Bare Metal
Single-tenant bare-metal nodes for the largest training runs or the strictest isolation requirements.
What it is
Bare metal exposes whole physical nodes — full GPU mesh, full NIC, full NVMe — with no hypervisor between you and the silicon. Used for the largest training runs (1024+ GPUs over InfiniBand), regulated workloads that mandate single-tenant compute, or anyone squeezing the last 5% of performance out of a custom kernel.
When to use it
- Multi-node distributed training across full racks.
- Workloads with strict single-tenant compliance requirements.
- Custom kernels, PCIe-pass-through experiments, hardware tracing.
Quickstart
asc bare-metal request --nodes 16 --gpu h100 --interconnect ib-400g --duration 14dclient.bare_metal.request(nodes=16, gpu='h100', interconnect='ib-400g', duration_days=14)await client.bareMetal.request({ nodes: 16, gpu: 'h100', interconnect: 'ib-400g', durationDays: 14 });Limits & quotas
| Limit | Default | Burst | Notes |
|---|---|---|---|
| Min booking | 1 day | — | |
| Min nodes per request | 1 | — | |
| Max nodes per region | 1024 | — | Lift via sales |
| Interconnect | IB-400G / IB-200G / Ethernet-400G | — |
Pricing
See pricing. Pay-as-you-go, billed monthly via Stripe.
API surface
POST /v1/bare-metal/requestsGET /v1/bare-metal/requests
Required scope(s): pods:write, bare-metal:write. See Scopes.
Security
All access is authenticated and scoped. See Auth & scopes and Network controls.