Nvidia Unleashes Earth-2: Open-Source AI Weather Forecasting for Everyone

The Dawn of Democratic Weather Forecasting

Nvidia has unveiled Earth-2, a groundbreaking suite of open-source artificial intelligence models designed to transform weather forecasting from an exclusive supercomputer domain into accessible technology for organisations worldwide. The announcement came at the American Meteorological Society's annual meeting in Houston, coinciding with a severe winter storm hitting much of the United States.

Three Models, One Mission

The Earth-2 family comprises three distinct AI models, each targeting different forecasting needs. Earth-2 Medium Range, built on the new Atlas architecture, delivers predictions for over seventy weather variables up to fifteen days ahead. Nvidia claims it outperforms Google DeepMind's GenCast model on standard benchmarks.

Earth-2 Nowcasting, powered by StormScope architecture, focuses on short-term predictions, generating kilometre-resolution storm forecasts for windows of zero to six hours. This model is trained on globally available satellite observations, making it adaptable to any region with decent satellite coverage.

The third component, Earth-2 Global Data Assimilation using HealDA architecture, tackles the computationally intensive process of establishing initial atmospheric conditions. What traditionally consumed roughly half of all supercomputing resources in weather forecasting can now be completed in seconds on standard graphics processing units.

Real-World Impact Already Underway

Several organisations have already adopted the platform. The Israel Meteorological Service reported a ninety percent reduction in compute time compared to traditional methods. Weather AI provider Brightband is using Earth-2 Medium Range for daily global forecasts. Energy giants including TotalEnergies and Eni are testing the models for demand forecasting, whilst Chinese solar manufacturer GCL Technology has deployed Earth-2 operationally.

Nvidia's Mike Pritchard emphasised the sovereignty implications, noting that weather is fundamentally a national security issue. The full model family will be available on GitHub and Hugging Face under commercial and non-commercial licences.