Dedicated bare metal
A whole machine for one tenant: 8x NVIDIA H100 SXM on a full NVLink mesh, 192 vCPU, root over SSH, no hypervisor. No enclave, not confidential: memory is not encrypted against the operator and there is no Intel TDX quote or GPU attestation to generate. You check the hardware yourself. Need the data sealed from the operator? Use a Confidential VM (Intel TDX, attestation you verify yourself).
Bare metal vs the other tiers
| Dedicated bare metal | Confidential VM | Standard GPU | |
|---|---|---|---|
| Hardware enclave (Intel TDX) | No | Yes, memory encryption, tenant attestation | No |
| Hypervisor | None: the OS runs on the hardware | Yes (the trust domain is a VM by design) | Container on a shared host |
| GPUs | 8x H100 SXM, full NVLink mesh, one tenant | 1 GPU per VM, or 8-GPU nodes in Protected PCIe mode | Whole machines from a distributed provider network |
| What you verify | hwcheck, systemd-detect-virt, nvidia-smi topo -m | TDX quote and NVIDIA attestation, on your nonce | hwcheck |
| Automatic release | Mandatory: 24 h by default, 72 h at most | Optional | Optional |
| Best for | Multi-GPU training on one quiet node, benchmarks | Regulated or proprietary data | Public data, experiments |
Who can deploy (as of 2026-10-24)
- A verified account (email confirmed) at least one day old, with one successful payment.
- One machine at a time per account.
- A prepaid balance covering the first hour of the machine; one hour is charged at start.
- At least one SSH key: the keys are bound at creation and are the only way in.
A request that does not meet these rules answers 403 with {"error", "code": "NOT_ELIGIBLE"}. Other codes: BALANCE_TOO_LOW, SOLD_OUT, FEATURE_DISABLED. Nothing is charged on an error.
Deploy from the dashboard
- Open voltagegpu.com/dedicated for the live list (price, region, stock), then Deploy.
- Pick the machine, an image (Ubuntu 24.04 LTS, delivered without the NVIDIA driver: install the data center driver and Fabric Manager first; you log in as ubuntu), your SSH keys and the automatic release (24 h by default, 72 h at most).
- Read the summary: price per hour for the whole machine, price per GPU-hour, the hour charged now, your balance.
- Deploy. The pod page shows the SSH address once the machine answers, and the three checks to run first.
Deploy with the API (as of 2026-10-24)
Same deployment as the dashboard, with your API key (X-API-Key). The machines are in the public pricing snapshot under dedicated.tiers; deploy through POST /api/volt/pods with tier: "dedicated".
# Machines, price and stock (public, no key needed)
curl -s https://voltagegpu.com/api/pricing/snapshot | jq '.dedicated.tiers'
# [{"resourceName":"bm-a5-h100-sxm","name":"8x H100 SXM","gpuModel":"H100","gpuCount":8,
# "vcpu":192,"pricePerHour":49.8,"available":2,"region":"eu-west"}]
# Deploy a whole machine (one hour charged at start, then per second)
curl -X POST https://api.voltagegpu.com/api/volt/pods \
-H "X-API-Key: $VOLT_API_KEY" -H "Content-Type: application/json" \
-d '{"tier": "dedicated", "resourceName": "bm-a5-h100-sxm",
"image": "ubuntu-24.04", "sshKeyIds": ["<ssh_key_id>"], "hardstopHours": 24}'
# {"pod": {"id": "<pod_id>", "status": "PENDING", "hourlyPrice": 49.8, "hardStopAt": "..."}}
# Poll until ssh_command is set, then release (unused part of the first hour refunded)
curl -H "X-API-Key: $VOLT_API_KEY" https://api.voltagegpu.com/api/volt/pods/<pod_id>
curl -X DELETE -H "X-API-Key: $VOLT_API_KEY" https://api.voltagegpu.com/api/volt/pods/<pod_id>hardstopHours is mandatory in effect: omitted, it is 24; above 72 it is refused. Body fields: tier, resourceName, image, sshKeyIds (at least one), hardstopHours. Full reference: API reference, Dedicated bare metal.
Check the machine, as root
Three read-only commands, in this order. Another reading than the expected one: do not put data on the machine, release it within the first hour and send us the output at contact@voltagegpu.com.
# 1. GPU model, count, VRAM and a matmul run, compared with the listing
curl -fsSL https://voltagegpu.com/hwcheck.sh | bash -s -- --gpu 'H100' --gpus 8 --vram 80
# expect: RESULT: MATCHES THE LISTING
# 2. No hypervisor
systemd-detect-virt
# expect: none
# 3. Full NVLink mesh between the eight GPUs
nvidia-smi topo -m
# expect: every GPU-to-GPU cell reads NV<n> (an NVLink path), never SYS, PHB or PIXNot proven by these commands: the BMC and its firmware (not readable from the OS, no published measurement), the VBIOS and system firmware versions (reported by nvidia-smi, attested by nothing), and the erasure of the disks between tenants. The hardware check script is public (MIT): github.com/VoltageGPU/hwcheck.
Billing
- One hour of the whole machine is charged at start (
pricePerHourof the tier, $49.80 for the 8x H100 SXM as of 2026-10-24, that is $6.23 per GPU-hour), then billing is per second. - Release early and the unused part of the first hour comes back to your balance.
- The automatic release is mandatory (24 h by default, 72 h at most) and destroys the disk with the machine: copy your checkpoints off before the timer.
- A
sudo poweroffstops the machine, not the rental: release it from the pod page or withDELETE. - The machine is released automatically if your balance reaches zero.