Open vs. Closed Weight Models and Why You Need Confidential Inference Either Way

The article contrasts open/API-based and closed/self-hosted LLM deployments, highlights the inference/runtime period as the critical risk where model weights and sensitive prompts can be exposed, and advocates confidential inference (hardware TEEs, attestation, and hardened key management) as a necessary control for protecting model IP and meeting regulatory and compliance needs in production AI systems.