Comment by FrasiertheLion

That's basically what we built at Tinfoil. We run open source models inside secure enclaves (also using Intel TDX/AMD SEV-SNP + NVIDIA Confidential Computing). All the code running inside the enclave is open source and the client SDKs (also open source) automatically verify that the pinned source code matches the runtime attestation. The protocol used is TLS (terminates in the enclave) + HPKE keys generated inside the enclave on boot. Docs walk you through the verification process: https://docs.tinfoil.sh/verification/verification-in-tinfoil

Of course, we can't support Claude or Grok as they are closed source, but there is no incentive for companies that need your data to train the next generation of models to allow for private inference. One day...