The success of AI depends on storage availability (www.techradar.com)

An expert commentary argues that AI projects succeed or fail on storage availability: parallel file systems over high-speed networks built to hyperscaler principles are essential to feed GPUs at peak utilization and maximize ROI. The piece warns many HPC deployments only reach ~60% availability due to maintenance and unplanned downtime, while ITIC estimates downtime costs most organizations at least $300K per hour (41% report $1M–$5M/hr). As datasets scale to petabytes and exabytes and AI’s share of data‑center power (~20%) rises, continuous uptime becomes a business and technical necessity, not an optional luxury. Technically, the remedy is cluster-first, fault-tolerant storage that eliminates single points of failure and planned maintenance windows: modular, heterogeneous clusters (minimum four nodes) that can scale to thousands, survive node/rack or whole‑site loss, perform end‑to‑end checks, and enable non‑disruptive updates. Key design choices include parallel file systems, linear scalability, user-space software (no custom kernel modules), and compatibility with hybrid hardware to protect prior investments. The upshot for AI/ML teams: prioritize resilient, scalable storage architectures so GPUs aren’t idle, maintenance doesn’t interrupt training/production, and large-scale AI workloads remain cost‑effective and operationally practical.