Supermicro NVIDIA BlueField-4 STX storage server for AI inference

Supermicro unveiled one of the industry's first context memory storage servers as part of the NVIDIA STX reference architecture announced at NVIDIA GTC 2026. The BlueField-4 STX storage server combines the NVIDIA Vera CPU and NVIDIA ConnectX-9 SuperNIC. It builds upon the Petascale JBOF all-flash array powered by NVIDIA BlueField-3 that the company introduced last year. The server addresses the challenge of long-lived AI queries and multi-stage chain-of-thought agentic workloads that require prior and intermediate tokens associated with the user's query to be accessed. The solution accelerates results and reduces power consumption by avoiding recomputation when local storage for the tokens is exceeded. The key value cache for the tokens is managed by NVIDIA Dynamo, NVIDIA's inference orchestration layer. Supermicro will work with software partners on porting and validation. The company will also enable testing with SSD providers such as Micron, Samsung and Phison for the specific STX architecture requirements. The CMX server is being shown in Supermicro booth 1113 and at the NVIDIA exhibit at NVIDIA GTC 2026 March 16 through 19. Charles Liang, president and chief executive officer of Supermicro, said the company continues to be first to market with new rack scale architectures designed to exceed the needs of a rapidly evolving AI Factory customer base. He said the prototype demonstrates the level of collaboration with NVIDIA and the commitment to be first to market with new technologies.