NOAA announced this week that Google Cloud will serve as the primary high-performance computing (HPC) provider for the Weather and Climate Operational Supercomputing System (WCOSS) — the infrastructure that generates every US national weather forecast, hurricane warning, and aviation weather alert. The agency becomes one of the first operational numerical weather prediction (NWP) centers globally to move tightly-coupled atmospheric simulation workloads from dedicated government supercomputers to public cloud. Consequently, the full transition is targeted for December 2027.
What WCOSS Does — and Why Moving It to Google Cloud Matters
WCOSS is not a side project. It runs more than 20 operational NWP models continuously, feeding output into every National Weather Service bulletin — from daily forecasts to severe weather warnings. The existing system consists of twin HPE Cray supercomputers at Manassas, Virginia and Phoenix, Arizona, delivering 29 petaflops combined under a $505 million, 10-year contract with General Dynamics Information Technology signed in 2020. Therefore, replacing this with cloud VMs is not a minor infrastructure update. It is one of the largest migrations of government-critical HPC workloads to public cloud ever attempted.
Moreover, Google Cloud’s H4D virtual machines, powered by AMD EPYC 5th-generation processors, will replace this hardware as the main computing backbone. NOAA stated the cloud transition will eliminate traditional bottlenecks of on-premises systems, enabling a more flexible, updatable modeling environment. Workloads — including the Global Forecast System and Global Ensemble Forecast System — will begin migrating in early 2027, according to Google’s official announcement.
Cloud RDMA: How Google Solved the Tightly-Coupled Problem
The reason cloud HPC skeptics always pointed to weather prediction as their strongest example is valid: NWP models are tightly-coupled. During each time step of an atmospheric simulation, thousands of nodes must exchange data constantly. The inter-node latency requirements are strict. Traditional HPC clusters use proprietary interconnects — InfiniBand, Cray Slingshot — specifically built for this communication pattern. However, standard cloud networking historically could not compete.
H4D VMs change that equation. Google’s Cloud RDMA runs over the Titanium network adapter, delivering 200 Gbps inter-node bandwidth with low latency — competitive with dedicated HPC interconnects. Furthermore, the instances scale to tens of thousands of cores for large MPI workloads, which is exactly the communication pattern numerical weather models use. Each H4D node delivers more than 12,000 GFLOPS of compute and more than 950 GB/s of memory bandwidth. In fact, independent benchmarks found cloud HPC generating NWP forecasts in roughly 53 minutes — about half the time of comparable on-premises runs. HPCwire’s coverage of H4D’s general availability noted the platform was designed precisely for tightly-coupled workloads: computational fluid dynamics, molecular dynamics, and weather modeling.
WeatherNext Already Proved the Cloud-AI Combination Works
Before the WCOSS migration was announced, the Google-NOAA collaboration already delivered a tangible result: NOAA used Google DeepMind’s WeatherNext AI model to predict Hurricane Melissa’s Category 5 landfall in Jamaica five days in advance. That lead time exceeded what traditional physics-based NWP models provided. Additionally, WeatherNext excels at predicting both track and intensity by training on decades of global weather patterns alongside specialized tropical cyclone datasets — capabilities that complement, rather than replace, physics-based simulation.
Running both WeatherNext and conventional NWP models in the same cloud environment creates a practical hybrid: AI models provide rapid initial forecasts during fast-evolving storms, while physics models produce the detailed 7-day outlooks that aviation and emergency management depend on. In contrast, this combination is not practical when your AI infrastructure is at Google and your NWP system is at a government data center in Virginia. Cloud consolidation is what makes it operationally viable. The Hurricane Melissa result is detailed in Google DeepMind’s writeup on the prediction.
The Validation Event Cloud HPC Has Been Waiting For
Architects debating cloud versus on-premises for demanding HPC workloads typically face three objections: latency, sovereignty, and cost. NOAA’s decision addresses all three. Cloud RDMA answers the latency objection directly. Meanwhile, the US government moving critical national safety infrastructure to Google Cloud addresses the sovereignty objection — or at least demonstrates the government itself has made that call. Moreover, replacing a fixed $505 million hardware contract with cloud’s variable cost model suggests NOAA’s own economic analysis came out in favor of cloud.
Furthermore, this is not an isolated case. The European Centre for Medium-Range Weather Forecasts (ECMWF) is building cloud-based infrastructure through the European Weather Cloud initiative. The National Weather Service is separately migrating data and applications to NWS HIVE and NWS CIRRUS cloud platforms, under contracts awarded in March 2026. As a result, the pattern across major weather centers globally is consistent: on-premises dedicated HPC is being phased out in favor of cloud infrastructure. NOAA’s WCOSS migration is the largest and most operationally significant step in that trend to date.
Key Takeaways
- NOAA selected Google Cloud to replace on-premises Cray supercomputers for WCOSS — 29 petaflops of critical national weather infrastructure moving to H4D VMs by December 2027
- Cloud RDMA on Google’s H4D instances delivers 200 Gbps inter-node bandwidth, directly addressing the latency barrier that previously blocked tightly-coupled NWP workloads from public cloud
- The NOAA-Google partnership already demonstrated value: WeatherNext AI predicted Hurricane Melissa’s Category 5 Jamaica landfall five days in advance — ahead of traditional NWP models
- For enterprise architects, the combination of government mandate, Cloud RDMA performance data, and economic evidence removes the three strongest objections to cloud HPC migration
