Nvidia opens ‘Earth-2’ AI weather models, promising faster forecasts on smaller GPU clusters

A winter storm backdrop for a new pitch

HOUSTON — As a winter storm disrupted flights and snarled traffic across the United States this week, hundreds of meteorologists packed into a conference hall here to hear a pitch that could reshape how the world predicts weather.

On Jan. 26 at the American Meteorological Society’s annual meeting, Nvidia Corp. unveiled what it calls the Earth-2 family of open models — a suite of artificial intelligence systems, code libraries and tools that the company says can deliver operational weather forecasts at supercomputer scale on far smaller, cheaper GPU clusters.

The company is billing Earth-2 as “the world’s first fully open, accelerated weather AI software stack.” It is releasing model weights and source code for free on platforms such as GitHub and Hugging Face, aiming the technology at national weather services, research institutes, energy companies and insurers.

“AI-powered weather forecasting saves significant computational time and costs, allowing more nations, weather enterprises and businesses to run application-specific forecasting systems,” Mike Pritchard, Nvidia’s director of climate simulation, said in a company blog post announcing the launch.

The move pushes AI weather forecasting from laboratory experiments into a form that governments and utilities are already testing for day-to-day operations. It also tightens the link between critical public infrastructure and a single private chipmaker whose hardware and software underpin the system’s promised performance.

Three models aimed at the forecasting pipeline

At the center of the new stack are three models designed to cover the main stages of modern forecasting: setting the current state of the atmosphere, predicting its evolution over days and weeks, and tracking hazardous storms over the next few hours.

Earth-2 Medium Range (Atlas)

The first, Earth-2 Medium Range, is a global model built on a new “Atlas” architecture. It generates forecasts up to 15 days ahead, predicting more than 70 variables including temperature, wind, humidity and pressure.

Nvidia says Atlas, which uses a latent diffusion transformer to simulate changes in the atmosphere over time, matches or outperforms leading open AI systems on standard benchmarks of medium-range skill while running far faster than traditional numerical weather prediction models. Those comparisons include models such as Google DeepMind’s GenCast, though GenCast itself is not released with open weights.

Earth-2 Nowcasting (StormScope)

The second, Earth-2 Nowcasting, code-named StormScope, targets the zero- to six-hour window that matters for flash floods, thunderstorms and aviation. It ingests geostationary satellite images and ground-based radar data to generate kilometer-scale forecasts of precipitation and storm structure.

Nvidia says StormScope is the first AI model to surpass physics-based short-range forecast systems on key measures of rainfall accuracy. A technical paper posted by company researchers reports that its 10-minute, roughly 6-kilometer forecasts are competitive with or better than state-of-the-art mesoscale numerical models out to six hours, and substantially outperform extrapolation-based nowcasting methods beyond one to two hours.

Earth-2 Global Data Assimilation (HealDA)

The third new component, Earth-2 Global Data Assimilation, branded HealDA, is intended to replace the sprawling numerical schemes that turn raw observations into a coherent picture of the atmosphere. Those systems, known as data assimilation, often consume roughly half of a weather center’s supercomputing time.

HealDA uses machine learning to map a short window of global satellite and conventional observations directly to a gridded three-dimensional atmospheric state, without needing a prior forecast as background during runtime. Nvidia says the model can produce an analysis in seconds on GPUs, compared with hours on large supercomputers, and that when paired with its medium-range AI model it yields “the most skillful forecasting predictions produced by an open, entirely AI pipeline.” The company plans to release HealDA later this year.

Existing Earth-2 components and speed claims

Those new tools join existing Earth-2 components, including FourCastNet3, a global forecast model that can generate a 60-day, quarter-degree-resolution forecast in under four minutes on a single Nvidia H100 GPU. Company tests show that run time to be about 60 times faster than a comparable forecast produced by the European Centre for Medium-Range Weather Forecasts’ flagship physics-based ensemble system.

Earth-2 also includes CorrDiff, a generative downscaling model that refines coarse continental-scale predictions into high-resolution local fields, and Earth2Studio, an open-source Python framework under an Apache-2.0 license for building pipelines that mix Nvidia and third-party models such as ECMWF’s AI Forecasting System.

Early adopters: agencies, energy companies and insurers

Beyond benchmark claims, some public agencies and companies say they are already seeing operational benefits.

The Israel Meteorological Service has been running Nvidia’s CorrDiff model in production. Director Amir Givati said the agency has seen about a 90% reduction in compute time for 2.5-kilometer-resolution simulations compared with a classic numerical model on a CPU cluster.

“Nvidia Earth-2 models give us a 90% reduction in compute time at 2.5-kilometer resolution compared with running a classic numerical weather prediction model without AI on a CPU cluster,” Givati said in remarks released by the company. After a recent rainstorm, he said, an AI system trained with CorrDiff delivered the best performance of all the agency’s operational models for a six-hour precipitation period.

Taiwan’s Central Weather Administration is using Earth-2 tools including FourCastNet and CorrDiff for typhoon prediction and localized forecasts. The U.S. National Weather Service is evaluating the medium-range and nowcasting models for potential integration into its workflow, though any shift in official forecasts would require extensive independent validation.

Private firms have moved quickly. Brightband, a weather technology startup, runs Earth-2 Medium Range daily to produce global forecasts.

“The model being open source speeds up innovation, allowing easier comparison and improvements by other members of the weather enterprise,” Brightband chief executive Julian Green said.

Energy and grid operators, who rely heavily on weather to manage wind and solar power, are also testing the stack. TotalEnergies is piloting Earth-2 Nowcasting to improve short-term risk awareness for its energy systems.

“Models like Earth-2 Nowcasting are groundbreaking for our business because they improve short-term risk awareness and decision-making in energy systems where minutes and local impacts matter,” Emmanuel Le Borgne, a senior manager at TotalEnergies, said.

Italian energy company Eni is using FourCastNet and CorrDiff for probabilistic, high-resolution forecasts weeks ahead. U.S. grid operator Southwest Power Pool, together with Hitachi, has been exploring Earth-2 models to refine intraday and day-ahead wind forecasts.

Insurers and risk analysts including AXA, S&P Global and U.K.-based JBA Risk Management are using or testing Earth-2 models to generate large ensembles of hurricane and flood scenarios for pricing and stress-testing portfolios.

Promise—and a new dependency question

Nvidia is framing the opening of Earth-2 as a contribution to what it calls “sovereign AI” weather infrastructure. Pritchard has argued that weather prediction is intertwined with national security, saying in public remarks that “sovereignty and weather are inseparable.” By releasing models and code, the company says, it wants to help countries that cannot afford multi-petaflop supercomputers run forecasts that approach the quality of top global centers.

Some meteorologists and policy experts see potential for lower-income countries, especially in regions where climate change is amplifying extreme events. A midrange GPU cluster is often cheaper and easier to maintain than a bespoke national supercomputer, and open models can shorten the path from research to operations.

At the same time, the system’s most dramatic speed and efficiency gains depend on Nvidia’s own hardware and software stack, including its CUDA platform. While the Earth-2 code and weights are open, replicating advertised performance on alternative accelerators is likely to be difficult in the near term.

Weather and climate scientists also caution that strong performance on historical reanalysis data does not guarantee robust behavior under future climate regimes or during unusual events that fall outside the training distribution. Data limitations are another concern: StormScope, for example, has been trained mainly on the continental United States, where dense radar and satellite coverage exist, and may not generalize as well to regions with sparse observations.

National centers such as the U.S. National Weather Service and ECMWF are running their own verification studies and, in many cases, treating AI models as complements to, rather than replacements for, physics-based systems. Many forecasters expect a hybrid era in which machine learning provides rapid ensembles, specialized nowcasts and downscaling, while traditional numerical models remain the backbone for global guidance and for checking AI-driven outputs.

A turning point for open AI forecasting

For now, Earth-2 marks a turning point: high-end AI forecast models once locked inside big technology companies are being released with the code and weights needed to run and adapt them. Whether that leads to a more level global playing field in weather prediction, or to a new form of dependence on commercial hardware providers, will be tested as agencies and companies move from pilot projects to decisions that affect lives and infrastructure.

As the Houston storm cleared, meteorologists left the conference with terabytes of new code to download and a familiar challenge: turning mathematical skill scores into forecasts people can trust, in an atmosphere that is changing faster than the institutions that seek to predict it.