Skip to content

Clusters

A cluster is one rack. It has one NVLink fabric, one power budget, one InfiniBand uplink to storage and one scheduler queue. We do not stripe jobs across racks. When you ask for 8 GPUs you get one machine, not a collection.

C1

C1 is the first cluster. It was commissioned on 2026-08-19 in Hillsboro, Oregon, and listed after 72 hours of soak testing.

ComponentSpecification
GPU8x NVIDIA H100 SXM 80 GB
FabricNVLink 4, 900 GB/s per GPU
Network3.2 Tb/s InfiniBand to storage
CPU2x AMD EPYC 9454, 96 cores
Memory2 TB DDR5
Storage30 TB NVMe local, 400 TB shared
Regionus-west
Commissioned2026-08-19

All 8 GPUs sit on one NVLink switch fabric. Any GPU can reach any other at 900 GB/s bidirectional. For a single node this is the fastest interconnect you can buy, and it is why we start with SXM parts rather than PCIe.

The InfiniBand link goes to the shared storage cluster only. There is no second rack to talk to yet.

Power and cooling

C1 draws up to 10.2 kW at full load. It is on a dedicated 15 kW circuit with A and B feeds. Cooling is rear door heat exchange. GPU temperature under sustained load sits between 58 C and 72 C. The telemetry feed shows it per GPU.

Commissioning

A new cluster goes through the same steps before it is listed. Hardware acceptance from the vendor. Firmware and driver pinning to the versions we run everywhere. 72 hours of soak testing at 95 % or higher utilization on all GPUs, with NVLink and storage bandwidth checks every hour. Telemetry agent installed and reporting to the public feed. Only then does it appear in the capacity endpoint.

If soak testing finds a fault, the clock restarts after the fix. The next cluster is listed when it passes, not on a date.

Software

Ubuntu 24.04 LTS, NVIDIA driver 560, CUDA 12.6 on the host. Containers run under the NVIDIA container runtime and can bring their own CUDA 12.x userspace. Host software is upgraded during a maintenance window announced 7 days ahead in the status history.

What is next

The next rack is eu-central in Frankfurt. It uses the same bill of materials. See regions.

RACK SYSTEMS, us-west. Console and API access is by invitation during preview.