Bring your own GPU
Run SCSA on hardware you own or rent — and be clear-eyed about who can unseal in each arrangement.
You don't have to run on anyone else's hardware. SCSA is software that runs on a GPU host you control: on-prem, in your own cloud account, or rented on your own provider account.
Deploying
Build the node binaries on the GPU host from the bundled source. Only a CUDA
12.x driver is required — no nvcc and no external download, which is what makes
an air-gapped install possible.
scsa-vault attest # confirm the GPU attests, and at which tier
scsa-vault cc-check # confirm whether Confidential Computing is activeRun these first. They tell you the honest ceiling of what this hardware can enforce before you seal anything to it. See Sovereign mode for the full set of node binaries and Proofs for how to read the verdicts.
Rented GPUs
Renting works the same way: rent the card on your own provider account and run SCSA on it. 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 arrangements, and they differ by who can unseal:
| Arrangement | Who drives it | Who can unseal | You trust |
|---|---|---|---|
| Managed — a provider rents and runs the GPU for you | provider | provider | the provider and their infrastructure |
| BYO, provider-driven — you own or rent, but authorize their access | provider | provider | the provider |
| Sovereign / self-hosted — you run everything | you | you | nobody but the hardware |
The middle arrangement is convenient, but it means someone else has shell access on your box and could, in principle, derive the key. That is fine for a trial. For a bank or a sovereign workload, use sovereign mode, or put the model into threshold custody so that no single party, including the operator, can unseal alone.
If you also need each use to be individually authorized, logged, and checkable by a third party afterwards, that is the Authority Fabric.
What stops fraud
- Faking a passing attestation: you can't. The attestation report is signed by the GPU vendor's key, verified against a pinned embedded root. The nonce is read from inside the signed body and compared in constant time, so a captured report cannot be replayed against a fresh challenge. On a box without real Confidential Computing the gate fails closed.
- Minting extra identities from one certificate chain: the chain is parsed with no trailing slack, and the identity measurement folds in only vendor-signed material — not values the host can claim for itself.
- 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, statelessly verifiable by anyone. There is no trapdoor to fake one.
One escape hatch remains by design: SCSA_INSECURE_SKIP_ATTESTATION bypasses
the attestation gate. It exists for development on machines without a GPU. Treat
its presence in any production environment as a finding.