Nvidia Launches BlueField-4 STX and CMX Platform to Accelerate AI Inference Storage

Agentic AI repeatedly reads and writes KV caches when reasoning over long contexts. Moving that data between GPU memory and conventional storage can slow inference and reduce GPU utilization. Nvidia designed BlueField-4 STX as a modular reference architecture, while CMX provides a high-speed, rack-scale context layer. Together, they give cloud and server providers a common set of deployment specifications.

Nvidia unveiled BlueField-4 STX and CMX, its first rack-scale implementation, at GTC 2026 on March 16, 2026. The company said the technology can deliver up to 5 times the token throughput, 4 times the energy efficiency and 2 times the data-ingestion speed of conventional storage. AIC, Supermicro and Quanta Cloud Technology, or QCT, have begun manufacturing the systems. The platforms are scheduled to become available in the second half of 2026, with pricing yet to be announced.