Intel TDX: I Benchmarked Encrypted vs Regular Inference — 5.2% Overhead. That Is It.

Quick Answer: Running AI inference inside Intel TDX enclaves adds just 5.2% latency overhead compared to non-encrypted inference. On an H200 GPU, encrypted inference took 630ms vs 600ms for the regular model. Cost per inference? $0.55 for TDX, $0.50 ...