Model Serving Capacity Changes Storage Ratios Entirely
Inference and serving infrastructure built in India carries storage per compute unit at ratios enterprise workloads never produced, since model weights, embeddings, and retrieval corpora all live on fast media that must be read constantly. That pulls demand toward scale-out architecture rather than toward arrays, and toward capacity purchased by the rack rather than by the system. Software-defined and scale-out storage grows at 22.8% on this. Around 58% of installed petabytes now sits with platform operators, and the proportion continues to rise with each facility that opens. Arrays address almost none of this.
Market Impact: Drives 23% of capacity








