ETH Zurich’s year-long transfer was not a model launch. It was a decision to keep a continuously updated Earth-observation archive close to Swiss compute, and less dependent on upstream US data services.
On 17 July 2026, ETH Zurich disclosed that researchers had copied approximately 100 petabytes of publicly available NASA Earth-observation data to the Swiss National Supercomputing Centre, or CSCS, in Lugano.
The collection contains roughly 6.5 billion files, according to ETH Zurich. Climate scientist Reto Knutti separately told AFP that completing the initial transfer took about a year.
No new weather model shipped with the announcement. ETH published no training result or benchmark. The milestone was more fundamental: Switzerland now maintains a large, continuously updated replica of NASA data at the same institution that operates Alps, its most powerful supercomputing platform.
That decision has two advantages. The first is data locality. Large AI and scientific workloads benefit from having frequently accessed data close to the systems processing it. The second is operational continuity. A Swiss replica reduces dependence on the continued availability and funding of upstream distribution services for data already copied.
A 100-petabyte archive can cross the Atlantic. Switzerland has just demonstrated that.
The more important question is what happens after the first transfer.
Spread over a year, 100 petabytes represents an average transfer rate of roughly 25 gigabits per second. The reported file count implies an average object size of only about 15 megabytes.
Those figures suggest that the difficulty was not simply moving bytes. Metadata operations, source throttling, validation, retries and the management of billions of objects may have mattered as much as raw bandwidth.
Training and analysis produce different access patterns from archiving. Researchers may repeatedly scan selected parts of the collection, combine products from multiple missions, generate derived datasets and feed distributed jobs running across many accelerators.
Doing that entirely through remote services leaves every workload dependent on external bandwidth, service availability and access interfaces. A local copy turns those dependencies into storage and data-engineering problems that CSCS can manage directly.
Thomas Schulthess, the director of CSCS, told SWI swissinfo that placing the data close to Alps would “greatly simplify” the development of large AI models for climate and weather research.
“Close,” however, does not mean that the full archive occupies Alps’ fastest filesystem.
CSCS operates several storage systems optimized for different workloads, including high-capacity Lustre storage, NVMe-based storage and archival services. ETH and CSCS have not disclosed where the NASA copy resides, whether it is replicated onto tape or how selected datasets will be staged for training.
The defensible conclusion is narrower: CSCS has the infrastructure to retain a collection at this scale and make relevant subsets available to Alps-based workloads.
Alps was inaugurated in September 2024 as an HPE Cray EX platform built around NVIDIA Grace Hopper architecture.
Its main AI partition contains 2,688 four-socket nodes, for 10,752 GH200-class sockets. The system has recorded 434.9 petaflops on the High-Performance Linpack benchmark. It debuted at seventh place on the November 2024 TOP500 list and stood at tenth in the June 2026 ranking.
Alps was designed as a multi-tenant platform. MeteoSwiss runs forecasting workloads on its infrastructure, and Switzerland’s open large language model, Apertus, was trained there.
The GPUs and interconnect were already in place. What changed was the decision to make a foreign-sourced Earth-observation archive a continuously maintained local resource.
It would be easy to describe the project as Switzerland entering the AI weather-model race. That is premature.
The field is already crowded. NVIDIA has its open Earth-2 model family. Google has GraphCast and GenCast. Microsoft has Aurora. The European Centre for Medium-Range Weather Forecasts operates its AIFS alongside physics-based forecasting systems.
These models do not all use the same data or solve the same problem. Many global weather models depend heavily on ERA5 reanalysis, while geospatial foundation models often train on satellite imagery and derived observation products.
The shared lesson is that useful Earth-system AI depends on large, well-maintained collections assembled from a limited number of foundational sources.
The durable asset is therefore not only the model architecture. It is the work beneath it:
Open model weights can be copied quickly. A governed, continuously updated scientific archive connected to national compute infrastructure is harder to reproduce.
The project also has a political dimension, but it should not be overstated.
Thomas Zurbuchen, the ETH professor and former NASA science chief who initiated the transfer, said recent US cost-cutting had “called the future of some data services into question.”
Knutti was more cautious. He told AFP that the United States had not restricted access, while noting that administrative decisions could happen quickly.
The project is therefore not a response to an existing NASA data cutoff. It is insurance against future changes to budgets, portals, processing services or distribution priorities.
A Swiss replica reduces that exposure for historical data already mirrored. It does not make Swiss science independent of NASA.
Future observations still depend on spacecraft, instruments, calibration, mission operations and processing pipelines. The Swiss archive also remains dependent on NASA for new releases.
The benefit is specific: researchers are less dependent on upstream distribution services when accessing data already held at CSCS.
That is better described as operational continuity than complete data sovereignty.
ETH is establishing Swiss GeoLab as part of a broader Earth-observation and climate-modeling strategy. The institution says the GeoLab will be developed over ten years, with one goal being a geospatial AI model trained on diverse observation data.
That does not mean the entire 100-petabyte archive will become one training corpus. Different projects will use different missions, variables, time periods and derived products. Much of the work will lie in selecting, cleaning and restructuring the data before a model sees it.
The value of the replica is that those choices can now be made locally and repeatedly.
The visible competition in Earth-system AI is over models. The quieter competition is over the infrastructure that determines which models can be trained reliably and at scale.
Switzerland has not launched the model yet. The petabytes are what shipped.
The SCN Staff is a small AI editorial squad working under human direction. Each agent owns one job.
Scout does the research. It runs down primary sources and checks what's already been published, on SCN and everywhere else, before a story gets written. If a claim can't be traced back to a real document, Scout flags it.
Forge writes. It takes what Scout found and turns it into a draft, argument and sentences and all. Every SCN piece starts here, then gets sharpened.
Cipher handles search: the titles, descriptions, and keyphrase work that decides whether a good article ever gets found. Least glamorous job on the squad. Also one that matters more than it looks.
Pixel makes the visuals. Images, charts, the occasional diagram, all built to SCN's brand instead of pulled from a stock library. When something's easier to see than to read, it goes to Pixel.
Editorial judgment and the final call stay with the humans. So does the fact-checking.
