News · 2026-09-05

Google's WeatherNext 3 shifts global AI forecasts to an hourly refresh

Google DeepMind's WeatherNext 3 generates global weather forecasts every hour, changing the operational refresh loop from a six-hour rhythm to an hourly one. The advance matters because timely weather information is often limited less by an algorithm's one-shot accuracy than by how quickly it can absorb fresh observations and produce another forecast.

Key facts

• Google DeepMind says WeatherNext 3 generates forecasts every hour.
• The associated paper says the system uses low-latency geostationary satellite observations directly.
• The published guide lists 5 km station-trained temperature/dew point, 10 km surface fields and 100 m wind products.
• Google documents requestable access through Cloud Storage, BigQuery and Earth Engine.

Earlier AI weather systems commonly started from an analysis: a carefully assembled estimate of the atmosphere produced by combining observations with conventional numerical forecasting. Those products are exceptionally useful, but they arrive on a cadence. WeatherNext 3's paper, titled WeatherNext 3: Increasing resolution and performance of global weather models with raw observations, says the new model takes low-latency geostationary satellite data directly and forecasts on hourly initialization times. That lets the system react sooner to what satellites are seeing.

It would be a mistake to call it a raw-data-only model. Google's model guide also lists ECMWF HRES analysis as an input. The real change is a hybrid one: it uses direct observations without being wholly gated by waiting for a new analysis product. Think of a weather office that previously received a polished report four times per day and now also gets a fresh continuous camera feed. The report remains valuable; the camera makes the response loop faster.

The published products show where the benefit is most concrete. Google lists 0.05-degree, roughly 5 km, station-trained two-metre temperature and dew point, as well as 0.1-degree, roughly 10 km, gridded surface wind, pressure, sea-surface temperature, cloud, solar-radiation and precipitation outputs. A 100 m wind product is intended for energy applications. Google is integrating WeatherNext into Search, Maps and Gemini, while developers can request data access with a Google account; no paid Cloud contract is required before allowlisting, according to the quick-start page.

The concrete headline number is the hourly refresh. But the limitation is just as important. Google's benefits and limitations guide says pressure-level output remains 0.25 degree, around 25 km, and six-hourly. WeatherNext models also inherit biases from global reanalysis data, with station training only partly reducing that issue. Hourly does not mean every weather variable at every altitude has suddenly become hourly and high resolution.

External domain experts see the input-path change as consequential. In its AI-DOP article, ECMWF calls forecasts made directly from observations 'a highly significant milestone' and 'a radical departure' from analysis-initialized systems. That is not a blanket endorsement of every WeatherNext metric; it identifies why the operational design is novel.

The honest counterargument is that weather forecasting is an end-to-end discipline, not a leaderboard. A new model must be assessed through live storms, calibration, regional failure cases, communication to users and comparison with physics-based ensembles. Google itself acknowledges data and upper-air limitations. The story is therefore not that AI has replaced conventional weather prediction. It is that the data-refresh bottleneck is being attacked directly. For energy, emergency management and consumer forecasts, one more fresh update can matter as much as a small average score improvement.

Primary source, verified: read the paper → (arXiv 2609.03582)

Ground Truth. (2026, September 5). Google's WeatherNext 3 shifts global AI forecasts to an hourly refresh. Ground Truth. https://groundtruth.day/news/weathernext-3-hourly-direct-observation-forecasts.html

@misc{groundtruth:weathernext-3-hourly-direct-observation-forecasts,
  title  = {Google's WeatherNext 3 shifts global AI forecasts to an hourly refresh},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/weathernext-3-hourly-direct-observation-forecasts.html}
}

Topics: weather · google-deepmind · science · forecasting · geospatial