NVIDIA Unveils BlueField-4 STX Architecture to Address AI Storage Bottlenecks

Quick Facts

• NVIDIA BlueField-4 STX delivers up to 5x faster token throughput and 4x better energy efficiency than traditional storage systems
• The modular architecture combines CPU, networking, and storage processing to eliminate bottlenecks in agentic AI workloads
• STX-based platforms will be available from partners including Dell, HPE, IBM, and NetApp in the second half of 2026

NVIDIA announced BlueField-4 STX, a modular reference architecture designed to solve storage bottlenecks in agentic AI systems at the GTC developer event on March 16, 2026. The architecture delivers performance improvements critical for AI agents that require continuous context memory across multiple sessions and reasoning steps.

The STX system combines the NVIDIA Vera CPU for complex logic processing, ConnectX-9 SuperNIC for ultra-low-latency data transfer, and Spectrum-X Ethernet for AI factory-scale networking. This integrated approach targets the key-value cache data bottleneck that occurs when AI models store intermediate calculations to maintain context during inference.

“Agentic AI is redefining what software can do — and the computing infrastructure behind it must be reinvented to keep pace,” said Jensen Huang, NVIDIA founder and CEO. “AI systems that reason across massive context and continuously learn require a new class of storage.”

The architecture’s first implementation is the NVIDIA CMX context memory storage platform, which extends GPU memory with a high-performance layer specifically designed for storing and retrieving data generated by large language models. This rack-scale system can ingest 2x more pages per second for enterprise AI data compared to traditional storage systems.

NVIDIA’s approach represents what the company calls “extreme co-design,” treating the entire data center as a single integrated unit rather than separate networking and storage components. The STX architecture bypasses host CPUs by routing data through a dedicated accelerated storage layer using RDMA over Spectrum-X networking.

The announcement comes as NVIDIA reported record Q4 2026 revenue of $68.1 billion, up 73% year-over-year, with fiscal 2026 revenue reaching $215.9 billion. The company guided Q1 FY27 revenue to $78.0 billion with gross margins near 75%.

Storage providers including Cloudian, DDN, Dell Technologies, Hitachi Vantara, HPE, IBM, NetApp, and VAST Data are co-designing next-generation AI infrastructure based on STX. Manufacturing partners AIC, Supermicro, and Quanta Cloud Technology will build STX-based systems.

Early adopters planning STX implementation for context memory storage include CoreWeave, Crusoe, Lambda, Mistral AI, and Oracle Cloud Infrastructure. The convergence of networking and storage administration reflects the infrastructure demands of AI agents that require microsecond-level response times.

Read more: The convergence of context: Why Nvidia’s BlueField-4 STX marries the network and storage admin