Accelerators
GPUs you can actually get
Single nodes by the hour, clusters by the month, and a queue that tells you the truth about waiting times.
Accelerator fleet
| GPU | Per node | VRAM | Interconnect | Availability | €/GPU/hour |
|---|---|---|---|---|---|
| L4 | 4 | 24 GB | PCIe 4.0 | immediate | 0.42 |
| L40S | 4 or 8 | 48 GB | PCIe 5.0 | immediate | 1.55 |
| H100 SXM | 8 | 80 GB | NVLink + 400G IB | 2–5 days | 2.30 |
| H200 SXM | 8 | 141 GB | NVLink + 400G IB | 1–3 weeks | 3.10 |
| MI300X | 8 | 192 GB | Infinity Fabric + 400G IB | immediate | 2.05 |
Cluster details
400G InfiniBand
Non-blocking fat tree within a pod of 64 nodes. Rail-optimised, with SHARP enabled for collectives.
Shared scratch
A 2 PB NVMe parallel filesystem at 400 GB/s aggregate, mounted on every node in the pod.
Dataset ingress
Pull training data at 100 Gbit/s per node. Ingress and egress are both unmetered — moving 200 TB of images in costs nothing.
Real telemetry
Per-GPU utilisation, memory, ECC counts and NVLink bandwidth in Prometheus format, retained for a year.
Honest queues
The availability column is a measured median, updated hourly. If it says three weeks, it means three weeks.
Single tenant
No GPU is ever shared or partitioned across customers. When you release a node it is wiped and the firmware re-flashed.