Scality has published new research arguing that production AI is increasingly constrained by data infrastructure, with object storage now serving as a core component in most private AI environments.
The study, conducted independently by Freeform Dynamics and based on a survey of 504 senior IT and data professionals in medium to large enterprises, found that 91% of organisations running private AI in production report meaningful use of object storage. Of those, 44% said they use it extensively and 47% said they use it “quite a bit”, putting object storage slightly ahead of file-based approaches and well above block-based storage in these environments.
The report also suggests infrastructure constraints are broadening beyond compute. While GPUs often dominate the AI infrastructure discussion, 57% of respondents said storage performance is a priority for avoiding AI bottlenecks, compared with 54% citing compute or GPU availability and 52% pointing to network bandwidth. Elsewhere, 81% said infrastructure they control is critical to success in private AI, while 40% cited metadata handling at scale as a bottleneck risk and 38% pointed to mixed workload handling challenges.
Tony Lock, Director of Engagement and Distinguished Analyst at Freeform Dynamics, said: “Most industry discussion frames AI infrastructure as primarily a compute challenge. This research makes clear that enterprises running private AI in production are dealing with a broader systems reality. Many see a need for simple, scalable architectures that keep data close, support multiple AI genres, and balance performance with governance and cyber resilience across the full pipeline.”
The findings add weight to an increasingly common enterprise view that production AI is as much a data pipeline problem as a model problem. As more workloads shift from experimentation to inference-heavy operational use, storage design, lifecycle control, and governance are moving closer to the centre of AI infrastructure planning.
The full report is available on Scality’s AI report page.
