GenCast's Real Numbers, and a New Open-Source Challenger

Climate & Earth Science has been one of the fastest-moving corners of AI-for-science coverage, and it just got more competitive. Here’s what’s actually verified, versus what’s still marketing claim.

What DeepMind actually published

Google DeepMind’s GenCast — a diffusion-based ensemble forecasting model — was published in Nature in December 2024, and the peer-reviewed numbers are genuinely strong: the model beat ECMWF’s operational ensemble system on 97.2% of 1,320 verification targets, with that win rate climbing to 99.8% at lead times beyond 36 hours. DeepMind released the model’s code, weights, and forecasts openly, explicitly framing it as a complement to — not a replacement for — traditional numerical weather prediction.

Why this matters for researchers: if you’re benchmarking a new climate or weather method, GenCast’s published Nature results are the legitimate comparison point — not the 2023 GraphCast numbers that still get cited in many papers as the current AI weather baseline.

A caution worth flagging: several 2026 news write-ups have gone further than DeepMind’s own claims, describing GenCast as having “solved” extreme weather prediction or fully replaced physics-based systems. As of 2026, no major meteorological agency has decommissioned its numerical weather prediction system — AI models are being used alongside physics-based forecasting, as one input among several that forecasters weigh, not a wholesale replacement.

A new open-source competitor: NVIDIA’s Earth-2

In early 2026, NVIDIA open-sourced a five-model suite called Earth-2, announced at an American Meteorological Society meeting, claiming it outperforms GenCast across more than 70 weather variables while running on standard GPUs rather than requiring specialized hardware.

Real-world adoption is already happening in a limited way: Israel’s Meteorological Service reports a 90% reduction in compute time using Earth-2 at high resolution, and has moved one of its component models into daily operational use. The U.S. National Weather Service and some energy companies are reportedly evaluating the suite as well.

The important caveat: NVIDIA’s performance comparisons against GenCast have not yet been independently peer-reviewed. Treat the “outperforms GenCast” claim the way you’d treat any vendor benchmark — plausible, worth testing yourself, but not yet verified the way GenCast’s own Nature-published numbers are.

What to watch next

Whether Earth-2’s comparison claims hold up under independent, peer-reviewed scrutiny
Whether any agency moves a GenCast-based system to the same kind of full operational status ECMWF’s AIFS or NOAA’s AIGFS already have
Continued “GenCast for X” attempts applying the same diffusion-based approach to adjacent problems like ocean forecasting or wildfire spread, which some researchers expect to see more of over the next year

This piece follows our sourcing standards: DeepMind’s own Nature publication and blog post anchor the verified claims; NVIDIA’s own announcement plus agency adoption reports anchor the Earth-2 section; commentary describing either as having “solved” weather forecasting is treated as unverified framing, not fact.