Deploying Proprietary Models Securely with NVIDIA Confidential Computing on Self-Hosted Kubernetes

So turns out that this whole thing with “The DRM-protected movie gets decrypted on the GPU and the GPU composites that into the output invisible to the OS” doesn’t just work for video but any kind of data. LLM weights for example.

Could probably even go so far as to having to buy a per-GPU per-model license? Maybe?

My uninformed guess is that we’re just not seeing this yet, because none of the labs want to hardware-lock themselves to NVIDIA specifically more than they already are.