Dev.to Security 🔐 Cybersecurity 👁 0 📖 8 min read

RTX PRO 6000 Blackwell in an Intel TDX VM: the proof files, re-checked from my laptop

"Confidential" is an easy word to put on a product page. This post shows the raw files behind it for one product, then what happened when I verified the same evidence again today on my own machine. Most checks passed; on

"Confidential" is an easy word to put on a product page. This post shows the raw files behind it for one product, then what happened when I verified the same evidence again today on my own machine. Most checks passed; one failed in offline mode, and I left it in because it is the most useful part.

I run VoltageGPU, which rents an NVIDIA RTX PRO 6000 Blackwell inside an Intel TDX virtual machine. The files below were captured inside one of those VMs on 17 September 2026 and are public.

What you rent

From our public pricing feed (/api/pricing/snapshot), built on 22 November 2026 at 11:30 UTC:

rtx6000b-small
GPU 1x RTX PRO 6000 Blackwell Server Edition
GPU memory (as listed) 96 GB
vCPUs 32
RAM 128 GB
Price $3.80 per hour for the whole VM
In stock at that reading 10
Enclave Intel TDX

Billing is per second: the first hour is charged when the VM starts, and the unused part is refunded when you stop it. A run on 9 October that deployed a VM through the API, attested it and stopped it used 161 seconds of VM time; $3.63 of the $3.80 came back, so it cost $0.17. A VM left running for a full month of 730 hours costs $2,774.

One VM, one GPU. Nothing here covers multi-GPU setups.

What the VM read about itself

Three small files, captured as a tenant inside the VM.

kernel.txt (uname -r, ls -l /dev/tdx_guest, systemd-detect-virt):

6.8.0-110-generic
crw------- 1 root root 10, 262 Sep 16 23:45 /dev/tdx_guest
kvm

gpu.txt:

name, driver_version, vbios_version, memory.total [MiB]
NVIDIA RTX PRO 6000 Blackwell Server Edition, 595.71.05, 98.02.9E.00.01, 97887 MiB

conf-compute.txt (nvidia-smi conf-compute -q and -gm):

==============NVSMI CONF-COMPUTE LOG==============

    CC State                   : ON
    Multi-GPU Mode             : None
    CPU CC Capabilities        : INTEL TDX
    GPU CC Capabilities        : CC Capable
    CC GPUs Ready State        : Ready

GPU 0: NVIDIA RTX PRO 6000 Blackwell Server Edition (UUID: GPU-9ab167db-eeab-d1bd-d7fd-e9c21f845a58)
Protected memory size:   99463296 KiB
Unprotected memory size: 0 KiB

The lines that matter: the GPU runs with confidential computing on (CC State : ON), it is alone in the VM (Multi-GPU Mode : None), and none of its memory is reported as unprotected.

All of these lines are printed by software inside the VM. They tell you what the guest saw; they are not something you can check from outside. The two signed proofs are.

The two proofs, and what binds them

Inside the VM, the tenant writes a manifest with a random challenge, then asks for both proofs at once. These are the commands from the capture's README (the VM image ships without pip):

curl -sS https://bootstrap.pypa.io/get-pip.py | python3 - --user --break-system-packages
python3 -m pip install --user --break-system-packages "voltage-verify[attest]"
export PATH=$HOME/.local/bin:$PATH
voltage-verify manifest --challenge auto -o manifest.json
sudo env PATH=$PATH PYTHONPATH=$(python3 -c 'import site;print(site.getusersitepackages())') \
  voltage-verify attest --manifest manifest.json --mode single-gpu -o bundle.json

The bundle holds:

  1. An Intel TDX quote (version 4, 4,940 bytes) whose report_data field is the SHA-512 of the manifest. It is signed through a chain that ends at Intel's SGX root CA.
  2. NVIDIA attestation tokens from NVIDIA's remote attestation service (NRAS) whose eat_nonce is the SHA-256 of the same manifest.

Both proofs carry a hash of the same manifest, and the manifest carries the challenge. If you pick the challenge yourself, an old bundle can't be passed off as a new one.

Re-checking it today

I downloaded the bundle and manifest from the evidence folder and checked them against the checksums computed inside the VM (22 November 2026, 11:46 UTC, Windows 11, Git Bash, GNU coreutils 8.32):

curl -fsSO https://voltagegpu.com/blog/two-proofs/evidence/rtx6000b-2026-09-17/bundle.json
curl -fsSO https://voltagegpu.com/blog/two-proofs/evidence/rtx6000b-2026-09-17/manifest.json
curl -fsSO https://voltagegpu.com/blog/two-proofs/evidence/rtx6000b-2026-09-17/SHA256SUMS
sha256sum -c --ignore-missing SHA256SUMS
bundle.json: OK
manifest.json: OK

Then I verified it with voltage-verify 0.3.0 from PyPI (MIT), in a fresh venv, with the challenge from manifest.json (11:44:55 UTC):

pip install "voltage-verify>=0.3.0"
voltage-verify verify bundle.json --challenge d807768657e40ee62d430d9dc5eb924d520ba1d07e351bc44ee59dbe611085a7 --hwmodel GB20X --gpus 1
[PASS] manifest.format       manifest is well formed
[PASS] manifest.commitments  bundle commitments match the manifest
[PASS] manifest.challenge    challenge is the one this verifier issued
[PASS] tdx.structure         TDX quote v4, TEE 0x81, 4940 bytes
[PASS] tdx.debug             TD attributes 0000001000000000 (0x10000000), DEBUG (bit 0) off: not a debug TD
[PASS] tdx.report_data       quote report_data equals SHA-512 of the manifest
[WARN] tdx.measurements      TD software not pinned: pass --expect-mrtd to check it
[PASS] tdx.signatures        attestation key, QE binding, QE report and PCK chain verify up to the pinned Intel SGX Root CA
[PASS] tdx.collateral        TCB info and QE identity signed by Intel (Intel PCS, fetched now)
[PASS] tdx.tcb               platform TCB UpToDate, QE UpToDate, TDX_03 UpToDate
[PASS] tdx.revocation        PCK certificate not revoked (57 entries in CRL)
[PASS] tdx.revocation        PCK CA certificate not revoked (0 entries in CRL)
[PASS] nvidia.signatures     2 NVIDIA-signed tokens verify, each GPU token listed by its digest in the verifier token, keys from the bundle chained to a pinned NRAS intermediate (pinned intermediates only)
[PASS] nvidia.nonce          NRAS eat_nonce equals SHA-256 of the manifest (verifier token and 1 GPU token(s))
[PASS] nvidia.policy         overall result true, every GPU: measurements ok, secure boot on, debug off, nonce matched, report signature and chain validated
[PASS] nvidia.freshness      token issued 0s from bundle creation
[PASS] nvidia.mode           attestation mode 'single-gpu'

TDX: MRTD eea8b6a814569a52bd1e12f6b869bb2d...  TCB UpToDate  QE UpToDate
GPU: GPU-0 GB20X driver 595.71.05 vbios 98.02.9E.00.01 measurements success

RESULT: VERIFIED

In that run the Intel collateral (TCB info, QE identity, revocation lists) was fetched from Intel at verification time. --hwmodel GB20X --gpus 1 makes the verifier also require the GPU model claim GB20X and exactly one attested GPU.

The offline check that failed

The same command with --offline uses the Intel collateral stored in the bundle instead (11:44:48 UTC):

voltage-verify verify bundle.json --offline --challenge d807768657e40ee62d430d9dc5eb924d520ba1d07e351bc44ee59dbe611085a7 --hwmodel GB20X --gpus 1
[PASS] manifest.format       manifest is well formed
[PASS] manifest.commitments  bundle commitments match the manifest
[PASS] manifest.challenge    challenge is the one this verifier issued
[PASS] tdx.structure         TDX quote v4, TEE 0x81, 4940 bytes
[PASS] tdx.debug             TD attributes 0000001000000000 (0x10000000), DEBUG (bit 0) off: not a debug TD
[PASS] tdx.report_data       quote report_data equals SHA-512 of the manifest
[WARN] tdx.measurements      TD software not pinned: pass --expect-mrtd to check it
[PASS] tdx.signatures        attestation key, QE binding, QE report and PCK chain verify up to the pinned Intel SGX Root CA
[PASS] tdx.collateral        TCB info and QE identity signed by Intel (bundle (fetched 2026-09-17T00:00:26+00:00))
[FAIL] tdx.collateral_fresh  Intel collateral is 66 days old (TCB info issued 2026-09-16), over the 30-day limit: refetch online, or accept it with --max-collateral-age 67
[PASS] tdx.tcb               platform TCB UpToDate, QE UpToDate, TDX_03 UpToDate
[PASS] tdx.revocation        CRL for the PCK certificate is past its next update (2026-10-16)
[PASS] tdx.revocation        PCK certificate not revoked (57 entries in CRL)
[PASS] tdx.revocation        PCK CA certificate not revoked (0 entries in CRL)
[PASS] nvidia.signatures     2 NVIDIA-signed tokens verify, each GPU token listed by its digest in the verifier token, keys from the bundle chained to a pinned NRAS intermediate (offline)
[PASS] nvidia.nonce          NRAS eat_nonce equals SHA-256 of the manifest (verifier token and 1 GPU token(s))
[PASS] nvidia.policy         overall result true, every GPU: measurements ok, secure boot on, debug off, nonce matched, report signature and chain validated
[PASS] nvidia.freshness      token issued 0s from bundle creation
[PASS] nvidia.mode           attestation mode 'single-gpu'

TDX: MRTD eea8b6a814569a52bd1e12f6b869bb2d...  TCB UpToDate  QE UpToDate
GPU: GPU-0 GB20X driver 595.71.05 vbios 98.02.9E.00.01 measurements success

RESULT: NOT VERIFIED (tdx.collateral_fresh)

That is the tool doing its job. The collateral in the bundle was fetched on 17 September, and the revocation list it carries is past its next-update date (16 October), as one line says. voltage-verify verify --help describes the 30-day default as "the validity Intel gives TCB info, QE identity and the PCK CRL". Past that, an offline verifier can't tell whether Intel has since downgraded the platform or revoked a certificate. Verify online, or accept an older reading knowingly with --max-collateral-age. For comparison, on 9 October a new bundle from another VM of the same size was verified offline that same night (voltage-verify 0.2.1) and returned VERIFIED.

With a wrong challenge, the replay check fails as it should (offline, all zeros as the challenge; the two lines that changed):

[FAIL] manifest.challenge    challenge differs from the one this verifier issued (replay or foreign bundle)
RESULT: NOT VERIFIED (manifest.challenge, tdx.collateral_fresh)

The warning

[WARN] tdx.measurements TD software not pinned. The quote includes measurements of what booted in the trust domain: MRTD for the initial firmware image, RTMR0 to RTMR3 for firmware configuration, boot loader, kernel command line and what the OS adds (from --help). Unless you pass the values you expect with --expect-mrtd and --expect-rtmr*, the verifier checks that the quote is genuine, not which software is running.

What this proves, and what it doesn't

It shows that on 17 September 2026 a VM of this size was an Intel TDX trust domain, not in debug mode, on a platform Intel rated up to date. It also shows that its RTX PRO 6000 ran in confidential-computing mode with NVIDIA-signed tokens bound to the same manifest. Today, the bundle still checks out online.

It does not show:

  • That your VM is in the same state today. nvidia.freshness compares the token with the bundle's creation time, not with today. This bundle is from September. Run attest in your own VM with a challenge you pick.
  • Which software runs in the VM, unless you pin the measurements (the warning above).
  • Protection against someone with physical access to the server.
  • That the trust domain and the GPU are in the same machine, or that your SSH session ends inside the attested trust domain. Our proofs page explains each of these limits and cites its sources.
  • Anything about performance. I ran no benchmark for this post. 96 GB is the listed size; the measured figures are the ones above.
  • Anything about multi-GPU. This is a one-GPU VM.

I did not rent a VM for this post: the commands inside the VM are the ones from the September capture, and only the verification side was run today. The product page lists the same commands pinned to voltage-verify>=0.3.0.

Drafted with AI assistance, then checked: the checksum and verify commands were run on 22 November 2026 and their output is copied as printed (for the wrong-challenge run, only the two lines that changed); the VM files are copied from the evidence published on 17 September 2026; the price was read on our pricing feed on 22 November 2026.

📰 Read the original article on Dev.to Security

Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.