Why the Best AI Clouds Don't Run on Flash Alone (www.backblaze.com)

Neoclouds, which rely heavily on GPUs for AI workloads, are facing challenges with conventional storage solutions due to the significant data throughput demands of AI training. While flash storage is ideal for the active training phase—delivering the necessary speed and low latency—most other stages of the AI lifecycle require less expensive, high-capacity storage options. This mismatch often leads customers to hyperscaler storage solutions, which charge high fees for data retrieval, creating a "data gravity" effect that ultimately consolidates client relationships with hyperscalers instead of the neocloud provider. To effectively compete, neoclouds should adopt a tiered storage architecture that optimally allocates storage resources based on the specific needs of each phase in the AI lifecycle. The recent introduction of B2 Neo allows neoclouds to incorporate a white-label solution that provides the necessary storage without needing to develop it from scratch. This approach ensures that the full range of storage requirements—from data ingestion to checkpointing and long-term storage—remains in-house, enabling neoclouds to not only enhance service reliability but also retain client relationships by avoiding dependency on hyperscalers.