Private AI: Building Trust Through Verifiable Computation
summary
Claims that user data never becomes visible to the provider while still running LLM workloads on cloud GPUs and that auditors can verify that claim remotely. Architecture uses confidential computing TEEs with Private Verifiable Compute for CPU and NVIDIA H100 attestation, plus user-controlled key-encryption for stateful data, differential privacy and oblivious HTTP for metadata protection. Demo shows browser receives a hardware-signed attestation and verifies runtime and model; project is open source with monthly releases and supports cloud, on‑prem and hybrid deployment. 230 million weekly health queries motivates the design.
AI-written summary · about
chapters (8)
- 0:00Private AI introduction
- 3:40Sensitive data challenges
- 6:32Confidential computing
- 8:19Private Compute goals
- 12:15Private AI architecture
- 13:20Verifiable privacy trust
- 15:27Private AI business value
- 17:15Open source attestation demo
related talks
13mProvably Secure AI for Health Data – Powered by Confidential Computing
24mStrictly Confidential: Use Cases & Demos of AI Workflows in TEEs
15mKeynote: Trustworthy AI with Confidential Computing
20mBuilding Trusted AI Factories with NVIDIA Accelerated Confidential Computing
47mTowards Confidential AI for the Masses!
28mWhy Private AI Matters: Data Sovereignty as the Foundation for Trustworthy AI on the Open Web
17mThe Roadmap to a User-Owned AI Model
41mAgentic AI, confidential computing and attested TLS