Google's New AI Weather Model Updates Every Hour From Raw Satellite Data

A thunderstorm rolling onto the coast at 4 p.m. is not the same forecast problem as one rolling in at 7. For the last several years, the AI weather models that power some of the better apps on your phone have run on a roughly six-hour delay, because they learned from a finished forecast rather than the raw satellite imagery underneath it. Google DeepMind's new model, WeatherNext 3, closes most of that delay by reading the satellites directly, and the company says the result can refresh the entire planet every hour.

"AI weather model" is a phrase that gets used loosely. The ones that have shipped from DeepMind, Huawei, and others aren't simulations of the atmosphere. They're neural networks trained to look at inputs and produce a forecast as output, more like a translator than a physics engine. The question that distinguishes one from another is what the inputs are. Earlier DeepMind models and most competitors trained on "analysis" fields, the polished output of a traditional numerical weather prediction (NWP) system, the kind that has run for decades on supercomputers at places like the European Centre for Medium-Range Weather Forecasts. That data is high-quality but stale by the time an AI model consumes it, which is the roughly six-hour lag DeepMind attributes to the prior generation.

WeatherNext 3 takes a different approach. It ingests raw geostationary satellite observations directly alongside the analysis fields, and it was trained on sparse real-world weather station data as well, per the developer documentation. The model is still trained on analysis data, so it isn't a wholesale replacement for NWP, but the satellite stream is now a first-class input. That's the mechanism shift underneath the marketing.

The resolution numbers tell the other half of the story. According to the technical paper and developer docs, the model resolves surface variables such as temperature and moisture down to about 0.05 degrees, roughly 5 kilometers, and gridded surface fields at about 10 kilometers. Pressure-level winds are output at 0.25 degrees, around 25 kilometers. For a coastal squall line that flips a beach from sunny to flooded in twenty minutes, a 5-kilometer grid is the difference between "your town" and "the next town over." For a temperature gradient across a mountain valley, it's the difference between "bring a jacket" and "don't bother."

Cadence matters as much as resolution. WeatherNext 3 produces a new global forecast every hour, with 15-day outlooks at the four synoptic cycles (00, 06, 12, 18 UTC) and 48-hour outlooks for the interim hourly runs. A forecast that updates every hour, rather than every six, gives a normal person a better answer to the question most people are asking at 6 p.m.: is it going to rain at 9. DeepMind positions the gains as especially large in Latin America, Africa, and Asia-Pacific, where high-resolution NWP coverage has historically been sparser.

The accuracy claim that has drawn the most attention is also the one a reader should hold most loosely. Google says WeatherNext 3 delivers up to 50% more accurate precipitation forecasts at day-plus lead times, citing an "independent live evaluation." The evaluator isn't named in the press materials circulated so far, and "up to" is a number that picks its own examples. Earlier AI weather models, including DeepMind's own GraphCast, posted similar headline gains on specific metrics that didn't always translate into better forecasts for the variables most people care about. A 50% number on precipitation at one lead time and one region is a data point, not a verdict.

What is genuinely new is the data diet. The model isn't a new physics engine; it's a new way of feeding the old one. The next time any vendor, Google, Huawei, or whoever's next, claims a percentage improvement in forecast skill, the questions worth asking are the ones WeatherNext 3 makes unavoidable: what does the model train on, how often does it refresh, and how fine is the grid. The answer, increasingly, won't be in kilometers of benchmark score. It will be in whether the model can see what is happening in the sky right now, or whether it's still paraphrasing yesterday's forecast.