The brief

Analyst view

• Alps: 10,752 GH200 superchips, 434.9 PFlop/s, No. 10 on the June 2026 TOP500.
• Trained Apertus, a fully open 8B/70B model, on about 40% of Alps for three to four months.
• Cooled with Lake Lugano water; waste heat feeds district heating.

Technical approach

As reported

vClusters

One machine, many clusters

A software layer splits Alps into partitions tailored to each user group, from MeteoSwiss forecasts to AI training, like a private cloud.

Hardware

Grace Hopper at scale

CSCS was the first customer for NVIDIA's Grace Hopper superchips, which pair a CPU and GPU sharing fast memory on one module.

Cooling

Lake water, zero consumption

Water drawn from Lake Lugano cools the machine and is returned; the waste heat goes into the local district heating network.