Nvidia launches Earth-2 AI models for weather forecasting

On Monday, Nvidia announced its new Earth-2 weather forecasting models at the American Meteorological Society meeting in the U.S. The company claims that these new models are accurate, suggesting that maybe they can improve forecasting speed and precision. Specifically, Nvidia asserts that its Earth-2 Medium Range model, which is based on a new architecture called Atlas, outperforms Google DeepMind’s GenCast AI weather model on more than 70 variables. Google released GenCast in December 2024, which itself improved upon existing weather models that were capable of generating forecasts up to 15 days out.

The Earth-2 suite includes Nowcasting—a model for short-term weather prediction from zero to six hours into the future—and the Global Data Assimilation model, which is a process of integrating data into weather models. Mike Pritchard, director of climate simulation at Nvidia, noted that the suite represents a shift away from customized niche setups toward simple, scalable transformer architectures, which are a type of AI model architecture.

According to Pritchard, the Global Data Assimilation model requires roughly 50% of the supercomputing loads typically consumed by traditional weather forecasting, running on GPUs in minutes rather than hours on supercomputers. Meanwhile, the Nowcasting model is designed to help meteorologists predict the impacts of hazardous weather. Because the Nowcasting model is trained directly on globally available geostationary satellite observations instead of region-specific physics outputs, Pritchard stated that the approach can be adapted anywhere on the planet that has good satellite coverage.

Nvidia is positioning these tools as critical infrastructure for national sovereignty. Pritchard explained that while some users might prefer subscribing to centralized enterprise systems, sovereignty is a key concern for individual countries. “Weather is a national security issue, and sovereignty and weather are inseparable,” said Mike Pritchard, director of climate simulation at Nvidia. The tools are intended to provide fundamental building blocks for national meteorological services, energy companies, and financial firms. These new tools join Nvidia’s existing models, CorrDiff and FourCastNet 3. Early adoption is already underway; meteorologists in Israel and Taiwan are using Earth-2 CorrDiff, while companies like The Weather Company and Total Energies are currently evaluating the Nowcasting model.

Why it matters

Nvidia is positioning its new Earth-2 AI weather models as a more efficient, accessible alternative to traditional supercomputer-based forecasting. By framing the technology around national sovereignty and security, the company aims to expand its footprint from commercial enterprise clients to sovereign state infrastructure.