Nvidia Says Blackwell GPUs Eliminate 30%-40% AI Security Hit for Sensitive Data
Updated
Updated · Computerworld · Jun 30
Nvidia Says Blackwell GPUs Eliminate 30%-40% AI Security Hit for Sensitive Data
3 articles · Updated · Computerworld · Jun 30
Summary
Blackwell and newer Nvidia GPUs let enterprises run confidential computing for AI without the 30% to 40% throughput penalty that previously made secure deployments uneconomic, senior director Dion Harris said.
Confidential computing keeps data encrypted in memory and between CPUs and GPUs, decrypting it only inside a hardware-rooted trusted zone during computation to protect models and sensitive records from unauthorized access.
Nvidia says that matters as agentic AI moves into cloud and hybrid environments where companies want to use customer, medical and financial data without exposing it to administrators or cloud operators.
About 70% of data still sits outside the cloud, Harris said, making confidential computing increasingly central for regulated sectors and broader enterprise AI adoption.