As a winter storm battered large swaths of the U.S., weather forecasts in the days leading up to it were anything but consistent. In some regions, snowfall projections swung dramatically, underscoring how uncertain even modern forecasting can be.

 

Against that backdrop, Nvidia’s timing looked impeccable. The company unveiled a new set of AI-powered weather forecasting models just as the storm arrived. Or, given how accurate Nvidia says the models are, perhaps the company had a clearer view of what was coming all along.

 

Nvidia claims its new Earth-2 weather models can generate forecasts both faster and more accurately than existing systems. One model in particular, Earth-2 Medium Range, outperforms Google DeepMind’s GenCast across more than 70 variables, according to the company. Google released GenCast in December 2024, and it was already a significant leap forward, delivering forecasts up to 15 days out with greater accuracy than traditional approaches.

 

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

 

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

 

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

 

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

 

Nowcasting focuses on short-term forecasts, from the present moment out to six hours, and is designed to help meteorologists better predict the impacts of severe storms and other hazardous weather events. Because the model is trained directly on globally available geostationary satellite data — rather than region-specific physics models — it can be adapted anywhere in the world with sufficient satellite coverage, Pritchard said. That could help governments, including those of smaller countries or individual states, better assess localized risks.

 

The Global Data Assimilation model tackles another major bottleneck in forecasting: assembling real-time snapshots of global weather conditions using data from sources such as weather stations and balloons. Traditionally, generating those snapshots has consumed enormous computing resources before forecasting even begins.

 

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

 

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

 

Together, Pritchard said, the models could broaden access to advanced weather forecasting, which has long been dominated by wealthy nations and large organizations that can afford expensive supercomputer time.

 

“These are foundational building blocks for everyone in the ecosystem — national meteorological agencies, financial firms, energy companies — anyone looking to develop or refine weather models,” he said. Some organizations are already putting the tools to use. Meteorologists in Israel and Taiwan are using CorrDiff, while companies including The Weather Company and TotalEnergies are evaluating the Nowcasting model, Nvidia said.

 

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