Sealed before it thinks.
Your question is answered inside a machine that encrypts its own memory. The company that hosts it can't look in, and the machine proves it to anyone who asks.
The path of one question
You ask. It leaves encrypted.
What the host sees
It can see
- That a request arrived, and when
- How much computing time it used
- Encrypted memory it has no key for
It never sees
- Your question
- The model's answer
- Anything left in the machine's memory
Check it yourself
Live attestation launch dayA report anyone can read
A public report the machine generates about itself, signed by its hardware.
Intel signs the machine
The quote proves the code runs in Intel TDX, with memory encrypted by the processor itself.
NVIDIA signs the GPU
NVIDIA's attestation service confirms the GPU is genuine and running in confidential mode.
Fresh, not replayed
Each report carries a new random value, so an old report can't be passed off as today's.
What the proof doesn't cover yet
Today the attestation proves the machine and its GPU are sealed. It doesn't yet name the exact model weights inside. That's the next step: a fingerprint of our weights in every report, so you can check which model answered you.
ComingFive free messages. No wallet.
Ask anything