As a major winter storm batters much of the United States, recent forecasts for several regions showed considerable inconsistency, with predictions for snowfall varying dramatically. The timing of Nvidia’s unveiling of its new Earth-2 weather forecasting models could not have been more fitting. Or perhaps, given the company's claims about the models' accuracy, it had foreseen the need.

These new AI models are designed to deliver weather forecasts more rapidly and with greater precision. Nvidia asserts that one model, Earth-2 Medium Range, outperforms Google DeepMind’s AI weather model, GenCast, across more than 70 different variables. Google released GenCast in December 2024, and it was already notably more accurate than existing models capable of 15-day forecasts.

The tools were announced on Monday at the American Meteorological Society meeting in Houston. "Philosophically and scientifically, this is a return to simplicity," said Mike Pritchard, Nvidia’s director of climate simulation, in a call with reporters before the meeting. "We’re moving away from hand-tailored, niche AI architectures and leaning into the future of simple, scalable transformer architectures."

Conventional weather forecasting has largely depended on simulating physics based on real-world observations, with AI models being a more recent innovation. The Earth-2 Medium Range model is built on a new Nvidia architecture named Atlas, with further details promised for Monday.

The Earth-2 suite also includes a Nowcasting model and a Global Data Assimilation model. Nowcasting generates short-term predictions from zero to six hours ahead, intended to help meteorologists anticipate the impacts of storms and other dangerous weather. "Because this model is trained directly on globally available geostationary satellite observations, rather than region-specific physics model outputs, Nowcasting’s approach can be adapted anywhere on the planet with good satellite coverage," Pritchard explained. This capability should assist state governments and smaller nations in understanding how severe weather could affect their regions.

The Global Data Assimilation model integrates data from sources like weather stations and balloons to create continuous snapshots of conditions at thousands of global locations. These snapshots then serve as the starting point for other models to generate forecasts. Historically, producing these snapshots demanded enormous computing resources before any forecasting could begin. "It consumes roughly 50% of the total supercomputing loads of traditional weather forecasting," Pritchard noted. "This model can do that in minutes on GPUs instead of hours on supercomputers."

These three new models join two existing ones: CorrDiff, which uses coarse forecasts to produce rapid, high-resolution predictions, and FourCastNet3, which models individual variables like temperature, wind, and humidity. According to Pritchard, the new suite should democratize access to powerful forecasting tools, which have traditionally been limited to wealthier countries and large corporations that can afford expensive supercomputer time. "This provides the fundamental building blocks used by everyone in the ecosystem—national meteorological services, financial service firms, energy companies—anyone who wants to build and refine weather forecasting models," he said.

Some tools are already in operation. For instance, meteorologists in Israel and Taiwan have been using Earth-2 CorrDiff, while The Weather Company and Total Energies are evaluating the Nowcasting model. "For some users, it makes sense to subscribe to an enterprise centralized weather forecasting system. But for others like countries, sovereignty matters," Pritchard stated. "Weather is a national security issue, and sovereignty and weather are inseparable."