In the run-up to the winter storm currently hitting much of the United States, weather forecasts in some areas have changed across the board, and snowfall forecasts have fluctuated wildly.
Nvidia couldn’t have released a new Earth-2 weather forecast model at a better time. Or perhaps the company knew something we didn’t, given how accurate the company claims its new model is?
New AI models promise to make weather forecasts faster and more accurate. Nvidia claims that one model in particular, Earth-2 Medium Range, outperforms Google DeepMind’s AI weather model GenCast on more than 70 variables. GenCast itself, which Google released in December 2024, was much more accurate than existing weather models, which can generate forecasts up to 15 days in advance.
Nvidia announced the new tool Monday at the American Meteorological Society meeting in Houston.
“Philosophically and scientifically, this is a return to simplicity,” Mike Pritchard, Nvidia’s director of climate simulation, told reporters on the phone before the conference. “We are leaning away from hand-customized, niche AI architectures and toward a future of simple, scalable Transformer architectures.”
Traditionally, most weather forecasts rely on simulations of the physics observed in the real world. AI models are a relatively recent addition. The Earth-2 medium-range model is based on a new Nvidia architecture called Atlas, and the company said it will announce details on Monday.
In addition to the Medium Range, Nvidia’s Earth-2 suite includes Nowcasting and Global Data Assimilation models.
tech crunch event
san francisco
|
October 13-15, 2026
Nowcasting is intended to produce short-term forecasts from 0 to 6 hours in advance to help meteorologists predict the effects of storms and other hazardous weather.
“Because the model is trained directly on globally available geostationary satellite observations, rather than the output of region-specific physical models, the nowcasting approach can be applied anywhere on Earth with good satellite coverage,” Pritchard said. This should help state and small governments understand how severe weather systems will affect their territories.
Global data assimilation models use data from sources such as weather stations and balloons to create continuous snapshots of weather conditions at thousands of locations around the world. These snapshots are used by weather models as a starting point to make predictions.
Traditionally, these snapshots require significant amounts of computing power before any predictive work can begin. “It consumes about 50% of the total traditional weather (forecasting) supercomputing load,” Pritchard said. “This model can run it in minutes on a GPU instead of hours on a supercomputer.”
The three new models join two existing models. CorrDiff uses coarse-grained forecasts to generate fast, high-resolution forecasts, and the other is FourCastNet3, which models individual weather variables such as temperature, wind, and humidity.
Pritchard said the new model will give more users access to powerful weather forecasting tools. This has historically been the domain of wealthy countries and large corporations who can afford to spend expensive time on supercomputers.
“This provides the basic building blocks that are used by everyone in the ecosystem, including national weather services, financial services companies, energy companies, and anyone else who wants to build and improve weather forecasting models,” Pritchard said. Some tools are already in use. Meteorologists in Israel and Taiwan, for example, are using Earth-2 CorrDiff, and The Weather Company and Total Energies are evaluating nowcasting, according to Nvidia.
“For some users, it makes sense to subscribe to a company’s centralized weather forecasting system. But for other users, such as countries, sovereignty is important,” Pritchard said. “Weather is a national security issue, and sovereignty and weather are inseparable.”
