NOAA is moving America's weather forecasting supercomputer to Google Cloud

The National Oceanic and Atmospheric Administration is moving the computers behind its daily weather forecasts from machines it owns to capacity it rents from Google Cloud, in a procurement shift the agency calls its cloud-first doctrine for forecasting.

The platform at the center of the move is the Weather and Climate Operational Supercomputing System, the federal supercomputing stack that runs the Global Forecast System and the Global Ensemble Forecast System. These are the models that produce the daily outlooks, hurricane tracks, and severe-weather warnings Americans see on television and in their weather apps. The cloud transition covers both systems, with cutover running through early 2027 and full completion by December 2027.

Why rent instead of own? NOAA Administrator Neil Jacobs said the on-premise supercomputers have become a bottleneck, with fixed capacity and slow chip refreshes, and that moving to the cloud gives the agency access to the latest processors, on-demand scale, and a single compute stack for research and operations. That last point is the structural one: a promising experimental model does not have to wait for a procurement cycle to be tested at operational scale.

The underlying hardware is Google Cloud's H4D virtual machines, powered by fifth-generation AMD EPYC processors, designed for the tightly coupled, compute-intensive workloads that numerical weather prediction generates. National Weather Service Director Ken Graham said the move charts a course toward global leadership in cloud-based forecasting and gives the agency operational modeling that can absorb new science as it emerges, from higher-resolution storm models to machine-learning emulators.

NOAA and Google have worked together since at least 2011, when the agency moved its internal collaboration tools to Google Workspace. Since then the partnership has expanded to shared petabytes of environmental data, real-time wildfire tracking, marine conservation analytics, and a cloud-compute pilot for NOAA Fisheries. More recently, Google DeepMind's weather research has become the framework for NOAA's AIGFS model suite, the agency's first set of machine-learning-powered forecast models. Google's own weather lab, WeatherNext, has shown what that stack can do in production: the company says it predicted Category 5 Hurricane Melissa's landfall five days in advance.

WCOSS being first matters beyond the Google name. Google Cloud has framed it as one of the first operational numerical weather prediction centers globally to move to the public cloud, a procurement experiment other federal high-performance computing workloads will be measured against. If the model holds, with elastic capacity, faster chip refreshes, and research and operations on one stack, other agencies running large scientific workloads are likely to face the same pressure. If it does not, the second-order questions will be sharper: what happens when a single commercial cloud outage coincides with an active hurricane, what the egress bill looks like as forecast datasets are redistributed, and how the agency preserves the option to move back to owned hardware if terms change.

For now, the watch items are concrete. The Global Forecast System and its ensemble are the first operational models on the new stack, and full cutover is scheduled for December 2027. The agency has not publicly disclosed the contract value or the multi-year cost envelope of the arrangement; that number, when it surfaces in procurement records, will be the cleanest test of whether renting beats owning for federal forecasting.