ETH Zurich spent a year copying 100 petabytes of NASA data onto servers next to Alps

Divide 100 petabytes by a year and the copy ran at an average of about 3.2 gigabytes a second, sustained, for twelve months ^[21]. The file count sets a second pace: roughly 6 billion files over the same period is about 190 files a second, at an average size near 17 megabytes ^{[22] [23]}. Neither rate is exotic for a national facility. Both had to be provisioned and paid for before a single model was trained.

Copying at all was a choice. Thomas Schulthess, the ETH computational physics professor who heads the Swiss National Supercomputing Centre in Lugano ^[2], made the case for locality in terms of wait time. "It matters whether you can move the data within a few seconds or whether you have to wait days for the data to come," he said ^[8]. The servers holding the archive sit a few meters from Alps ^[24]. Training models of this kind takes both large volumes of Earth observation data and significant computing power, and ETH says it now has the pair in one place ^[11].

Reto Knutti, who heads ETH's Center for Climate Systems Modeling, said the models are expensive to train but "cheap to run", so "you can do more iterations, maybe every few minutes, to see if some specific weather pattern exists" ^{[13] [4]}. "That allows us to do early warning systems to save lives," he said ^[14]. The interviews give no figure for the training bill and none for the cost of a single run.

Nature reported last week that satellite image analysis showed some warning signs before the Aug. 26 glacial collapse on the Nepal-China border, and that spotting them could have flagged the area "as a hotspot warranting closer attention and monitoring" ^[18]. That work was done after the collapse. What it does not say is how many other slopes carried the same signature and stayed intact. That count is what sets a false alarm rate.

Blatten is the closest thing here to a positive control. It was done without AI. Knutti said the Swiss village's destruction by a glacier collapse in May last year "was visible in satellite data more than a year before it actually happened" ^[16]. The evacuation that avoided mass casualties came a week ahead of the collapse, out of close monitoring by Swiss authorities ^[17].

Knutti said "Data is essentially everything," and that "The next step will be making sense of the data" ^{[9] [10]}. Schulthess called it "a huge scientific opportunity" ^[6] and said it is "really enabling scientists to do things we would not even have thought of before" ^[7]. The account given to AFP describes the archive and the machine it sits beside. No forecast produced by these models appears in it.