If you’ve ever watched snowfall forecasts swing from “dusting” to “feet of snow” in the span of a day, Nvidia’s timing couldn’t be better. Right as a recent winter storm had predictions all over the place, the company showed up with a new set of AI weather models it says can make forecasting faster, more accurate, and less dependent on the traditional supercomputer-heavy approach.
Nvidia just introduced the latest additions to its Earth-2 forecasting suite at the American Meteorological Society meeting in Houston, including a new model called Earth-2 Medium Range. Nvidia claims it outperforms Google DeepMind’s GenCast model across more than 70 different weather variables — a notable flex considering GenCast only launched in December 2024 and was already seen as a major leap for forecasts up to 15 days out.
The bigger story here isn’t just “AI is getting better at weather.” It’s that Nvidia is pushing a simpler, scalable approach built around transformer-style architectures (instead of highly specialized AI designs), and packaging these models in a way that could put serious forecasting tools into more hands. Traditionally, the best weather prediction has been dominated by countries and companies that can afford huge amounts of supercomputing time. Nvidia’s pitch is that GPUs — plus the right AI models — can level that playing field.
Alongside Medium Range, Nvidia also rolled out two other models with very different jobs. Nowcasting is aimed at the immediate window — roughly 0 to 6 hours out — helping meteorologists track storms and other dangerous weather in near real time. And because it’s trained directly on globally available geostationary satellite data, Nvidia says it can be adapted anywhere there’s strong satellite coverage, not just regions with massive local forecasting infrastructure.
Then there’s Global Data Assimilation, which tackles one of the least glamorous but most expensive parts of forecasting: pulling in data from weather stations, balloons, and other sources to create a constantly updated snapshot of conditions around the world. Nvidia says this step can eat up around half of traditional forecasting compute loads, but its new model can do it in minutes on GPUs instead of hours on supercomputers.
These models join Nvidia’s existing Earth-2 lineup, including CorrDiff (for high-resolution forecasts from coarse inputs) and FourCastNet3 (for predicting individual variables like wind, temperature, and humidity). Nvidia says some of this is already seeing real-world testing — with meteorologists in Israel and Taiwan using CorrDiff, and groups like The Weather Company and TotalEnergies evaluating Nowcasting.
And Nvidia isn’t shy about why this matters: weather isn’t just a convenience feature — it’s infrastructure. As one exec put it, forecasting ties directly to national security, which is why “sovereignty” came up more than once. The idea is that not every country wants to rely on a centralized, black-box forecast provider if they can run powerful models themselves.
Bottom line: Nvidia is trying to turn weather prediction into something faster, cheaper, and more widely available — and if these models perform anywhere close to the hype, they won’t just help meteorologists. They’ll reshape how energy companies, insurers, governments, and financial firms prepare for what’s coming next.
This analysis is based on reporting from TechCrunch.
Image courtesy of Unsplash.
This article was generated with AI assistance and reviewed for accuracy and quality.
