By ByteBot
NVIDIA launched Earth-2 on January 26, making professional-grade weather forecasting accessible to any developer with a GPU. The platform includes five open-source AI models that replace supercomputer-based forecasting with GPU-accelerated predictions—generating 15-day global forecasts in minutes instead of hours. Israel’s Meteorological Service reports 90% reductions in compute costs. Here’s how to get started with Earth2Studio, the Python framework that makes these models accessible.
Five Models, One Complete Weather Stack
Earth-2 isn’t a single model—it’s an end-to-end pipeline. Each of the five models targets a specific forecasting need.
Earth-2 Medium Range (Atlas) generates 15-day global forecasts across 70+ variables (temperature, pressure, wind, humidity). NVIDIA claims it outperforms Google’s GenCast on standard benchmarks. Earth-2 Nowcasting (StormScope) predicts severe weather at kilometer-scale resolution for 0-6 hours ahead—the first AI model to beat traditional physics-based models for short-term precipitation. Earth-2 Global Data Assimilation (HealDA) generates initial atmospheric conditions in seconds on GPUs versus hours on supercomputers.
The downscaling and ensemble models complete the stack. CorrDiff takes coarse forecasts and downscales them to 200-meter detail 500x faster than traditional methods. FourCastNet3 runs ensemble forecasts 60x faster for probabilistic predictions. Together, these five models cover everything from hyperlocal city forecasts to global medium-range predictions.
Real-world deployments prove this works in production. Israel’s Meteorological Service cut compute time 90% at 2.5-kilometer resolution compared to traditional numerical weather prediction. The UAE deployed a custom Earth-2 variant that outputs 200-meter resolution forecasts for Abu Dhabi, capturing the influence of coastal islands and urban development on local weather.
Getting Started with Earth2Studio
Earth2Studio is the Python framework that makes Earth-2 accessible. Installation takes one command:
pip install earth2studio[dlwp]Running your first forecast requires under 10 lines of code:
from earth2studio.models.px import DLWP
from earth2studio.data import GFS
from earth2studio.io import NetCDF4Backend
from earth2studio.run import deterministic as run
model = DLWP.load_model(DLWP.load_default_package())
ds = GFS()
io = NetCDF4Backend("output.nc")
run(["2024-01-01"], 10, model, ds, io)Model checkpoints live on Hugging Face. StormScope (nowcasting) and Atlas (medium-range) are available now, with HealDA coming soon. The examples gallery includes downloadable Jupyter notebooks covering deterministic forecasting, ensemble runs, and downscaling workflows.
System requirements aren’t trivial—you need an NVIDIA GPU with at least 16GB VRAM for basic forecasts. DGX systems or cloud instances with A100/H100 GPUs work best for production workloads. This hardware dependency is Earth-2’s main limitation compared to device-agnostic alternatives like GraphCast.
Who’s Already Using Earth-2
Weather agencies deployed Earth-2 within days of release. Beyond Israel and Taiwan’s Central Weather Administration, the U.S. National Weather Service is evaluating the models for operational workflows. The Weather Company is testing Nowcasting for localized severe-weather applications.
Energy companies dominate early adoption. TotalEnergies uses Nowcasting to improve short-term risk awareness for grid operations. Eni, GCL, and Southwest Powerool (partnered with Hitachi) all run Earth-2 for energy forecasting. Insurance and finance firms like AXA, JBA Risk Management, and S&P Global Energy use the models for hindcasting and flood risk assessment.
These aren’t pilot projects—they’re production deployments replacing expensive traditional infrastructure. That 90% cost reduction matters when you’re running forecasts continuously.
Earth-2 vs GraphCast: When to Choose What
Google DeepMind’s GraphCast remains the established open-source weather AI. It generates 10-day forecasts at 28-kilometer resolution, outperforming traditional HRES on 90% of test variables. GraphCast uses Graph Neural Networks and runs on any hardware—no NVIDIA GPU lock-in.
Earth-2 trades broader compatibility for higher resolution and deeper capabilities. Where GraphCast maxes out at 28km, Earth-2 can downscale to 200 meters. GraphCast delivers one forecast; Earth-2 provides an entire pipeline from data assimilation through downscaling. NVIDIA claims Earth-2 Medium Range beats GenCast (Google’s probabilistic model) on 70+ meteorological variables, though independent benchmarks would clarify the comparison.
Choose Earth-2 if you need hyperlocal forecasts, run NVIDIA infrastructure, or want end-to-end control. Choose GraphCast if you need simple 10-day global forecasts or run non-NVIDIA hardware. For most developers experimenting with weather AI, GraphCast’s lower barrier to entry wins. For production applications requiring city-scale resolution, Earth-2’s capabilities justify the GPU investment.
What This Enables
Earth-2 democratizes weather forecasting categories that were previously impractical. Hyperlocal 200-meter forecasts enable city-block level predictions for urban planning, drone operations, and air taxi routing. Real-time processing (minutes vs hours) unlocks reactive systems—renewable energy optimization that responds to weather changes, agricultural automation that triggers irrigation based on upcoming rainfall.
“Sovereign forecasting” matters geopolitically. Countries can now run professional-grade forecasts locally without depending on ECMWF or NOAA infrastructure. That 90% cost reduction means startups can afford what only national weather services could previously deploy.
The open-source release accelerates innovation faster than any proprietary platform could. Researchers can customize models for specific regions (like the UAE’s 200-meter Abu Dhabi variant), vertical-specific applications emerge (agriculture, maritime, aviation), and community contributions improve the core models. NVIDIA wins by selling more GPUs; developers win by accessing tools that were locked behind supercomputer paywalls six months ago.
Key Takeaways
- Earth-2 is fully open source and accessible via Python (Earth2Studio framework)
- Five models cover data assimilation through downscaling, enabling end-to-end forecasting pipelines
- Early adopters report 90% compute cost reductions versus traditional numerical weather prediction
- Requires NVIDIA GPU (16GB+ VRAM) unlike device-agnostic GraphCast
- Production-ready: Weather agencies, energy companies, and insurers deployed within days of release
- Get started at github.com/NVIDIA/earth2studio
