The part I would not gloss over in OpenGradient is accounting for a private AI call without opening it.
I kept looking at the relay flow because the tension is sharp.
A user sends a sealed OHTTP payload. The relay forwards it and attaches an X-Payment header. The gateway verifies the x402 payment, decrypts inside the enclave, sends the request upstream, then signs the response before sealing it back.
The relay can bill its own users separately, by subscription or per call. But the relay is not supposed to casually see the prompt it is paying to route.
That is where the production burden shows up.
If a user disputes a charge or says an agent made the wrong private request, the builder cannot solve it by exposing the prompt logs. That would break the point of the private path. But they also cannot shrug and say the call happened somewhere inside the system.
They need enough evidence to connect the user charge, the relay payment, and the signed enclave response without turning the private request into a public receipt.
That is the OpenGradient bottleneck I care about here.
Private inference only works if billing can be defended without becoming surveillance.
#OPG $OPG @OpenGradient $VELVET $O