Google says its AI weather model is getting better

Justine Calma reports Google is rolling out WeatherNext 3, pitched as more accurate especially on rain and snow, with 'unprecedented resolution' from real-time observations. Research engineer Samier Merchant says the main development is going beyond the data most global AI models train on, using fresher observational datasets. WeatherNext 3 produces a forecast each hour from the latest satellite observations. WeatherNext 2 ran every six hours on a 25-kilometer grid; the new model can visualize temperature and moisture at up to 5 kilometers.

Calma flags that satellite input also fills gaps where rain gauges are sparse, primarily outside the U.S. and Europe, which Google says is where improvements will be largest. Users are promised precipitation forecasts up to 50 percent more accurate at least a day ahead. The model also forecasts 100-meter wind speeds, about turbine height. Ferran Alet of DeepMind told The Verge that as Google's energy needs — and humanity's — rise, making renewables appealing matters. WeatherNext 3 is in Search, Maps, Gemini, and other products. Google has worked with the U.S. National Hurricane Center and agencies in Asia. Calma notes the model is still trained on physics-based data and that agencies still look at a range of predictions for warnings.

Calma's angle is the right one for a science desk: skill where the observing system is thin, and a power-hungry company building turbine-height winds into the same model that tells you to bring an umbrella. The 50 percent precipitation figure is Google's, aimed at consumer day-ahead plans. The equity claim — biggest gains outside the U.S. and Europe — is the part that would matter if it survives contact with tropical convection, not a UK 2-meter temperature graphic.

Alet's energy quote is unusually blunt. Google is not a weather NGO. It is a buyer of electrons. High-resolution wind and cloud products that make renewables 'appealing' are aligned with data-center growth. That does not make the science fake. It does mean the product roadmap has an internal customer.

Signal Desk's read: keep the tandem-with-physics caveat in the foreground. Calma is careful: WeatherNext 3 still trains on NWP-style data, and warning responsibility stays with agencies. Google is inserting an hourly, sharper layer into consumer apps and calling it a revolution. The revolution, if it exists, is distribution plus satellite cadence, not the retirement of physics codes.

The NHC and Asia agency collaborations are how you get legitimacy. Watch whether those agencies start citing WeatherNext in public products, or only appear as name-drops in a Google launch.

Physics-based global models remain the backbone of official warnings and carry a time lag. AI models skip some of that cost by learning patterns. WeatherNext 3's bet is that live satellite data can close the remaining gap on fast, local precipitation — the failure mode that makes people distrust the app.