WeatherNext 2 by Google DeepMind

WeatherNext 3 is one of the few AI weather releases where the pitch matches the plumbing. Hourly global output is a genuine change from once-a-day numerical runs, and DeepMind's claim that accuracy holds in locations the model didn't see in training is the part that matters for anyone operating outside well-instrumented regions. Pick it when you're already on Google Cloud. BigQuery, Earth Engine and Cloud Storage access means no model setup and no separate data pipeline — you join forecasts to your own generation or logistics tables and run. Energy operators get the most obvious fit, since radiation and cloud cover are forecast variables rather than something you infer. Pass if what you actually need is a forecast console. There's no radar layer, no push alerts, no street-level nowcasting here, and the 5-10km grid is the wrong tool for hyper-local decisions. DeepMind is explicit that official warnings come from national meteorological agencies; Weather Lab is a research surface, not an operations product. Where it bites: this is data engineering territory. A small team without Cloud infrastructure or someone who can query gridded datasets will spend more time on plumbing than on forecasting. Budget for that before you assume the hard part is done. The closest alternative is operational NWP from a national met service or a commercial forecast vendor — slower to run, costlier to operate, but wrapped in services and support. WeatherNext competes on cadence and access, not on hand-holding. One caveat on currency: the family is more than one generation. WeatherNext 2 still sits on the DeepMind blog as a fast, accurate forecaster, so check which generation a given data surface is actually serving before you build around it.