Request capacity

Overview / Compute

GPU infrastructure

Current-generation accelerators in whole machines, built into clusters with a fabric that belongs to one customer at a time. Every part of the system is named, published and inspectable.

Open the catalogPricing
Rack elevation: two top-of-rack switches, a management switch and four 8U GPU nodes; nodes B and C belong to one tenant inside a hatched boundaryU48U44U40U36U32U28U24U20U16U12U8U4U0NODE ANODE BNODE CNODE DPOWER · COOLING2× LEAF SWITCHMANAGEMENTONE TENANTWHOLE NODES,OWN PORTS8U HGX NODE48U · 600 MM
Elevation · 48U rack · schematic

The node is the
smallest thing we rent.

An eight-GPU host is never split between customers. A single-GPU instance is a whole, smaller machine — not a slice of a large one.

Plan of an eight-GPU H200 node enclosed by a single-tenant boundary (schematic)ABCDEFGH123456TENANT BOUNDARY · ONE CUSTOMERGPU0GPU1GPU2GPU3GPU4GPU5GPU6GPU7CPU0CPU1NIC0NIC1NIC2NIC3NIC4NIC5NIC6NIC7123456
Plan · HGX H200 node · schematic

cl.h200.x8

Sample host spec
Accelerators
8× NVIDIA H200 SXM
GPU memory
8 × 141 GB HBM3e · 1.1 TB total
GPU interconnect
NVSwitch · 900 GB/s per GPU
Host CPU
2× Xeon Platinum 8480+ · 224 threads
System memory
2 TB
Local NVMe
30.7 TB
Compute fabric
8× 400 Gb/s, one per GPU
Front-end network
2× 100 Gb/s Ethernet
All H200 configurations

A fabric with
one tenant on it.

Clusters use a rail-optimised two-tier fabric: GPU n in every node connects to leaf switch n, so collective operations stay one hop apart. Each cluster gets its own fabric partition, and its leaf ports are not shared.

Rail-optimised fabric: four nodes, eight leaf switches, four spine switches. NIC number i on every node connects to leaf i.SPINE 0SPINE 1SPINE 2SPINE 3LEAF 0LEAF 1LEAF 2LEAF 3LEAF 4LEAF 5LEAF 6LEAF 7NODE 0 · 8 GPUNODE 1 · 8 GPUNODE 2 · 8 GPUNODE 3 · 8 GPU400 Gb/s PER GPU · NON-BLOCKING
Highlight rail
Per-GPU bandwidth
400 Gb/s, InfiniBand NDR or RoCEv2
Oversubscription
None, leaf to spine
Isolation
Partition keys or VRFs per tenant
Front end
Separate Ethernet for storage and ingress

Deployment forms

The same hardware, held three ways. What changes is the size of the boundary and how long you keep it.

NodeClusterEnclosure
Unit1–8 GPUs on one host16–1,024 GPUsDedicated racks or a caged area
Tenancy boundaryThe whole hostHosts plus their fabric ports and partitionPhysical space, power feed and network edge
Compute fabricNVLink / NVSwitch inside the nodeRail-optimised, 400 Gb/s per GPU, non-blockingDesigned with you; your edge routers
ProvisioningMinutes, self-serveDays, scheduled with youWeeks, built to specification
TermPer second, or 1–12 month reservation1–12 month reservation12 months and longer
Confidential modeOn supported GPUsOn supported GPUsOn supported GPUs, plus your own HSM
OrchestrationYour choiceManaged Slurm or Kubernetes, or bring your ownYour choice, with optional management
Physical accessOurs, loggedOurs, loggedEscorted access for your staff
Configure a nodePlan a cluster or enclosure

Software and storage

Images
Ubuntu 24.04 LTS with the NVIDIA driver, CUDA, NCCL and Fabric Manager pre-installed — or your own signed image.
Orchestration
Managed Slurm or Kubernetes on clusters, with topology-aware scheduling that keeps jobs on the same rails.
Storage
Local NVMe on every node; network block volumes and S3-compatible object storage in the same facility.
Observability
GPU, fabric and thermal telemetry exported to your Prometheus endpoint. We keep only what operating the hardware requires.