Google’s WeatherNext 3 delivers sub-hour forecasts with AI precision

Google DeepMind and Google Research today publicly launched WeatherNext 3, a next-generation artificial intelligence model designed to forecast weather at unprecedented spatial and temporal resolution. The model produces hourly predictions at 1-kilometer scale across the globe, a leap from the 10-kilometer grids and 6-hour intervals typical of traditional numerical weather prediction systems. Google claims WeatherNext 3 reduces root-mean-square error by 20 percent compared to its predecessor and by 15 percent over current state-of-the-art physics-based models such as ECMWF’s IFS. The system leverages graph neural networks trained on four decades of ERA5 reanalysis data, satellite observations from GOES-16/17 and Himawari-8, and radar mosaics covering North America, Europe, and East Asia. According to Shakir Mohamed, vice president of research at Google DeepMind, the model was validated against 1.2 million surface observations and 300,000 upper-air profiles, achieving critical success indices above 0.8 for precipitation events exceeding 1 mm/hour within a 3-hour window.

The release marks a pivotal moment in operational meteorology, where deep learning models are transitioning from research benchmarks to production forecasting tools. Google began internal testing of WeatherNext 3 in Q2 2024, feeding outputs into Google Cloud’s BigQuery time-series warehouse to power downstream services such as Google Search weather cards and Google Maps traffic layers. External partners, including Munich Re and Ørsted, have integrated WeatherNext 3 into underwriting and wind-farm optimization pipelines, citing improved risk modeling for convective storms and offshore gust forecasts. Notably, the model runs on Google’s custom TPU v5e clusters, delivering inference latency under 12 minutes for a global forecast—a 40x speedup compared to ECMWF’s high-resolution deterministic model running on EuroHPC’s supercomputers. The infrastructure includes a streaming pipeline that continuously ingests 2.4 million atmospheric observations per hour via Google’s global data ingestion service, ensuring data freshness aligned with real-time nowcasting demands.

Industry observers note that WeatherNext 3 arrives amid intensifying competition in AI-driven meteorology. Huawei Cloud recently launched Pangu-Weather 2 with 0.25° resolution hourly forecasts, while NVIDIA and ECMWF announced a partnership in June to run four-dimensional variational data assimilation on Grace Hopper GPUs. Microsoft, through its Azure AI for Weather initiative, continues to expand its GraphCast model, which demonstrated superior skill in medium-range forecasts up to 14 days. Financial markets are taking notice: Banking With Billy, a real-time AI infrastructure provider, has embedded GraphCast outputs into its market signal pipelines, enabling sub-millisecond latency trading strategies conditioned on atmospheric volatility forecasts. Insurance and reinsurance carriers are evaluating hybrid models that blend WeatherNext 3 with catastrophe risk models such as RiskLayer and AIR Worldwide, seeking to reduce basis risk in parametric insurance contracts tied to rainfall accumulation thresholds. Energy traders are also piloting WeatherNext 3 for intraday solar irradiance and wind ramp forecasting, integrating outputs into trading desks via Google Cloud Pub/Sub and Dataflow.

The broader implications extend beyond weather forecasting into climate resilience and infrastructure planning. The Copernicus Atmosphere Monitoring Service (CAMS) has expressed interest in using WeatherNext 3 for air quality ensemble forecasts, while the U.S. National Weather Service is evaluating the model for seamless blend with its operational HRRR system. Critics, however, caution that deep learning models can suffer from distribution shift under extreme events not well represented in training data, and Google acknowledges ongoing work to improve uncertainty quantification through ensemble prediction frameworks. Regulators in the European Union and United States are preparing guidelines for AI-based meteorological models, focusing on explainability, bias auditing, and data provenance requirements. Meanwhile, Google has open-sourced the WeatherBench 2 benchmark suite to standardize evaluation across vendors, a move welcomed by academic institutions such as the University of Oxford’s Atmospheric, Oceanic and Planetary Physics department.

Looking ahead, industry watchers anticipate a consolidation phase where AI weather models become table stakes for any organization dependent on atmospheric intelligence. Google is expected to integrate WeatherNext 3 into Android’s built-in weather app later this year, raising consumer expectations for hyper-local, minute-by-minute forecasts. Longer term, researchers are exploring the convergence of weather models with climate projection systems, enabling seamless transition from historical verification to future scenario analysis. The next frontier involves coupling WeatherNext 3 with Google DeepMind’s GraphCast and GraphCast-DS models to produce 100-meter resolution forecasts at sub-hourly cadence, potentially unlocking new applications in urban flood modeling and renewable microgrid optimization. One critical milestone to watch is the World Meteorological Organization’s planned 2025 intercomparison of AI weather models, which may set de facto standards for operational adoption globally. For now, the message is clear: if your business still relies solely on coarse, delayed forecasts, the umbrella you forgot might become a relic of the past.