Google WeatherNext 3 skips preprocessing and reads raw satellite data directly

Google has introduced WeatherNext 3, a model that cuts the time required to issue a forecast from the usual six down to three or four hours. It does this by cutting out a step that meteorology has historically relied upon.

The end of reliance on data assimilation

Previous AI models worked with data that had already been processed. Meteorological institutes have to take raw signals from satellites and radars and complexly recalculate (assimilate) them into a grid representing the state of the atmosphere. WeatherNext 3 learns directly from raw satellite images and reports from ground stations, lowering latency.

Hardware requirements shift for local deployment

By bypassing traditional assimilation systems running on government supercomputers, the model alters the barrier for commercial entities. Logistics companies or farmers could theoretically generate their own forecasts much closer to real time, provided they can access the stream of raw satellite data.

Physical anomalies remain a black box

Removing physical assimilation comes at a price. When you feed raw data directly into the model, you lose the control layer that checks whether the input violates basic laws of physics. If a sensor starts sending noise, the AI has to deal with it on its own.

Reliability during extreme events will prove the real value

Routine daily operations will be faster. The key question is whether the model, without physical “supervision”, can maintain accuracy during hurricanes and rapid weather changes, where raw data often contain massive fluctuations.

Lilith's verdict

Google didn’t optimize weather forecasting. It turned an expensive government monopoly on data cleaning into a software problem.