Google and DeepMind just threw out the physics textbook. WeatherNext 3, their newest AI weather model, skips traditional numerical weather prediction entirely and learns straight from live geostationary satellite data. That’s a genuine departure from how weather models have worked for decades — and from Google’s own WeatherNext 2, which trained on NWP outputs that come with a built-in six-hour delay thanks to supercomputer processing time.
That lag matters. Fast-changing variables like rainfall and temperature drift in those six hours. WeatherNext 3 closes the gap by generating hourly forecasts at five-kilometer resolution for temperature and humidity — roughly five times sharper than its predecessor’s 25-kilometer grid. Google’s comparison over the UK tells the story: a pixelated map versus one that actually resolves valleys and coastlines. The model also trains on individual weather station data, which means regions in Latin America, Africa, and Asia-Pacific that couldn’t afford expensive regional models finally get forecasts tuned to local terrain.
Precipitation has always been AI weather’s weak spot. Rain and snow depend on fast cloud processes at small scales, where global models fall apart. Google attacked this by training on NASA’s IMERG satellite dataset alongside its own satellite radar analysis. The Continuous Ranked Probability Score shows improvements up to 60 percent over IMERG and 30 percent over MRMS. For anyone checking whether to carry an umbrella, that’s up to 50 percent better precipitation forecasts in Google Search, Maps, and Gemini. The model also predicts wind speeds at 100 meters — turbine height — and improved solar irradiance, giving grid operators harder numbers for balancing renewable supply and demand.
This builds on a fast-moving lineage. WeatherNext 2 arrived in November 2025, beating the first generation on 99.9 percent of meteorological variables while running eight times faster. GenCast, released in late 2024, was the first probabilistic AI model to outperform ECMWF’s ensemble. The speed of iteration is striking — three major model generations in under two years. Google is being honest about limits, though: the atmosphere remains fundamentally unpredictable, and they still point users to national weather services for severe warnings. The real question is whether physics-free learning holds up in extreme events it has never seen. Satellite data only shows what has already happened, not what a warming atmosphere might do next.
