Dedicated bare metal · No enclave, not confidential
A whole 8x H100 SXM machine. No hypervisor, nothing shared, and you check it yourself.
One tenant per machine: eight NVIDIA H100 SXM GPUs on a full NVLink mesh, 192 vCPU, root over SSH, no hypervisor between you and the hardware. Billed per second from a prepaid balance, with an automatic release you set at deploy (24 hours by default, 72 at most).
$49.80/hour for the whole 8x H100 SXM machine, that is $6.23 per GPU-hour, billed per second. 2 available now.
§ 01 / machines
Machines, price and stock
One price for the whole machine, per hour, billed per second. Stock is read live from the inventory.
| Machine | Region | Available | Per GPU-hour | Whole machine | Deploy |
|---|---|---|---|---|---|
| 8x H100 SXM8x H100 · 192 vCPU · No enclave, not confidential | aus | 2 available | $6.23/GPU-h | $49.80/h | Deploy |
One hour of the machine is charged when it starts, then billing is per second; release early and the unused part of that hour comes back. The automatic release is mandatory (24 hours by default, 72 at most) and destroys the disk with the machine: copy your checkpoints off before the timer. Images: ubuntu-24.04. Bare metal is open to verified accounts at least one day old with one successful payment, one machine at a time.
§ 02 / what you can verify
Three commands, run by you, as root
Do not take our word for the machine. The first script compares the GPUs with the listing and runs a matmul; the second shows there is no hypervisor; the third shows the NVLink mesh between the eight GPUs.
Model, count, VRAM and a matmul run
curl -fsSL https://voltagegpu.com/hwcheck.sh | bash -s -- --gpu 'H100' --gpus 8 --vram 80Expect: RESULT: MATCHES THE LISTING
No hypervisor
systemd-detect-virtExpect: none
Full NVLink mesh
nvidia-smi topo -mExpect: every GPU-to-GPU cell reads NV<n> (an NVLink path), never SYS, PHB or PIX
Another reading than the expected one: do not put data on the machine, release it within the first hour (the unused part is refunded) and send us the output at contact@voltagegpu.com. The hardware check script is public: read it, it is short.
§ 03 / honest limits
What these checks do not prove
§ 04 / use cases
What a whole machine is for
Multi-GPU training on one node
Data or tensor parallelism across eight H100 SXM on NVLink, with nobody else on the machine and no hypervisor in the path.
Inference that needs all the memory
Eight GPUs of HBM3 for models that do not fit on one or two cards, served from a machine you benchmarked yourself.
Benchmarks on a quiet machine
Measure what the hardware does, not what a neighbour leaves you. Root lets you install the drivers and tooling you want.
FAQ
Common questions
Is dedicated bare metal confidential?
No. Not confidential: this machine has no enclave. Its memory is not encrypted against the operator and there is no quote or GPU attestation to generate, and we do not claim otherwise. Need the data sealed from the operator? Use a Confidential VM (Intel TDX, attestation you verify yourself).
What can I verify on the machine myself?
Three things, with commands you run as root: curl -fsSL https://voltagegpu.com/hwcheck.sh | bash -s -- --gpu 'H100' --gpus 8 --vram 80 compares the GPU model, count and VRAM with the listing and runs a matmul; systemd-detect-virt must print "none"; nvidia-smi topo -m must show every GPU-to-GPU cell reads NV<n> (an NVLink path), never SYS, PHB or PIX. Not proven by these commands: the BMC and its firmware, the VBIOS and system firmware versions, and the erasure of the disks between tenants.
How is dedicated bare metal billed?
$49.80 per hour for the whole 8x H100 SXM machine, that is $6.23 per GPU-hour. One hour of the machine is charged when it starts, then billing is per second; release early and the unused part of that hour comes back. The automatic release is mandatory (24 hours by default, 72 at most) and destroys the disk with the machine: copy your checkpoints off before the timer.
Who can deploy a bare metal machine?
Bare metal is open to verified accounts at least one day old with one successful payment, one machine at a time. Deploy from the dashboard with your SSH key, or with an API key: POST /api/volt/pods with "tier": "dedicated" (reference: docs.voltagegpu.com/pods/dedicated).
Can I keep the machine longer than 72 hours?
The automatic release is capped at 72 hours per deployment and the disk does not survive it. For longer runs, save your checkpoints elsewhere and deploy again.
Which operating system does it boot?
Ubuntu 24.04 LTS (or 22.04). You have root, so you install the drivers and tooling you need; the hardware check script tells you what the GPUs report.
Deploy it, check it, keep it or release it
Pick the machine, an image and your SSH key; the automatic release is set to 24 hours unless you choose another duration up to 72. One hour is charged at start, then per second.