LLM_Deploy
Talk to an engineerMenu

Run private AI on GPUs you don't own.

NVIDIA confidential computing on H100, H200 and Blackwell. Your model runs inside a hardware-sealed enclave, and the company that owns the server sees encrypted traffic and an attestation record. It does not see your prompts, the answers or the weights.

Have your own GPUs? The same service runs inside your network.

Type a prompt. See who can read it.

The same request, seen from both ends. You get a plain answer. The company that owns the GPU sees encrypted memory traffic and an attestation record, and nothing it can read.

Your view

Sample

> Summarise the open claims for policy 4471-B

Sample response. This page runs no model, so this is a canned reply. In production your model answers inside the sealed enclave, and the reply is readable only at your end of the encrypted session.

Plain text, at your endpoint only.

Attestation

Demo
  1. CPU TEE report verifiedPass
  2. GPU in confidential-computing modePass
  3. GPU attestation report signed by NVIDIAPass
  4. Driver and firmware measurements matchPass
  5. CPU to GPU transfers encryptedPass
  6. Keys released to the enclave onlyPass

6 of 6 checks passed

Hardware owner's view

Sample

Unreadable encrypted bytes. The hardware owner cannot see your prompt or the answer.

Readable by the host: nothing above.

Illustration only. No model runs on this page and nothing leaves your browser. The bytes on the right are a SHA-256 stream of your prompt, so they change with every keystroke, the way real ciphertext would.

What the hardware owner can and cannot see.

Confidential computing changes whom you have to trust. You trust the CPU and GPU hardware, and the attestation that proves it was switched on. You no longer have to trust the company that owns the server, its administrators or its hypervisor.

The host cannot see

  • Your prompts and the answers. They are decrypted only inside the enclave and at your endpoint.
  • Your model weights and fine-tunes. They load into the GPU’s protected memory region, which the host cannot read.
  • The memory of the confidential VM. The CPU’s trusted execution environment encrypts it, so the hypervisor sees ciphertext.
  • The session keys. They are released to the enclave only after attestation passes.

The host can still see

  • That a workload is running. Power draw, GPU utilisation and uptime stay visible to the operator.
  • The shape of the traffic. The size and timing of encrypted requests, and how many arrive.
  • The attestation record. The signed proof that the enclave was sealed when it started.
  • Whether the server stays on. Confidential computing protects confidentiality, not availability. The owner can still switch it off.

Out of scope: physical attacks on the GPU package, side channels beyond the vendors’ published mitigations, and flaws in your own application code. This is a threat model, not a promise that nothing can ever be broken. Ask an engineer to walk your security team through it for your deployment.

The parts that make it work

GPUs
NVIDIA H100, H200 and Blackwell, running in confidential-computing mode.
GPU memory
Your data and weights sit in a protected VRAM region that hardware firewalls close to the host CPU and to other GPUs. Everything crossing PCIe between the CPU and the GPU is encrypted with AES-GCM.
CPU TEE
The confidential VM runs in a CPU trusted execution environment, AMD SEV-SNP or Intel TDX, which encrypts its memory.
Attestation
The CPU and the GPU each sign a report of their firmware and mode. Both are checked against the vendors’ certificates before any key or data is released.

Confidential compute

By quote

Each confidential deployment is priced on four things:

  • The GPU generation and how many you need.
  • Where the hardware runs: a cloud’s confidential VMs, a colocation host or a partner’s rack.
  • The model, its size and the traffic it serves.
  • The attestation policy your security team signs off.
Ask for a quote

Have your own GPUs?

The same service runs inside your network, on hardware you control. No enclave is needed when nobody else holds the server.

How an on-prem deployment works
Pilot
$4,900
Production deployment
from $25,000
Managed ops
from $1,500
All prices

Tell us whose GPUs.

Tell us the model, the data and where the hardware is: a cloud, a colocation host or a partner’s rack. An engineer from Limassol or Auckland replies with a confidential-compute plan and a quote.

limassol.officeLocal time

 

Limassol HQ

Cyprus, 24/7

+357 25 123877
auckland.officeLocal time

 

Auckland

New Zealand, 24/7

+64 9 887 9088