I noticed the problem on a documentation detail that most people skim past.

@OpenGradient Chat routes prompts through OHTTP relays before they reach a TEE-isolated LLM proxy node. The relay sits between the user and the TEE specifically so the TEE node never sees the originating IP. Anonymity and privacy, handled in sequence.

I assumed those two things were the same guarantee. That felt obvious.

That was the first mismatch.

The OHTTP relay decouples your network identity from the request. The TEE enclave ensures the node operator cannot read your prompt. But those are two separate protections with two separate failure surfaces. The relay sees your IP but not your prompt. The TEE node sees your prompt but not your IP. The privacy guarantee only holds if both layers work correctly, simultaneously, and neither is compromised at its seam.

Geography isn't coordination. That's what I kept turning over.

The dependency chain runs like this. User sends a request. OHTTP relay strips IP metadata and forwards the encapsulated payload. TEE proxy node receives it, decrypts inside the enclave, calls the upstream LLM provider via attested connection, returns the response. Neither layer alone delivers the full guarantee. Both have to hold.

What I can't resolve is the relay itself. It's not a TEE. It's a standard network intermediary. Apparently that's intentional OHTTP relays are designed to be trust-minimized, not trust-eliminated.

I'm less certain now about what happens when the relay is operated by a single entity during a congestion event and becomes a bottleneck.

If the relay layer slows down under load, does the anonymity property degrade before the privacy property does?

What's more important for AI infrastructure?

#OPG #opg $OPG