Breaking your AI storage bottlenecks

The future of AI infrastructure is being reshaped by a fundamental bottleneck: data starvation. As GPUs and AI models grow more powerful, the traditional data storage layer built on commodity x86 hardware can't keep up, creating a performance bottleneck that stalls training and inference. In this episode, Garima Kapoor and Anand Babu Pariyasamy of MinIO reveal how they're working with NVIDIA on a new reference architecture called STX — a purpose-built, ARM-based DPU system designed to eliminate the PCIe, memory, and network bottlenecks that plague legacy systems. By leveraging NVIDIA’s 800G networking, PCIe Gen 6, and HBM-like memory bandwidth in a single system-on-chip, STX enables data to flow to GPUs at unprecedented speeds. MinIO’s software, already optimized for ARM and low-memory environments, is uniquely positioned to run on this new architecture, delivering up to 5x faster read performance and sub-millisecond latency. The conversation also dives into the emerging concept of G3.5 memory — a hybrid layer between storage and memory that stores massive AI context without the cost of HBM, enabling long-term, persistent memory for AI agents. This shift marks a pivotal move from legacy file and block storage toward a software-defined, open-standard object store model that’s essential for scalable, efficient AI factories. The episode underscores a critical truth: the next wave of AI innovation won’t be driven by bigger models alone, but by smarter infrastructure.