Traditional weather forecasting runs on supercomputers that most countries can't afford. NVIDIA just open-sourced an AI stack that replaces them. The package is called Earth-2. It covers every stage of weather prediction: processing raw observation data, generating 15-day global forecasts, and producing kilometre-resolution storm predictions. All open. All free. Announced at the American Meteorological Society's annual meeting in January. Three new models sit at the core: 1. Earth-2 Medium Range (architecture: Atlas) forecasts up to 15 days across 70+ weather variables and outperforms leading open models on standard benchmarks. 2. Earth-2 Nowcasting (architecture: StormScope) generates zero-to-six-hour local storm predictions at kilometre resolution, the first AI model to beat traditional physics-based models on short-term precipitation. 3. Earth-2 Global Data Assimilation (architecture: HealDA) produces atmospheric snapshots in seconds on GPUs instead of hours on supercomputers. Chain the three together and you get the most skilful open, fully AI weather pipeline ever built. The cost gap is where the story gets concrete. The Israel Meteorological Service is already running Earth-2 CorrDiff in operation and reports a 90% reduction in compute time at 2.5km resolution compared with running classical numerical weather prediction on a CPU cluster. After a recent rainstorm, their AI model trained with CorrDiff was their best-performing operational model for six-hour precipitation verification. They generate high-resolution forecasts up to eight times daily now. The adoption list reads like a cross-section of the global economy: • The Weather Company (which operates the Weather Channel), • Taiwan's Central Weather Administration, • the U.S. National Weather Service. • S&P Global Energy, • JBA Risk Management. • AXA is using FourCastNet to generate thousands of hypothetical hurricane scenarios for risk modelling. • GCL, one of China's largest solar producers, runs Earth-2 in production for photovoltaic power prediction. The timing is pretty important as well. NOAA is facing staff cuts under the current administration. Access to weather data in the U.S. is becoming a live political question. Meanwhile, the countries hit hardest by flooding, drought and extreme heat tend to be the ones with the least forecasting infrastructure. This follows a pattern across big tech. Microsoft built Aurora. Google partnered NeuralGCM with the European Centre for Medium-Range Weather Forecasts to get granular forecasting data to small farms in India. But NVIDIA's play is different in scope. They haven't built one model. They've open-sourced the full pipeline, from raw satellite observations to local storm prediction, and made it run on standard GPU infrastructure. Physics-based weather models were one of the original justifications for building supercomputers. For decades, only wealthy nations and large agencies could afford to run them. NVIDIA just made the replacement available to anyone with a GPU and an internet connection.