Storage may receive less attention than power, cooling and compute, but Simon Ninan, Senior Vice President, Business Strategy at Hitachi Vantara, explains why it could quietly constrain the next generation of AI infrastructure.

The AI data centre conversation has been dominated by compute power. GPU capacity, accelerator roadmaps and the race to power ever-larger models have absorbed much of the industry’s attention and capital expenditure in recent months. Yet, as hyperscalers, neo-clouds, colocation operators and enterprises scale their AI deployments, a more foundational challenge is coming into focus: data storage.

The gap is not just one of raw capacity. The storage industry is also confronting a global shortage of NAND flash memory chips, which is increasing the prices and lead times of storage devices needed to meet enterprise demand. Even as the industry works to catch up on capacity over the next couple of years, it is being forced to confront a hard reality: the storage estates that have served enterprises reliably for a decade were not designed for the data movement patterns, latency demands or throughput requirements that AI workloads impose. Filling that gap, alongside addressing the capacity challenge, requires a different approach, not simply more of what already exists.

Where legacy architectures fall short

Traditional enterprise storage was optimised for transactional workloads, consistent input/output operations per second (IOPS), high availability and predictable access to structured data. AI training and inference pipelines operate on different assumptions.

A large language model fine-tuning run may need to ingest hundreds of terabytes of unstructured, multimodal data in parallel at sustained throughput, with minimal latency between the storage layer and the GPU cluster. Inference, on the other hand, shifts the focus towards low-latency access to geographically distributed models and feature data, including at the edge. Storage must deliver fast model and context reads while scaling horizontally with traffic.

Furthermore, data lifecycle assumptions change with AI pipelines, which expect the continuous ingestion, transformation and retention of growing datasets. This requires storage to scale capacity and bandwidth rapidly across heterogeneous workloads. In all of this, the quality of the data in the pipeline can have a significant effect on business value and return on investment.

Legacy storage area network (SAN) and network-attached storage (NAS) architectures can introduce bottlenecks at the data ingestion stage, leaving GPU resources sitting idle. When a single AI accelerator can cost tens of thousands of pounds, the economics of delays caused by storage or data pipelines become stark. Idle GPUs are not just a performance problem; they are a budget problem.

The market has started to recognise this. In a recent study of IT decision-makers across 15 countries, 67% of UK respondents identified high-quality data as a critical success factor for AI projects, up 26 percentage points in a single year and well above the global average of 48%. A further 41% cited robust infrastructure as a success factor, up 13 percentage points year-on-year. The effectiveness of AI depends heavily on the data pipeline feeding it, and that pipeline begins with storage.

The data movement problem

Data movement is an underappreciated cost centre within AI infrastructure. Before a model can train or perform inference, raw data must be ingested, cleaned, transformed and staged across multiple pipeline stages. Each movement consumes bandwidth, introduces latency and draws power.

EMEA operators face a particular challenge. Hybrid, multi-domain environments are becoming the norm, with data distributed across on-premises infrastructure, colocation facilities and the public cloud. In the UK specifically, 85% of organisations report data sovereignty requirements that influence where AI workloads are deployed, significantly above the global average of 68%. As a consequence, 73% now host sensitive data in private cloud environments. This places data sovereignty considerations for on-premises and colocation storage at the centre of emerging AI strategies, rather than at the periphery.

Performance, power and the scalability trade-off

Storage choices directly affect power and cooling budgets in ways that are often overlooked. High-density non-volatile memory express (NVMe) all-flash arrays deliver the performance AI workloads demand, but consume more power per rack unit than spinning-disk alternatives. For operators facing constrained power budgets in Western Europe, the energy profile of the storage tier matters.

Modern flash platforms are substantially more power-efficient per terabyte than earlier generations, while data reduction technologies such as deduplication and compression can ease both space and power constraints without sacrificing performance.

However, implementation and system design can have a significant bearing on the outcome. Storage technologies vary in how effectively they optimise capacity, data movement and device power consumption according to the workloads being handled at any given time. At scale, these differences can have a material effect on annual operating costs.

Practical considerations for enterprise and EMEA operators

When evaluating storage in the context of AI, three questions are worth addressing alongside compute and cooling planning.

First, what are the characteristics of the business use case, and what do they mean for the throughput and latency requirements of the workloads in scope? Training demands sustained throughput, whereas inference demands low latency; these profiles may require different storage tiers to be deployed together. Inference demands may also vary significantly depending on the industry and use case.

Second, how is the data pipeline managed, and how are both quality and security guaranteed? Infrastructure investment without data governance discipline consistently underdelivers.

Third, what is the operational model? Beyond the technical requirements, how will sovereignty be delivered through localised operations and ecosystems? AIOps capabilities and greater automation of routine operations may help to reduce the management burden as workload demands evolve.

Storage as a strategic decision

Storage architecture determines how effectively GPU investment is utilised, how efficiently data moves through AI pipelines, how much power the infrastructure consumes and how readily the environment can scale. For enterprise and EMEA operators, addressing the storage gap as part of an integrated infrastructure strategy, rather than as an afterthought to compute procurement, can help ensure that AI infrastructure delivers on its promise rather than quietly constraining it.