# Earth-2 by NVIDIA

Canonical page: https://technooptimist.io/ai/earth-2
Frontier: [AI](https://technooptimist.io/ai.md) · Field: Weather models · Company: [NVIDIA](https://technooptimist.io/companies/nvidia.md)
Stage: Deploy (4 of 5: Research → Proto → Pilot → Deploy → Scaling)
Updated: last update 2026-03-16, facts checked 2026-10-11
Goal: Give weather services, utilities and insurers a complete, open AI forecasting system they can run on their own GPUs, so fast, accurate forecasts no longer need hours on a national supercomputer.
Status: Open models since Jan 2026: 15-day global, 0–6 h storm nowcasts, and HealDA starting states.
Next milestone: Weather services and energy firms evaluate and adopt the models

## Milestones

| Status | Milestone | Date | Slip |
| --- | --- | --- | --- |
| Done | Earth-2 cloud platform and CorrDiff model launched at GTC | 18 Mar 2024 |   |
| Done | Open Medium Range (Atlas) and Nowcasting (StormScope) models | 26 Jan 2026 |   |
| Done | HealDA data-assimilation model released | 16 Mar 2026 |   |
| Current | Weather services and energy firms evaluate and adopt the models |   |   |

## Most important updates

- **16 Mar 2026**: NVIDIA publishes HealDA, an open AI model that builds a global weather starting state from raw observations. NVIDIA released HealDA on Hugging Face, the data-assimilation piece of Earth-2. ([our story](https://technooptimist.io/stories/nvidia-publishes-healda-an-open-ai-model-that-builds-a-global-weather-7f1c6b5c.md) · [original source](https://huggingface.co/nvidia/healda))
- **26 Jan 2026**: NVIDIA releases open Earth-2 weather models: 15-day Medium Range and 0–6 hour storm Nowcasting. At the American Meteorological Society's annual meeting, NVIDIA released Earth-2 as an open family of pretrained weather models, frameworks and inference libraries covering the whole forecast chain. ([our story](https://technooptimist.io/stories/nvidia-releases-open-earth-2-weather-models-15-day-medium-range-and-0-6-7a4916ed.md) · [original source](https://blogs.nvidia.com/blog/nvidia-earth-2-open-models))
- **18 Mar 2024**: NVIDIA launches Earth-2, a cloud platform for AI weather and climate simulation, with the CorrDiff model. NVIDIA launched Earth-2 at GTC, a cloud platform for simulating and visualising weather and climate with AI, offered as APIs on its DGX Cloud. ([our story](https://technooptimist.io/stories/nvidia-launches-earth-2-a-cloud-platform-for-ai-weather-and-climate-51ab38bd.md) · [original source](https://mezha.ua/en/2024/03/19/nvidia-introduces-earth-2-a-digital-twin-of-the-earth-for-weather-modeling-and-forecasting/))

## Current obstacles

- **Accuracy claims are still NVIDIA's own**: NVIDIA says Atlas beats other open models and StormScope beats physics on short-term rain; the US National Weather Service is still evaluating.
- **Coarse starting state**: HealDA outputs ~1° (~100 km) cells, four times coarser than the 0.25° grids forecast models start from, so its output is regridded.

## Physics limits

- **Learned only from the past**: AI forecasters learn from decades of recorded weather, so extremes beyond that record are where they are least tested; chaos in the atmosphere caps useful forecasts at about two weeks.
- **Blind where nobody measures**: A forecast is only as good as its starting state. Oceans and poles have few direct readings, so assimilation must infer them from satellites, which adds error.

## How it works

### Starting state: Observations into a global snapshot

HealDA, a 330M-parameter transformer, turns a day's ~10 million satellite and ground readings into a starting map of the atmosphere.

### Global forecast: Atlas steps out to 15 days

A simple, scalable transformer predicts 70+ variables (temperature, wind, humidity, pressure) step by step up to 15 days ahead.

### Nowcast: Storms straight from satellite images

StormScope is generative: trained on geostationary satellite images, it predicts km-scale satellite and radar pictures 0–6 hours ahead.

### Downscaling: CorrDiff sharpens local detail

A generative model turns coarse continental forecasts into fine regional fields; Israel's weather service runs it at 2.5 km.

## Spec sheet

| Spec | Earth-2 | Provenance |
| --- | --- | --- |
| Medium Range forecast length | Up to 15 days | reported |
| Medium Range variables | 70+ | reported |
| Nowcasting horizon | 0–6 hours | reported |
| Nowcasting resolution | Kilometre-scale | reported |
| HealDA model size | 330M parameters | reported |
| HealDA input per day | ~10M observations | reported |
| HealDA output grid | ~1° (~100 km) | reported |
| CorrDiff vs numerical models | 12.5x resolution, 1,000x faster | reported |

Provenance: reported = stated by the company; estimated = our estimate; sample = a sample figure.

## About NVIDIA

NVIDIA designs the GPUs, networking and software behind most AI training and inference. Its CUDA software made GPUs the default for AI. In October 2025 it became the first company worth $5 trillion.

- Vera Rubin in production in 2026; cloud partners get it from the second half of the year.
- Licensed Groq's inference technology and hired its leaders for about $20B (Dec 2025).
- Plans to invest up to $100B in OpenAI, tied to at least 10 GW of NVIDIA systems (Sep 2025).

- Founded: 1993
- Headquarters: US
- Status: public
- Listing: NVDA (NASDAQ)
- Valuation: $5.4T (market cap, 2026-09)
- Staff: ~36,000
- Website: https://www.nvidia.com
- People: Jensen Huang (Co-founder, President and CEO); Colette Kress (CFO); Bill Dally (Chief Scientist)
- Partners, customers, investors: TSMC (supplier); SK hynix (supplier); OpenAI (customer); Intel (partner); Groq (partner); CoreWeave (customer)

Company page: https://technooptimist.io/companies/nvidia.md

## Related programs

More from NVIDIA:

- [Vera Rubin](https://technooptimist.io/computing/vera-rubin.md) (Deploy)
- [Blackwell](https://technooptimist.io/computing/blackwell.md) (Scaling)
- [Isaac GR00T](https://technooptimist.io/robotics/gr00t.md) (Pilot)
- [Nemotron 3](https://technooptimist.io/ai/nemotron-3.md) (Deploy)
- [Cosmos 3](https://technooptimist.io/ai/cosmos-3.md) (Deploy)
- [Groq 3 LPX](https://technooptimist.io/computing/groq-3-lpx.md) (Proto)
- [Rubin CPX](https://technooptimist.io/computing/rubin-cpx.md) (Proto)
- [Spectrum-X and Quantum-X Photonics](https://technooptimist.io/computing/spectrum-x-quantum-x-photonics.md) (Deploy)
- [Omniverse (Factory Digital Twins)](https://technooptimist.io/manufacturing/omniverse-factory.md) (Deploy)
- [NVIDIA DRIVE](https://technooptimist.io/vehicles/nvidia-drive.md) (Deploy)
- [CUDA-Q and NVQLink](https://technooptimist.io/quantum/cuda-q-nvqlink.md) (Deploy)

Everything in AI: https://technooptimist.io/ai.md

---

The Techno Optimist tracks every frontier of technology, program by program: https://technooptimist.io/ (a Markdown version of any page: add .md to its URL; site map for agents: https://technooptimist.io/llms.txt).
Data API (JSON, realtime stream): https://technooptimist.io/developers. People read 5 pages free, then Reader $10/month; these Markdown pages are open to agents.