← Downlink 11 Jun 2025 · AIwire

NVIDIA Earth-2: Climate at Kilometer Scale

11 Jun 2025 Source: AIwire EO · AI

Digital twins of planets sound like science fiction until the hardware arrives. In June, NVIDIA detailed Earth-2, generative AI models that simulate the global climate at kilometer-scale resolution — the company's bid to make a full digital twin of Earth computable.

The physics background is sobering. Traditional climate models run on grids of tens of kilometers; resolving a kilometer means simulating cloud convection directly, which multiplies the computation by orders of magnitude. Conventional supercomputing simply does not reach, which is why Earth-2 uses learned emulators: neural models trained on physics to produce high-resolution fields at a fraction of the cost.

The architecture combines the two AI tribes deliberately. A physics-based model keeps the large-scale dynamics honest; generative downscaling models add the fine structure; a data-assimilation layer continuously anchors the twin to reality — including satellite observations.

The near-term customers are insurers, energy traders and governments that need to price wind, solar and storm risk at the scale of actual assets rather than grid cells.

The design pattern matters beyond weather: emulate expensive physics with learned models, verify against observation, never let the twin drift from the data. That is the recipe for every digital twin now being attempted — of cities, grids and machines — and Earth-2 is its flagship deployment.

The strategic detail worth noting is the delivery mechanism: Earth-2 is a platform, not a paper. The models ship in NVIDIA's software stack, meaning any lab with a GPU cluster can run kilometer-scale climate experiments that would once have required a national supercomputing allocation — a democratization of a capability that used to belong to five institutions on Earth.


Read the original →