…/3

Chat, speech-to-text and image models run on NEAR AI Cloud in an Intel TDX VM with NVIDIA confidential GPUs. For each one this server sends a fresh nonce and checks: Intel signed the report (dcap-qvl), the model key and nonce are inside it, and NVIDIA NRAS vouches for every GPU. Every chat reply is signed by that key. Run it yourself: npm run attest.

Not attested: our app server and the Cartesia voice. Not checked: that the TDX measurements match a published NEAR AI build.