DeepMind Pitches WeatherNext 3 Power Output Forecasts to Grid Operators

The model itself arrived on September 3. Google DeepMind and Google Research introduced WeatherNext 3 in that day's announcement as the most advanced global weather AI model to date, citing independent live evaluations by Brightband. According to the announcement, key variables such as temperature and humidity are forecast on a 5-kilometer grid, other surface variables including wind on a 10-kilometer grid, and atmospheric variables on a 25-kilometer grid. The company says the overall picture is roughly five times sharper than the previous model, WeatherNext 2, which ran on a 25-kilometer grid in 6-hour steps.

In the announcement, which METAL read in full, the company singled out not the resolution but what the model learns from as the biggest leap. Most AI weather models train on the output of physics simulations run on supercomputers, and that data carries a six-hour lag. WeatherNext 3 ingests live imagery from geostationary satellites to produce a fresh forecast every hour, and learns temperature and humidity directly from weather station observations so its 5-kilometer grid reflects local topography. The announcement says this matters most for Latin America, Africa and Asia-Pacific, regions that have historically gone without high-resolution forecasting.

The paper released the same day spells the design out in more detail. The paper, which METAL reviewed, runs 43 pages, and its authors are 25 researchers from Google DeepMind, Google Research and Google, led by Stephan Rasp. According to the paper, the model produces 15-day forecasts as a 64-member ensemble and outputs single-level variables, including solar radiation and cloud cover, on a 0.1-degree grid every hour. Generating a single forecast takes 6.3 minutes on four TPUv5p chips. At the top of the paper the authors write that "WeatherNext 3's capabilities move operational AI-based weather forecasting beyond emulating the traditionally distinct stages of data assimilation, forecasting and post-processing."

The paper also puts a number on how much further ahead the hourly refresh lets users see. Compared with a 6-hourly forecast, the hourly forecasts gain two to three hours of lead time in the early part of the forecast, because the average latency of satellite observations is roughly three hours shorter than that of a 6-hourly analysis. According to reports, DeepMind senior research scientist Ilan Price said, "It gets much more accurate by not waiting for the next analysis date and using the most recent information," and the data lag has fallen from around seven hours to three to four.

The precipitation figures sit separately in the announcement. The company says it trained on NASA's satellite precipitation product IMERG and its own precipitation reanalysis, and reports medium-range score improvements of up to 60 percent against IMERG, 30 percent against the radar-based MRMS, and 10 percent against rain gauges. The three are different baselines and do not add up. The same announcement says that in Google Search, the Gemini app and Google Maps, precipitation forecasts a day or more ahead become up to 50 percent more accurate.

Power companies watch these numbers because of what a wrong forecast costs. According to reports, if an operator underestimates wind output it has to buy replacement electricity at short notice from gas plants kept on standby, and if it overestimates, the grid cannot absorb the supply and wind and solar farms end up being paid to switch off. Both are costs created by forecasting failures. Because renewables generate according to the weather rather than demand, every new gigawatt of capacity raises the value of a short-term forecast.

The demand side is shifting too. According to reports, S&P Global's 2026 US grid outlook projects solar at 51.2 gigawatts and storage at 25.7 gigawatts as the main additions among more than 90 gigawatts of planned new capacity this year, and names the spread of data centers across North America as the primary driver of the recent surge in electricity demand. Deloitte projects peak demand growing by roughly 26 percent by 2035, with data center demand alone potentially reaching 176 gigawatts. A 2024 International Energy Agency report found that failing to integrate variable renewables properly could leave solar and wind generation 15 percent below projections in 2030, roughly 2,000 terawatt-hours.

Companies already sell into this market. According to reports, Vaisala, Solcast, DNV's WindGEMINI and IBM's HyperWatch sell weather data to the power sector, and the Swiss firm Jua promotes its own model that updates 24 times a day. What sets Google apart is that the same forecast lands all at once as a table in BigQuery, a layer in Earth Engine, an API in Google Maps Platform, and the default answer in Google Search. The hourly refresh narrows the update-frequency gap the specialist vendors have used to set themselves apart.

Google is also a party to the problem. According to reports, the data center build-out driving the load growth is led by the hyperscalers, Google among them, and Google has signed multi-gigawatt renewable procurement agreements to supply its own facilities. For a company that has to match large volumes of clean power against a load that is both growing and variable, forecasting wind and solar output is immediately useful. That commercial logic goes some way to explaining why the energy variables shipped in this release.

METAL reported earlier on this model family catching a hurricane's rapid intensification a day ahead, and the paper cites the cyclone forecasting results of the 2025 hurricane season as this model's starting point. The announcement closes with a line telling readers to consult their national weather service for official forecasts and warnings, and the paper also records that individual ensemble members show honeycomb-shaped artifacts. Delivering a forecast that is rewritten every hour to the grid operator's desk: that is the story DeepMind pulled back out with its September 14 thread.