Neoclouds force a rethink of data storage infrastructure

The rapid rise of artificial intelligence has triggered a fundamental rethink of how cloud infrastructure is designed and deployed with a new class of specialist providers, known as neoclouds, challenging the dominance of traditional hyperscale platforms.

As demand for GPU-intensive computing accelerates, industry attention is shifting from compute alone to the often-overlooked constraint shaping AI performance at scale: data storage and movement.

Neoclouds emerge to meet AI’s infrastructure demands

AI workloads, particularly model training and inference, rely on dense GPU clusters capable of massive parallel processing. These environments require infrastructure that goes beyond the original design parameters of general-purpose cloud platforms, particularly in areas such as low-latency networking, high-throughput data transfer, and remote direct memory access (RDMA) optimisation.

This gap has given rise to neocloud providers such as Firmus, OneQode, Sharon AI, CoreWeave, Crusoe, and Nebius, which are building infrastructure specifically optimised for AI workloads. Their model centres on GPU-dense architectures, high-performance networking and consumption-based pricing designed to better align costs with usage intensity.

Market momentum behind these providers has been strong. The neocloud segment is forecast to grow rapidly during the coming decade, underpinned by surging demand for AI training and inference capacity. But as these platforms scale, attention is increasingly shifting to the supporting data infrastructure required to keep expensive GPU resources fully utilised.

Storage emerges as the critical bottleneck in AI performance

While compute capacity is often viewed as the primary constraint in AI systems, storage performance is becoming equally decisive. In practice, GPU utilisation is frequently limited not by processing power, but by the ability to move and deliver data efficiently across distributed systems.

AI workloads depend on a continuous pipeline of data movement. Large-scale ingestion requires high-capacity object storage, while data preparation processes transform information for model consumption.

Training and fine-tuning workloads relies on high-performance file and block storage capable of feeding GPUs at scale, while inference outputs and model artefacts must be retained in durable object storage for governance and compliance.

In many neocloud environments, these requirements are currently met through a patchwork of open-source tools and vendor-specific point solutions. While functional, this fragmented architecture introduces operational complexity, creates data movement bottlenecks and increases governance overheads.

The emerging alternative is a unified storage architecture that integrates block, file and object storage under a single control plane. This approach reduces friction across the AI lifecycle, improves data locality, simplifies governance and helps ensure GPU resources remain continuously fed with data.

GPU-as-a-Service reshapes infrastructure economics

The rise of neoclouds is also being driven by a structural shift in how AI infrastructure is consumed. GPUs are capital-intensive assets, placing them beyond the reach of most organisations if purchased outright. As a result, GPU-as-a-Service models are becoming the dominant procurement approach, enabling organisations to access high-performance compute on a pay-as-you-go basis.

This shift is changing how storage is evaluated. Traditional metrics such as terabytes stored or cost per gigabyte are becoming less relevant than whether storage systems can sustain continuous GPU utilisation. In this new model, idle compute capacity represents direct financial waste, making storage performance a core determinant of economic efficiency.

For storage providers, this evolution introduces a new requirement. Systems must not only deliver high performance but also provide granular visibility into data access patterns and workload behaviour.

Compliance, sovereignty and security

As neocloud platforms increasingly target regulated industries, such as financial services, healthcare and government, the bar for compliance has risen significantly. Customers now expect enterprise-grade security and governance features as standard, including zero-trust architectures, immutable snapshots and air-gapped backups. At the same time, governments are placing greater emphasis on data sovereignty, with a growing share introducing requirements for localised data storage, geo-fencing, and region-aware failover.

For neoclouds, compliance can no longer be treated as a customer-managed responsibility. Instead, it must be embedded directly into the infrastructure stack, integrated across compute, networking, and storage layers, and fully auditable end-to-end.

Storage becomes part of the AI execution fabric

The convergence of performance, compliance and economic pressures is redefining the role of storage in AI infrastructure. Once treated as a passive layer that simply held data, storage is now emerging as an active component of the AI execution fabric.

In GPU-dense environments, data is constantly being consumed and generated at high velocity. Any latency in storage systems directly impacts GPU efficiency and therefore the economic return on infrastructure investment. As a result, storage must now operate as a real-time participant in AI workflows rather than a background service. As the neocloud sector continues to mature, the competitive battleground is shifting. It is no longer enough to simply provide access to GPUs. The winners will be those that can ensure those GPUs are fully utilised, continuously fed with data and supported by storage systems capable of operating at AI-scale performance requirements.

Top image credit: iStock.com/primeimages