Every conversation about AI seems to begin with GPUs. Organizations compete to deploy the latest accelerators and increase compute capacity in pursuit of faster model training and better outcomes. GPUs are expensive, increasingly scarce, and fundamental to modern AI. But for most enterprises, they’re only one part of the equation.
Unlike AI-native companies building purpose-designed infrastructure, enterprises begin with years or even decades of existing data spread across Windows, Linux, macOS, NAS appliances, cloud storage, and departmental file servers. Before a model can learn anything, all that data must be located, centralized, and made available to the AI infrastructure fast enough to keep those GPUs busy.
Why AI training is fundamentally a file problem
Training an AI model is, at its core, an exercise in reading and writing files. Training datasets can consist of billions of files accumulated over many years, while long-running training jobs continuously read that data and periodically write checkpoints to preserve progress. If a training run is interrupted, those checkpoints allow the model to resume rather than starting from scratch.
The faster data can be delivered to GPUs and the faster checkpoints can be written, the more productive those expensive compute resources become. However, if storage cannot keep pace, GPUs spend time waiting rather than processing. This is why the discussion around AI infrastructure cannot stop at compute. GPUs only create value when they are consistently fed with data, and that responsibility sits with the storage, networking, and file service layers underneath them.
The enterprise challenge
Many discussions about AI infrastructure assume organizations are building entirely new environments designed specifically for AI workloads. This isn’t necessarily the case. Many organizations introduce AI into environments that have evolved over years to support line-of-business applications, engineering workloads, finance systems, collaboration platforms, and countless other services.
The underlying storage infrastructure was designed for predictable enterprise workloads that don’t involve hundreds of GPUs simultaneously reading petabytes of unstructured data. What was once considered reliable, background infrastructure is now being asked to support some of the most demanding workloads organizations have ever deployed.
Why the file service layer matters
For decades, remote file services have underpinned enterprise IT. Linux environments have relied on NFS for more than forty years while Windows and macOS environments have depended on SMB. Both remain essential because enterprise AI rarely operates within a single operating system.
Engineering teams may generate data on Linux, finance departments work primarily on Windows, creative teams rely on macOS, and historical information may exist across legacy platforms or the cloud. Bringing these datasets together for AI training requires file services capable of supporting multiple operating systems without creating unnecessary duplication or operational complexity.
Legacy approaches expose new compromises
Today’s file service architectures generally fall into three categories. The first is proprietary storage platforms. These systems are often mature, well supported, and capable of delivering strong performance. However, they frequently depend on specialized hardware, proprietary infrastructure, and flash storage that continue to increase in cost as demand for AI infrastructure grows. For example, Pure Storage (recently rebranded Everpure) disclosed in an open letter to customers that its average prices had risen roughly 70% since the start of 2026, driven by component costs that had surged 300–900% since mid-2025.
The second is legacy kernel-based file services remain familiar to many infrastructure teams. Because they operate directly within the operating system kernel, they introduce reliability and security issues while also making safe container deployment impossible.
The third is user-space alternatives. These improve isolation and support contemporary deployment models, including containers, but often introduce compromises around performance, scalability, and high availability.
Both kernel and user-mode open-source options also lack enterprise support and SLA options. Free software that isn’t working is extremely expensive.
None of these approaches is inherently wrong. The challenge is that AI workloads expose their strengths and weaknesses far more quickly than traditional enterprise applications ever did.
Heterogeneous environments are the new normal
Organizations increasingly need to train AI using information distributed across multiple operating systems, departments, and storage platforms. Rather than replacing these environments, most enterprises need infrastructure capable of bringing them together. That requires file services that provide seamless multi-protocol access to the same datasets while maintaining consistent performance, reliability, and operational simplicity. This means shared security and locking mechanisms to protect the data from corruption and leakage.
Infrastructure teams are no longer evaluating file services purely on compatibility. They increasingly care about deployment flexibility, horizontal scalability, support for containers, operational resilience, and the ability to integrate into modern orchestration environments without introducing unnecessary risk.
AI is changing how organizations evaluate infrastructure
One of the more interesting consequences of enterprise AI is that it is forcing organizations to revisit infrastructure decisions that may have remained unchanged for years. File services have traditionally been viewed as plumbing, i.e., essential but rarely strategic.
However, AI changes that perspective. Infrastructure teams are beginning to evaluate file services in the same way they evaluate networking or compute platforms. They want architectures that maximize throughput, minimize latency, scale efficiently, reduce operational complexity, and support the increasingly heterogeneous environments that define enterprise IT today.
Rather than asking whether a storage platform supports NFS or SMB, organizations are asking whether their file service architecture is capable of supporting the next generation of enterprise workloads.Â
Looking beyond compute
The AI industry will continue to focus on larger models, faster accelerators, and increasingly powerful GPUs. But for enterprises what will make the difference is moving data from storage to compute more efficiently.
An effective file service for enterprise AI must deliver extreme throughput, low latency, low resource consumption, high availability, and operate safely in user space. It must also provide seamless multi-protocol access that allows the same data to be shared across Linux, Windows, macOS, and other enterprise platforms without compromise.
While organizations will continue to invest in AI, ROI will depend on keeping expensive compute resources productive. The organizations that solve the data movement challenge will be the ones that realize the greatest return on their AI investment.
