variables and a Technical Breakthrough
The key difference of WeatherNext 3 is an expanded set of parameters it takes into account. The model now analyzes wind speed at a height of 100 meters, cloud cover levels, and the intensity of solar radiation at the surface. This data is critical for assessing the output of renewable energy sources (RES) and forecasting load on power grids.
The technical specifications are impressive: the global forecast is updated hourly, and spatial resolution has been improved to 5 km for surface variables (temperature, humidity) and 11 km for precipitation. For comparison, the previous version, WeatherNext 2, operated on a 25 km grid and updated data only once every six hours.
Training on Real Data
The architecture of the new model is fundamentally different. Instead of relying solely on simulations, WeatherNext 3 is trained directly on data from weather stations and current satellite imagery. This approach has reduced information latency from 7 to 3–4 hours, as confirmed by a senior researcher at Google DeepMind. This means traders and grid operators receive a fresher picture for decision-making.
The progress in precipitation forecasting accuracy is particularly noticeable:
- up to 60% improvement relative to the NASA IMERG reference dataset;
- up to 30% compared to the American MRMS system;
- up to 10% when cross-checked against rain gauge data at early forecast horizons.
Integration and Market Context
WeatherNext 3 is already integrated into Google's ecosystem: Search, Gemini, Google Maps, and the Weather API. Corporate clients have gained access via BigQuery, Earth Engine, and Google Cloud Storage, removing the barrier to adoption without the need for their own infrastructure deployment. The subscription cost for businesses has not yet been disclosed.
Interest in such solutions is fueled by structural changes in the energy sector. According to industry estimates, in 2026, solar generation and storage will become the main sources of new capacity in the U.S. — 51.2 GW and 25.7 GW, respectively, out of more than 90 GW planned. The cost of forecast errors is enormous: underestimating wind generation forces emergency purchases of energy from gas plants, while overestimation leads to forced curtailment of output.
In this arena, Google competes with Vaisala, Solcast, DNV (WindGEMINI), IBM HyperWatch, and Switzerland's Jua. However, for the corporation itself, WeatherNext 3 is also a tool for internal optimization: Google is actively purchasing "clean" energy for its data centers, and accurate forecasts directly reduce its operational costs.
My analysis: The release of WeatherNext 3 is a strategic move aimed at monetizing AI infrastructure in a niche where demand for accuracy is growing exponentially. However, the reliance on purely empirical training carries risks during extreme climate anomalies, where historical data loses relevance. A hybrid approach combining the physical laws of the atmosphere with machine learning is likely to become the next stage in the evolution of such systems.