VoltageGPU vs Phala

Short answer for the people who typed "phala pricing" or "phala gpu tee": on 17 September 2026 the Phala GPU TEE page listed the H100 at $3.08 per GPU-hour on demand ($2.38 reserved), the H200 at $4.80 ($3.20 reserved) and the B300 at $6.50 ($5.60 reserved), with a 24 hour minimum on the on-demand H100 and H200 and a 30 day minimum on the B300. Its pricing page headlines GPU TEE from $3.80 per hour. Per GPU-hour that is below VoltageGPU, where a confidential H100 VM is $5.00/hr and a confidential H200 VM is $6.58/hr. For a job that runs a week, Phala is the cheaper attested cloud, and this page says so.

The minimum is where the arithmetic turns. VoltageGPU charges one hour upfront and returns the unused part to your balance per second the moment you release the machine, so a ten minute attestation test on a confidential H100 costs about one sixth of $5.00/hr, and the same test on the cheapest attested SKU, a single RTX PRO 6000 Blackwell, costs well under a dollar. On a 24 hour minimum at $3.08 the same ten minutes cost $73.92. If your work is bursty, iterative, or you simply want to verify the enclave before committing, the minimum matters more than the hourly rate.

Both providers put the GPU inside an Intel TDX trust domain and switch on NVIDIA confidential computing, and both call it dual attestation: Intel signs a quote for the CPU trust domain, NVIDIA signs a report for the GPU. Phala's GPU TEE page states that Intel TDX and NVIDIA each emit a signed quote, and its attestation overview explains that the CVM quote is signed by Intel hardware and checked against Intel root certificates. That is the right design, and it is the same one we use.

Where VoltageGPU is specific: on the Confidential VM tier you hold /dev/tdx_guest yourself. You write a manifest with a random challenge, the TDX quote carries its SHA-512 in report_data, the NVIDIA tokens carry its SHA-256 as nonce, and voltage-verify, an MIT tool on PyPI, checks the whole bundle on your own laptop, offline if you want, and answers NOT VERIFIED on a replayed bundle. We publish the bundles we generated, with checksums, for every SKU we have attested: single H200, single H100, single RTX PRO 6000 Blackwell, and an 8x H100 node in NVIDIA Protected PCIe mode. The public index is at voltagegpu.com/api/attestation/evidence, readable without an account. Whether a Phala tenant can choose the nonce and download the raw quotes is not stated on the pages we read on 17 September 2026; if Phala documents it, we will link it here.

Phala wins on hourly price for long runs, on the B300 which VoltageGPU has never had in inventory to attest, on reserved slots for teams that plan months ahead, and on paper: its trust center title claims SOC 2 and HIPAA, and VoltageGPU has neither a SOC 2 report nor a HIPAA Business Associate Agreement today. If your reviewer starts with the report, Phala answers and we do not.

It is also a broader platform, with confidential containers and an application layer around the CVM, where VoltageGPU sells the raw machine, the API and the proof. A team that wants the platform should look at Phala first.

Three things we can show rather than say. One, the limits: on 8-GPU nodes NVIDIA runs Protected PCIe mode, nvidia-smi reads CC State OFF, all eight GPUs attest, and the NVSwitch fabric does not attest from inside the guest; we publish the failing log rather than the word "attested". Two, the API: since 16 September 2026 the whole sequence, tiers, SSH key, deploy, poll, SSH, attest, verify outside the VM, stop with a per-second refund, runs on an API key alone, and the bundle from that run is public. Three, the price of a test: less than a dollar, refunded to the second.

None of that makes Phala wrong. It makes the comparison concrete: pay less per hour with a day-long minimum and a trust center, or pay per second with the evidence in your hands.