NVIDIA adds Confidential Bare Metal for AI model makers

NVIDIA tells about Confidential Bare Metal, a new alternative in its Confidential Computing line. It lets makers of frontier models (the largest AI models) operate sensitive work on dedicated servers that a different company owns. The makers keep control of their models and data. Cohere has a plan to use it, and Google Cloud adds it to A4X Max instances. NVIDIA gives no test results or speed numbers.

How it works

NVIDIA adds controls for data in transit, a check of the platform and access limits in the hardware. Builders can operate their own software on bare metal, which is a server with no virtual layer. 3 alternatives are available:

• Confidential Bare Metal for bare metal servers
• Confidential Containers for Kubernetes systems
• Confidential Virtual Machines for virtual servers

The support

10 cloud providers have deployments of NVIDIA Confidential Computing. 6 systems makers will have certified systems with these functions.

What is not known

Cohere only has a plan to use it. Google Cloud tells that serving performance is top line, but it gives no number. The source does not tell the cost of the protection in speed.

Insight: NVIDIA sells trust in the hardware below the software walls

Perplexity made its SPACE sandbox open source. The sandbox keeps AI agents in set limits, and CoreWeave will give it to its customers. CoreWeave also has NVIDIA Confidential Computing deployments. Software walls are free, but the hardware protection comes from NVIDIA, which reads as a plan for NVIDIA to hold the gate below each wall. Find out which layer holds your protection.