Hi everyone,
I’m Zoheb, founder of Custodian. I’d like to explore an integration that combines NVIDIA confidential computing with shared authority over model release.
Custodian addresses a practical deployment challenge: model builders need to protect valuable models and retain release authority, while deploying customers need local operation and independently verifiable control.
Each party holds a separate share of the model-decryption key. Both must authorize release to the approved, attested runtime, where the key is reconstructed. An enforced lease bounds serving; either holder can refuse the next authorization.
Already tested: Custodian has been tested on NVIDIA H100 and RTX PRO 6000 Blackwell Server Edition, within Intel TDX confidential VMs. Laboratory runs demonstrated real inference, rejection of tampered evidence, and actual serving-worker shutdown following expiry or detected withdrawal.
Building on the H100 and Blackwell lab results, I’d like to qualify the complete deployment with independently administered holders and an approved measured serving image.
Proposed next step: qualify one named model on a supported confidential-GPU configuration, combining an owner-approved measured serving image, independently administered holders, and NVIDIA’s supported attestation tooling.
For model builders, the value is enforceable release conditions. For customers, it is an independent role in authorizing models running on their infrastructure.
I’d welcome a technical discussion with NVIDIA or an infrastructure partner to review this integration and identify a confidential GPU environment where I can deploy Custodian’s approved VM image and test the complete setup. Who would be the right contact?
Zoheb Shaik
Founder, Custodian · Dubai, UAE
https://custodylabs.co
