As a winter storm swept across large parts of the United States, weather forecasts in the days leading up to it varied widely. In some regions, snowfall projections fluctuated dramatically, highlighting the limits of even modern forecasting systems.

 

Against that backdrop, Nvidia rolled out a new set of AI-powered weather forecasting models. The company says the tools are designed to make forecasts faster and more accurate than existing approaches.

 

According to Nvidia, its Earth-2 weather models represent a significant advance in AI-driven forecasting. One model in particular, Earth-2 Medium Range, outperforms Google DeepMind’s GenCast across more than 70 variables, the company said. GenCast, released in December 2024, was itself a notable improvement over traditional forecasting systems, capable of producing more accurate forecasts up to 15 days in advance.

 

Nvidia announced the Earth-2 tools Monday at the American Meteorological Society’s annual meeting in Houston.

 

“Philosophically and scientifically, this is a return to simplicity,” said Mike Pritchard, Nvidia’s director of climate simulation, during a press call ahead of the event. “We’re moving away from hand-tailored, niche AI architectures and leaning into simple, scalable transformer architectures.”

 

Most weather forecasts today are still based primarily on physics-based simulations of the atmosphere, with AI playing a secondary role. Earth-2 Medium Range, however, is built on a new Nvidia architecture called Atlas, about which the company said it would share additional details.

 

Medium Range is part of a broader Earth-2 suite that also includes Nowcasting and Global Data Assimilation models.

 

The Nowcasting model generates short-term forecasts ranging from immediate conditions to six hours ahead and is intended to help meteorologists better assess the impacts of severe storms and other hazardous weather. Because the model is trained directly on globally available geostationary satellite data rather than region-specific physics models, it can be adapted for use anywhere with sufficient satellite coverage, Pritchard said. This approach could help governments, including smaller countries and subnational regions, better understand localized weather risks.

 

The Global Data Assimilation model addresses another computational challenge in forecasting: generating continuous snapshots of global weather conditions using data from sources such as weather stations and atmospheric balloons. Traditionally, this process requires large amounts of computing power before forecasting can even begin.

 

“Data assimilation accounts for roughly 50 percent of the total supercomputing load in traditional weather forecasting,” Pritchard said. “This model can do that work in minutes on GPUs instead of hours on supercomputers.”

 

The three new models join two existing Earth-2 tools: CorrDiff, which converts coarse forecasts into high-resolution predictions, and FourCastNet3, which models individual variables such as temperature, wind, and humidity.

 

Pritchard said the expanded Earth-2 suite could make advanced weather forecasting more accessible. Historically, such capabilities have been concentrated among wealthier nations and large organizations able to afford extensive supercomputing resources.

 

“These models provide core building blocks for the broader ecosystem — national meteorological agencies, financial firms, energy companies, and others looking to develop or refine weather forecasting systems,” he said. Nvidia noted that some organizations are already using the tools, including meteorologists in Israel and Taiwan, while companies such as The Weather Company and TotalEnergies are evaluating the Nowcasting model.

 

“For some users, subscribing to a centralized forecasting system makes sense,” Pritchard said. “But for countries, sovereignty matters. Weather is a national security issue, and sovereignty and weather are inseparable.”