Skip to content

Confidential VM · 8-GPU nodes

8x H100 and 8x H200 confidential GPU nodes on Intel TDX

The 8x H100 node costs $55.62 per hour for the whole VM ($6.95 per GPU-hour). The 8x H200 node costs $72.00 per hour for the whole VM ($9.00 per GPU-hour). Each node is a VoltageGPU Confidential VM (Intel TDX) with eight GPUs and root over SSH, billed per second with no commitment. Stock on 22 November 2026 at 13:40 UTC: 1 eight-GPU H100 VM available, 1 eight-GPU H200 VM available.

Both nodes run NVIDIA's multi-GPU Protected PCIe mode, where nvidia-smi reads CC State OFF next to Multi-GPU Mode Protected PCIe. On 4 November 2026 we ran the attestation from inside each node and published the files: the Intel TDX quote and NVIDIA's attestation of all eight GPUs.

Specs, price and stock

A Confidential VM price covers the whole VM, every GPU in it; a standard GPU price is per GPU-hour.

Prices and stock read from our live inventory on 22 November 2026 at 13:40 UTC, in US dollars.
Node8x H1008x H200
Resourceh100-xlargeh200-xlarge
GPUs8x NVIDIA H100 80 GB8x NVIDIA H200 141 GB
Total VRAM640 GB1,128 GB
vCPU96128
RAM1,400 GB1,400 GB
Price per hour, whole VM$55.62$72.00
Price per GPU-hour$6.95$9.00
Stock1 VM available1 VM available
TEE and GPU modeIntel TDX; NVIDIA multi-GPU Protected PCIe modeIntel TDX; NVIDIA multi-GPU Protected PCIe mode
Attestation published4 November 20264 November 2026
Single-GPU pageNVIDIA H100NVIDIA H200

What was attested on each node

Only what we ran from inside a tenant VM and published, file by file. Each folder holds the bundle, the verifier’s output and the checksums.

8x H100 node

h100-xlarge

All eight GPUs first attested on 10 September 2026; attested again on 4 November 2026, with the Intel TDX quote in the same bundle.

  • Intel TDX quote, captured on 4 November 2026 from 22:28 to 22:31 UTC inside a tenant VM (h100-xlarge): version 4, 4,936 bytes, not a debug trust domain, platform TCB rated UpToDate by Intel, report_data equal to the SHA-512 of the run's manifest.
  • NVIDIA's remote attestation service (NRAS) returned "Attestation Successful" for all eight GPUs, on a nonce drawn outside the VM: measurements success, secure boot on, debug disabled, driver 595.71.05, VBIOS 96.00.CF.00.01.
  • The quote and the GPU tokens sit in one voltage-verify bundle, verified outside the VM, online and offline; the same bundle under another challenge is rejected.
  • Each GPU reports 82,542,720 KiB of protected memory and 0 KiB unprotected.

8x H200 node

h200-xlarge

Attested on 4 November 2026: the Intel TDX quote and all eight GPUs, in one bundle.

  • Intel TDX quote, captured on 4 November 2026 from 20:09 to 20:12 UTC inside a tenant VM (h200-xlarge): version 4, 4,936 bytes, not a debug trust domain, platform TCB rated UpToDate by Intel, report_data equal to the SHA-512 of the run's manifest.
  • NVIDIA's remote attestation service (NRAS) returned "Attestation Successful" for all eight GPUs, on a nonce drawn outside the VM: measurements success, secure boot on, debug disabled, driver 595.71.05, VBIOS 96.00.CF.00.02.
  • The quote and the GPU tokens sit in one voltage-verify bundle, verified outside the VM, online and offline; the same bundle under another challenge is rejected.
  • Each GPU reports 146,115,584 KiB of protected memory and 0 KiB unprotected.

How NVIDIA’s multi-GPU Protected PCIe mode reads

Both nodes run NVIDIA's multi-GPU Protected PCIe mode. In this mode nvidia-smi reads CC State OFF next to Multi-GPU Mode Protected PCIe: that is the normal reading, and the Multi-GPU Mode line is the one to read.

nvidia-smi conf-compute -q, inside an 8-GPU node
CC State                   : OFF
Multi-GPU Mode             : Protected PCIe
CPU CC Capabilities        : INTEL TDX
GPU CC Capabilities        : CC Capable

On the runs of 4 November 2026 the GPUs read Not Ready at boot and were set Ready with "sudo nvidia-smi conf-compute -srs 1". On these nodes, NVIDIA's attestation SDK must be called with get_evidence(options={"ppcie_mode": False}).

What is not attested, and what this mode does not do

NVLink traffic between the GPUs

NVLink traffic between the GPUs is not encrypted in this mode. NVIDIA's white paper (WP-12554-001 v1.3, page 13) says that data between the CPU and the GPUs is encrypted through bounce buffers, "while data traffic between GPUs over NVLINK remains unencrypted".

NVIDIA's white paper (PDF)

The four NVSwitches of each node

On the 8x H200 and 8x H100 nodes, the 4 NVSwitches of each node passed NVIDIA attestation on 4 November 2026 once libnvidia-nscq matching the driver version was installed in the guest; on the image as delivered, switch attestation fails. Each switch result is a single run, a separate call on its own nonce, outside the voltage-verify bundle and not bound to the TDX quote. We do not claim the NVLink fabric.

The 8x H200 run and its NVSwitch test

The software in the VM, and any other VM

The quote and the tokens describe the hardware, the firmware and the trust domain at the moment of the challenge, not what ran afterwards, and not any other VM. The files we publish come from VMs created for the capture and deleted after it: the proof that counts for you is a fresh run on your own VM, on your own nonce.

Attacks by someone with physical access to the server

A Confidential VM proof does not establish protection against someone with physical access to the server. Each limit, with its sources, is on the proofs page.

What a proof establishes, and what it does not

Dedicated bare metal is another product

An 8x H100 SXM machine is also sold as dedicated bare metal: a whole machine with no hypervisor and no enclave, so no TDX trust domain and no attestation. It is not confidential, and nothing on this page applies to it.

Dedicated bare metal (no enclave, not confidential)

Verify it yourself

On your own node: run nvidia-smi conf-compute -q at first login to read the mode, then draw a nonce on your own machine and run both attestations from inside the VM. The method is in our Confidential VM guide, and the scripts of the 4 November runs are published as they ran, in each evidence folder. Confidential VM guide, GPU attestation

On the files we published: you need Python 3 with pip, no account and no GPU. Download bundle.json from the 8x H200 evidence folder and verify it online. voltage-verify is open source: it checks the TDX quote up to the Intel SGX Root CA pinned in the tool, and the NVIDIA tokens up to a pinned NRAS intermediate. The 8x H100 folder has its own bundle and challenge, listed in its README.

on your machine, with bundle.json of the 8x H200 run
pip install "voltage-verify>=0.3.0"
voltage-verify verify bundle.json --hwmodel GH100 --gpus 8 \
  --challenge 5fe630be5d8455ca3bca9f9bee88d66915c91df39f2136186fbcb2530ef6a206

On 5 November 2026, with voltage-verify 0.3.0, this command ended on "RESULT: VERIFIED", with two warnings that name what no check covers: the TD measurements, which the command does not pin, and the NVSwitch fabric, which the per-GPU tokens do not attest.

Billed per second, from a prepaid balance

Prices are shown per hour and billed per second. The first hour is charged when the machine starts; after that the running time is prorated to the second, and unused prepaid time is refunded when you stop it. At the prices above, the hour charged at start is $55.62 for the 8x H100 node and $72.00 for the 8x H200 node.

Card top-ups are limited to $1,000 per payment at first and can go up to $10,000 per payment for accounts with an established card payment history; the limit for your account is checked when you pay. Bitcoin top-ups go up to $10,000 per quote.

Common questions

How much does an 8x H100 or 8x H200 confidential node cost?

On 22 November 2026 at 13:40 UTC, the 8x H100 node (h100-xlarge) costs $55.62 per hour for the whole VM, $6.95 per GPU-hour, and the 8x H200 node (h200-xlarge) costs $72.00 per hour for the whole VM, $9.00 per GPU-hour. Prices are shown per hour and billed per second. The first hour is charged when the machine starts; after that the running time is prorated to the second, and unused prepaid time is refunded when you stop it.

What is attested on the 8x H100 and 8x H200 nodes?

Both nodes run NVIDIA's multi-GPU Protected PCIe mode. The 8x H200 node was attested on 4 November 2026 (Intel TDX quote and NVIDIA attestation of all 8 GPUs, Protected PCIe mode). The 8x H100 node, first attested on 10 September 2026, was attested again on 4 November 2026 (Intel TDX quote and NVIDIA attestation of all 8 GPUs, Protected PCIe mode). For each node, the Intel TDX quote and the NVIDIA tokens of the eight GPUs sit in one voltage-verify bundle, verified outside the VM, and the files are published with their checksums.

What is not attested on these nodes?

NVLink traffic between the GPUs is not encrypted in NVIDIA's multi-GPU Protected PCIe mode, and no attestation changes that. On the 8x H200 and 8x H100 nodes, the 4 NVSwitches of each node passed NVIDIA attestation on 4 November 2026 once libnvidia-nscq matching the driver version was installed in the guest; on the image as delivered, switch attestation fails. That switch result is a separate call on its own nonce, outside the voltage-verify bundle and not bound to the TDX quote. We do not claim the NVLink fabric.

Why does nvidia-smi show CC State OFF on an 8-GPU node?

Both nodes run NVIDIA's multi-GPU Protected PCIe mode. In this mode nvidia-smi reads CC State OFF next to Multi-GPU Mode Protected PCIe: that is the normal reading, and the Multi-GPU Mode line is the one to read.

How do I verify a node myself?

On your own node, run nvidia-smi conf-compute -q, then generate the Intel TDX quote and the NVIDIA attestation of the eight GPUs from inside the VM, on a nonce you drew on your own machine, and verify the bundle outside the VM with voltage-verify. It is open source: it checks the TDX quote up to the Intel SGX Root CA pinned in the tool, and the NVIDIA tokens up to a pinned NRAS intermediate. The files of our 4 November 2026 runs can be re-checked the same way, with no account and no GPU.

How is a node billed, and how can I pay?

Prices are shown per hour and billed per second. The first hour is charged when the machine starts; after that the running time is prorated to the second, and unused prepaid time is refunded when you stop it. Card top-ups are limited to $1,000 per payment at first and can go up to $10,000 per payment for accounts with an established card payment history; the limit for your account is checked when you pay. Bitcoin top-ups go up to $10,000 per quote.

Is this the same as the dedicated bare metal 8x H100?

No. Dedicated bare metal is a whole 8x H100 SXM machine with no hypervisor and no enclave: no TDX trust domain, no attestation, not confidential. The nodes on this page are Confidential VMs (Intel TDX).

Run the attestation on your own node

The deploy page lists each size our provider lists, at the price a deploy bills.