Nvidia launches Earth-2 to predict climate change with AI

Predicting the exact path of a severe typhoon usually means burning hours of expensive supercomputer time on dense physics equations. Nvidia wants to condense that process down to a few clicks and a handful of seconds.

At GTC 2024, CEO Jensen Huang officially rolled out Earth-2, a climate digital twin cloud platform. Nvidia first floated the idea back in 2021, but it is now finally ready for production.

With extreme weather causing an estimated $140 billion in economic losses each year, the goal is simple: forecast regional climate catastrophes before they hit, using generative AI instead of raw brute-force computing.

Kilometer-scale forecasting in seconds

Traditional weather forecasting relies heavily on numerical models running on massive CPU clusters. They are slow, energy-hungry, and often too coarse to catch micro-level disasters unless you throw obscene amounts of compute at them.

According to Rev Lebaredian, Nvidia’s VP of simulation, achieving impactful forecasts requires kilometer-scale resolution. Running classical simulations at that granularity can demand a million times more compute.

Earth-2 bridges that gap with AI. Sitting inside Nvidia CUDA-X and running on DGX Cloud, the platform features CorrDiff, a generative diffusion model designed specifically for high-resolution atmospheric physics.

Nvidia claims CorrDiff produces high-res simulations 1,000 times faster and up to 3,000 times more energy-efficient than existing numerical models. Combined with APIs linked into Nvidia Omniverse, users can visualize weather phenomena at a two-kilometer scale in seconds rather than hours.

Early adopters are already moving in. Taiwan’s Central Weather Administration is tapping CorrDiff to sharpen typhoon tracking for earlier evacuations. Commercial giants like The Weather Company, Spire, and Meteomatics, alongside startups like Tomorrow.io, flood-risk platform ClimaSens, and ocean-mapping outfit north.io, are integrating the APIs into their operational stacks.

Why it matters

We think Earth-2 is a textbook example of Nvidia packaging its full stack—hardware, AI frameworks, and Omniverse—into a specialized, high-margin cloud service.

For enterprise users, the immediate value is massive. Logistics operators, offshore wind farms, and regional insurers no longer need a national laboratory's budget to run hyper-local climate models. A startup can now query an API to simulate urban flood risks or route autonomous marine vehicles without maintaining physical server racks.

Strategically, it cements Nvidia's grip on scientific computing. Instead of trying to convince atmospheric scientists to throw away decades of physical modeling, Earth-2 blends standard numerical models (like ICON and IFS) with generative models like CorrDiff and GraphCast. Either way, the underlying compute runs on Nvidia GPUs.

That said, generative climate modeling comes with real questions. Diffusion models excel at sharpening images and downscaling data, but extreme climate events are inherently rare edge cases. If an AI hallucinates a storm track during an unprecedented crisis, the cost is measured in human lives. For now, Earth-2 looks like an incredible accelerator for legacy workflows, but meteorologists will likely keep traditional numerical baselines close at hand.