NVIDIA has unveiled BlueField-4 STX, a modular reference architecture designed to help enterprises and AI providers build accelerated storage infrastructure capable of supporting agentic AI. Traditional data centers lack the responsiveness needed for AI agents that work across multiple steps and sessions, creating a bottleneck for advanced AI applications that require real-time data access and contextual working memory.

The new architecture includes the NVIDIA CMX context memory storage platform, delivering up to 5x tokens per second compared to traditional storage. It provides 4x higher energy efficiency and can ingest 2x more pages per second. Major storage partners including Dell Technologies, HPE, NetApp, and Nutanix are codesigning next-generation infrastructure based on STX, with manufacturing partners like Supermicro and Quanta Cloud Technology building systems.

STX is accelerated by the NVIDIA Vera Rubin platform and uses a new BlueField-4 processor with ConnectX-9 SuperNIC. Leading AI labs and cloud providers including Mistral AI, Lambda, and Vultr plan to adopt STX for context memory storage. Platforms based on the architecture will be available from partners in the second half of this year.