Overview

Confidential AI runs open-source language models inside an Intel TDX trust domain with NVIDIA H100/H200 confidential-compute mode enabled. Prompts and responses are decrypted only inside the enclave, the model weights are loaded from a dm-verity protected disk, and every chat session can prove with a fresh remote-attestation handshake which exact code and which exact weights produced its answers.

Why this matters

Today's hosted LLMs require unconditional trust in the operator: the provider can read every prompt, log every response, fine-tune on user data, or substitute a quietly modified model. There is no cryptographic mechanism a user can run to verify what was on the other end of the API call.

Confidential AI removes that trust. The same hardware-rooted attestation chain that protects Privasys' confidential workloads also protects the model serving runtime:

• The TDX quote pins MRTD + RTMR0..3, proving exactly which kernel, initrd, and userspace booted.
• A custom x.509 extension on the enclave certificate (OID 1.3.6.1.4.1.65230.3.5) carries the dm-verity root hash of the model disk.
• The same RA-TLS handshake exposes the loaded model name, the proxy status, and any per-request challenge nonce, so a client can prove the response it received came from the attested binary.

What you get

• A drop-in OpenAI-compatible /v1/chat/completions endpoint that streams responses through the enclave.
• Cryptographic proof of the model weights, the inference server binary, the GPU firmware, and the host platform, returned alongside every session.
• A reproducible, signed image pipeline that lets you (or an auditor) rebuild the enclave bit-for-bit and confirm the measurements.
• Operator transparency: the management plane can lock model load/unload behind a fleet token but cannot read prompts, responses, or KV-cache contents.

Where to go next

• Architecture - the components inside and around the enclave.
• Attestation - the certificate extensions, RTMR replay, and challenge protocol.
• Models - dm-verity model disks and how they get published.
• API - the OpenAI-compatible endpoints exposed by the proxy.
• Privacy guarantees - what the operator can and cannot see.