Seven Layer
GPU Model Protection

Bring your own GPU

Register an attested GPU you already control — on-prem or in your own cloud — by its SSH endpoint.

You don't have to run on anyone else's hardware. Register a GPU you rent or own by its SSH endpoint, and the platform deploys and runs SCSA there.

The flow

# 1. get the control-plane key and authorize it on your GPU host
scsa gpu pubkey
#    -> run the printed line on the box:
#       mkdir -p ~/.ssh && echo '<key>' >> ~/.ssh/authorized_keys

# 2. register the endpoint
scsa gpu register gpu01.internal 22 --label my-h100

# 3. build SCSA on it (skipped automatically if already built)
scsa gpu deploy <id>

# 4. run the proofs
scsa prove <id>

register probes the box over SSH, reads its GPU identity, and — if SCSA is already built there — marks it ready without rebuilding. De-registering only removes it from your list; your hardware is never touched.

Rented GPUs

Renting works the same way: rent an H100 on your own provider account, authorize the key, and register the endpoint. You pay the provider for the GPU; SCSA is the software layer on top. The weight key is bound to that specific GPU's attestation, so a protected model can't be moved to a cheaper card.

Who trusts whom

Be clear-eyed about this. There are three models, and they differ by who can unseal:

ModelWho drives itWho can unsealYou trust
Managed — the operator rents + runs the GPUoperatoroperatoroperator + their infra
BYO, operator-driven — you rent/own, authorize their keyoperatoroperatoroperator
Sovereign / self-hosted — you run everythingyouyounobody but the hardware

The flow above is the middle model: convenient, but it means the operator has SSH on your box and could, in principle, derive the key. That's fine for a trial — but for a bank or a sovereign workload, use sovereign mode, or put the model into threshold custody so no single party, including the operator, can unseal alone.

What stops fraud

  • Faking a passing attestation: you can't. The attestation report is signed by the GPU vendor's key (a certificate chain to an embedded root). On a box without real Confidential Computing it fails closed. See Proofs.
  • Moving a protected model to another GPU: the key is bound to the attested measurement — a different GPU derives a different key, and decryption fails.
  • Forging a compliance proof: the audit proofs are transparent STARKs, stateless- verifiable by anyone. There is no trapdoor to fake one.

On this page