Loading page…
The agency aims to have its cloud-based forecasting environment operating at scale by early 2027, giving it more room to combine fast AI models with physics-based forecasting.
Listen to this story
Rather than replacing its aging supercomputers one-for-one, NWS will use Google Cloud as the primary HPC provider for WCOSS, with an at-scale environment expected by early 2027. The move could make it easier to allocate compute for AI and hybrid forecasts, especially during severe weather, and help advance cloud-native systems such as Warn-on-Forecast from demonstration toward operations.
NWS will begin with early access, then performance-test final scale; 2027 marks infrastructure readiness, not universal production deployment.
Existing AI models AIGFS, AIGEFS, and Hybrid-GEFS were developed partly in Google DeepMind’s cloud and deployed on-premises.
Cloud flexibility addresses differing compute footprints of numerical and AI forecasting, including demand spikes during severe weather.
The National Weather Service has selected Google Cloud as the primary high-performance-computing provider for a move that will take the backbone of its weather-modeling operations out of agency data centers. The shift is meant to give NWS more flexibility as AI-based forecasting develops, with a full-scale cloud environment targeted for the start of 2027.
The system being moved is the Weather and Climate Operational Supercomputing System, or WCOSS. NWS currently relies on two operational supercomputers in on-premises data centers to run the weather models that support its forecasts. Agency officials are treating the expiration of that equipment as an opening to modernize rather than simply replace it.
Traditional numerical weather prediction and AI prediction have different computing footprints, according to NWS central operations director David Michaud. An on-premises system has a finite set of compute resources; cloud infrastructure can, in the agency’s view, shift the balance of those resources more quickly as needs change, including during severe weather.
That flexibility does not mean NWS is abandoning physical simulation. The agency expects future forecasting systems to combine traditional physics-based models with AI models, seeking the strengths of both approaches. Its officials also caution that AI weather prediction remains at an early stage and can be limited in extreme-weather situations.
NWS already runs AI-based models: AIGFS, a global forecasting model; AIGEFS, a global ensemble model that provides probabilistic information; and Hybrid-GEFS, which combines a traditional numerical model with AI. Some of that AI-model work was developed in Google DeepMind’s cloud environment before being deployed to NWS’s existing on-premises infrastructure.
Warn-on-Forecast has been operating in demonstration mode, but NWS Meteorological Development Laboratory director Richard Bandy said the lack of an operational cloud environment had complicated putting it into production. Moving the broader production suite to the cloud gives the agency a path toward operationalizing projects built for that setting.
NWS plans to reduce risk with a smaller early-access environment before testing performance at final scale. The early-2027 target is therefore a milestone for infrastructure readiness, not evidence that every cloud-designed forecasting project will be operational by then. The immediate test is whether the agency can validate performance while moving a system that underpins its model suite.
Loading discussion...
Make your case
A view to debateHybrid forecasting is safer than relying on AI alone.
Explain what you would need to trust an AI forecast.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.