NVIDIA Launches Earth-2, an Open AI Stack for Weather Forecasting

• First fully open, accelerated weather AI stack (January 2026)

  Announced at the American Meteorological Society Annual Meeting, Earth-2 is an end-to-end, open weather AI software stack. It combines pretrained models, inference tools, developer frameworks, and customization recipes, covering the entire forecasting pipeline: from raw observations to global and hyper-local forecasts. By shifting workloads from CPU-based supercomputers to GPUs, it significantly reduces compute time and costs, broadening access for researchers, enterprises, and governments.

• Earth-2 Medium Range: 15-day global forecasts across 70+ variables

  Built on NVIDIA’s Atlas architecture, Earth-2 Medium Range delivers forecasts up to 15 days ahead across more than 70 atmospheric variables, including temperature, pressure, wind, and humidity. On standard benchmarks, NVIDIA reports it outperforms leading open models on core forecasting metrics.

• Earth-2 Nowcasting: kilometer-scale storm prediction in minutes

  Using the StormScope generative AI architecture, Earth-2 Nowcasting produces zero- to six-hour forecasts at kilometer resolution for local storms and hazardous weather. By directly predicting satellite and radar imagery, it outperforms traditional physics-based models in short-term precipitation forecasting and delivers country-scale results in minutes rather than hours.

• Earth-2 Global Data Assimilation: atmospheric snapshots in seconds

  Powered by the HealDA architecture, this model generates global atmospheric initial conditions (temperature, wind, humidity, and air pressure) in seconds on GPUs instead of hours on supercomputers. When paired with Earth-2 Medium Range, it improves both forecast speed and accuracy, forming a fully AI-based, open forecasting pipeline.

• Existing open models integrated into the Earth-2 stack

  Earth-2 incorporates earlier NVIDIA models:

  • CorrDiff, which downscales continental forecasts to high-resolution regional outputs up to 500× faster than traditional methods.

  • FourCastNet3, which delivers high-accuracy forecasts for key variables up to 60× faster than leading diffusion-based approaches.

    The platform also integrates open models from the European Centre for Medium-Range Weather Forecasts, Microsoft, Google, and others, enabling multiple approaches within one framework.

• Operational use by national weather agencies

  The Israel Meteorological Service reports a 90% reduction in compute time at 2.5-kilometer resolution using CorrDiff, with improved six-hour precipitation verification after major storms. It plans to deploy Nowcasting up to eight times daily. Other evaluators include the U.S. National Weather Service, Taiwan’s Central Weather Administration, and The Weather Company.

• Energy, finance, and risk assessment adoption

  Beyond meteorology, Earth-2 is used by energy firms such as TotalEnergies, Eni, GCL, and Southwest Power Pool (with Hitachi) to improve grid reliability, wind forecasting, and photovoltaic prediction. Financial and insurance groups including AXA, S&P Global Energy, and JBA Risk Management apply the models to climate risk and extreme-weather scenario analysis.

• Open licensing, availability, and deployment

  NVIDIA states that the Earth-2 model family is licensed for commercial and non-commercial use. Earth-2 Medium Range and Nowcasting are available via Earth2Studio, GitHub, and Hugging Face, allowing organizations to run and fine-tune models on their own infrastructure. Global Data Assimilation is expected later in 2026. Training and customization are supported by NVIDIA PhysicsNeMo, an open-source AI-physics framework.

Why This Matters

Accurate weather and climate forecasts are essential for agriculture, energy, disaster response, insurance, public health, and infrastructure planning. By making high-performance weather AI fully open, GPU-accelerated, and production-ready, Earth-2 lowers barriers to advanced forecasting while delivering faster, more localized insights. The result is broader global access to climate intelligence and a clear shift away from exclusive reliance on supercomputers toward scalable, AI-driven forecasting systems.

This article was drafted with the assistance of generative AI. All facts and details were reviewed and confirmed by an editor prior to publication.