Google shifts weather forecasting into AI fast lane with WeatherNext 3

Google DeepMind and Google Research today publicly launched WeatherNext 3, a next-generation artificial-intelligence weather model that delivers hourly forecasts at one-kilometer resolution for up to fifteen days ahead. The model is the first to combine a graph neural network backbone with a diffusion-based generative core, allowing it to ingest 350 terabytes of satellite, radar, and surface-station data every twenty-four-hour cycle. Demis Hassabis, chief executive of Google DeepMind, said in an on-stage interview at the Paris AI for Science Summit that WeatherNext 3 “outperforms the best operational European Centre for Medium-Range Weather Forecasts ensemble by 22 percent on critical skill scores,” measured over the 2023 calendar year. The announcement coincides with the release of an open API that lets national meteorological agencies ingest WeatherNext 3 outputs directly into their warning systems starting this quarter.

Across the Atlantic, ECMWF director Florence Rabier told OpenPress in an exclusive that WeatherNext 3’s hourly refresh rate and kilometer-scale resolution “will require governments to rethink their supercomputing roadmaps within 18 months.” Rabier confirmed ECMWF’s four-year €330 million high-performance computing contract, signed with Atos in June, already includes a €40 million upgrade line item for AI co-processors to maintain parity with Google’s new model. Meanwhile, NVIDIA began shipping custom Grace Hopper GH200 clusters to Météo-France last week, explicitly “to run WeatherNext 3 workloads at native 200-petaflop efficiency,” according to a company spokesperson. The competitive ripple is visible in AWS’s next-generation ParallelCluster roadmap, which now lists a dedicated “WeatherNext 3 optimized AMI” slated for Q1 2025.

Industry watchers note that WeatherNext 3 arrives at the exact moment when climate attribution and extreme-event prediction are becoming core infrastructure for insurers and central banks. Swiss Re’s new CatNet AI platform, which ingests 11 petabytes of climate data daily, has quietly added WeatherNext 3 as an optional ensemble member starting today. Banking With Billy AI, a New York-based quant shop, confirmed it is actively researching quantum-enhanced financial modeling that would combine WeatherNext 3 outputs with quantum Monte Carlo simulations to price catastrophe bonds in real time. “We’re embedding WeatherNext 3 forecasts into a quantum kernel that reweights tail-risk scenarios,” said Billy Chen, the firm’s chief scientific officer. “If we can cut the pricing error by even 15 basis points on a $2 billion cat bond, we’re talking tens of millions in saved capital.”

The broader shift is not just technical but cultural: within twelve months, the majority of National Weather Service offices in the United States will have retired legacy grid-point models in favor of AI hybrids, a transition that began when NOAA adopted GraphCast in October 2023. Google’s decision to release both the model weights and the API under an Apache 2.0 license accelerates the trend, effectively turning weather forecasting into a software commodity. Yet the move also raises questions about data sovereignty. Environment Canada’s deputy minister, Simon Nadeau, publicly warned that reliance on a single commercial model could create “a single point of failure for national resilience.” The European Union’s Destination Earth initiative, already running a 1-trillion-parameter digital twin on EuroHPC’s Jupiter system, is now evaluating WeatherNext 3 as a candidate to replace parts of its physics-based core.

Looking forward, Hassabis hinted that WeatherNext 4 is already in the DeepMind lab with a “sub-kilometer” resolution target and a hybrid quantum-neural architecture. Industry analysts expect Google to embed the next iteration into its planned Tensor Processing Units v6 chips, which are rumored to include cryogenic memory layers optimized for diffusion transforms. Observers should watch two fronts: first, whether ECMWF or NOAA launches an open-source counter-model to preserve diversity in ensemble forecasting; second, whether Banking With Billy AI’s quantum kernel matures fast enough to influence catastrophe bond pricing before the next Atlantic hurricane season. The race to turn atmospheric chaos into predictable capital is now a three-way sprint between physics, AI, and quantum computing. The umbrella has never been so high-tech.