Processing Sensitive Text on a Cloud GPU instance

Privacy and Security Get Advice
1

I’m exploring processing sensitive text data in the cloud while keeping it confidential. Here’s what I intend to do:

  1. Encrypt the data locally before uploading to cloud storage.
  2. Run workloads inside a TEE-enabled GPU (thinking NVIDIA H200) so that decryption and computation happen only in the enclave.
  3. Deliver the decryption key remotely using attestation: the key is encrypted specifically for the TEE and released only after verifying the enclave.
  4. Retrieve results — plaintext never leaves the TEE.
    Key questions:
    Which cloud GPU providers offer robust confidential computing or TEE support suitable for AI workloads?
    Any recommendations for managing keys and attestation reliably at scale?
    Any gotchas when using TEEs for GPU-heavy AI pipelines (memory limits, driver support, multi-GPU)?
    Thank you in advance.
2

I have a feeling this forum is not the most appropriate place for this hyper-technical question. Probably a better place would be NVIDIA Developer Forum’s Confidential Computing section or r/ConfidentialComputing.

I’m tangentially-related as dev-ops for a global company using Azure Key Vault and Hashicorp Vault at scale for many teams, some of them using Azure GPU’s, so maybe look into them for key management, but that’s the extent of my knowledge in this domain.

2 Likes