Nvidia launches Earth-2 open-source AI models to transform forecasting

Nvidia unveiled Earth-2 at the American Meteorological Society meeting in Houston, presenting a family of open-source artificial intelligence models aimed at reshaping how weather is predicted. The suite includes a medium-range global forecast model capable of producing 15-day outlooks and a short-range nowcasting model designed to track severe storms up to six hours ahead, signaling a move by a major technology company into operational meteorology with tools intended for broad community use.

The Earth-2 announcement highlights two parallel ambitions: to extend the reach of advanced forecasting beyond well-funded national centers and to accelerate innovation by opening model code and weights to researchers, agencies, and private developers. By making models and training artifacts public, Nvidia is betting that transparent, collaborative development will speed improvements in accuracy, lead times and local customization, areas where traditional numerical weather prediction has been slow to evolve for many users.

The 15-day model occupies an important niche between short-range rapid-updating systems and seasonal forecasts. If driven by high-frequency observations and robust data assimilation, a data-driven medium-range forecast could help emergency managers, utilities and agriculture planners refine decisions about resource allocation and risk mitigation during unfolding weather events. The nowcasting component aims to address a different, urgent problem: providing actionable warnings for convective storms, flash floods and hazardous wind events with a horizon of minutes to hours when human lives and infrastructure are most vulnerable.

Open-source release changes the incentives around verification, trust and deployment. Researchers can independently evaluate model performance across regions where observational networks vary widely, and local meteorological services and universities can adapt models to regional radar networks, satellite feeds and surface observations. That accessibility could narrow gaps in forecasting capability between wealthy and resource-constrained countries, where computational or licensing barriers have limited the use of state-of-the-art models.

But openness is no panacea. Operational adoption will depend on rigorous independent validation, consistent data inputs, and integration into existing forecasting workflows. Machine learning approaches have shown promise, especially at short ranges, but they can underperform when fed biased or sparse data. Models trained on historical patterns may struggle with novel events amplified by a changing climate. Ensuring robust uncertainty estimates, avoiding false alarms, and maintaining transparency about limitations will be essential for agencies that issue public warnings.

Earth-2 also raises questions about infrastructure and governance. Running large AI models at scale requires substantial compute and data pipelines; smaller agencies may need partnerships or cloud services. Because the models are open, they invite diverse actors but also demand clear licensing and stewardship to prevent fragmentation or misuse. Independent benchmarks and collaborative evaluation frameworks will be critical to translate research promise into reliable public safety tools.

Nvidia’s entry intensifies a broader shift toward hybrid forecasting systems that combine physical models, machine learning and richer observational streams. The company’s open-source strategy could accelerate that convergence, but the ultimate test will be whether Earth-2 demonstrably improves forecast skill in real-world operations and whether the meteorological community can govern and deploy these tools responsibly for societies increasingly at risk from extreme weather.