Facebook
Britain's News Portal
Around The Clock
BREAKING
Loading latest headlines…

AI data hunger exposes storage bottleneck as GPUs sit idle

UK businesses investing in AI are discovering that their expensive GPU clusters are often idle, waiting for data that is fragmented across legacy storage systems. Experts warn that data placement, not capacity, is now the critical constraint holding back AI projects.

  • GPU utilisation figures can be misleading as accelerators spend significant time waiting for data access across distributed storage systems.
  • Data fragmentation across departments, sites and clouds is a major barrier, with 57% of organisations believing their data is not AI-ready.
  • NVMe drives inside GPU servers are often left as isolated scratch space, representing 'stranded' storage capacity that could be pooled for faster AI training.
  • The Hammerspace Data Platform aims to create a global namespace that lets compute find data where it already lives, rather than moving data to the GPUs.
  • UK regulators including the ICO are watching data infrastructure developments closely as AI adoption accelerates under the EU AI Act framework.

The promise of artificial intelligence in UK industry is running into an unglamorous but costly reality: data is scattered, and the expensive GPUs bought to train models are spending too much time waiting for files. According to experts at data infrastructure firm Hammerspace, a GPU dashboard showing 70% utilisation can be misleading, as a significant portion of that time is spent idle while accelerators wait for data stored three network hops away on a legacy NAS box.

This bottleneck has shifted the conversation from storage capacity to data placement. 'The data is in disparate groups and disparate orgs and disparate silos within a company,' said Jonathan Flynn, director of applied systems at Hammerspace. 'Having the data in a curated data set for you just to go train is rare. It has to be collected. It has to be moved around from system to system.' Gartner estimates that 57% of organisations believe their data is not AI-ready, while two-thirds of executives say no one in their organisation fully understands all the data they have collected or how to access it.

The problem is compounded by vendor lock-in. Mike Bloom, who covers AR architecture at Hammerspace, warned that many vendors push companies to rip out legacy storage entirely. 'They'll go to a vendor that will promise them that if they sweep the floor and throw out all of their legacy storage arrays, their brand will solve the problem,' he said. He argued that this approach ignores the reality that valuable data sets already reside on those older systems, and that throwing them out is 'like throwing the baby out with the bath water.'

Another overlooked resource is the NVMe storage already installed inside GPU servers. Modern HGX and DGX boxes ship with eight to sixteen NVMe drives, but orchestration layers typically treat this as local scratch space for a single server. Hammerspace calls this 'stranded' capacity, noting it can amount to hundreds of terabytes per server, with two-petabyte GPU servers on the roadmap. By pulling that storage into a shared namespace, organisations could create what Hammerspace calls Tier 0, using hardware they already own and a network already deployed.

For UK businesses, the implications are significant. The Information Commissioner's Office (ICO) continues to emphasise data governance as AI regulation evolves, and the EU AI Act's extraterritorial scope affects British firms operating in Europe. The cost of idle GPU time, combined with inefficient data pipelines, directly impacts return on investment for AI projects across sectors from finance to healthcare. As Bloom put it, 'You can't orchestrate what you can't see' — and for many UK enterprises, the first step may be simply getting a clear view of where their data actually lives.

Why this matters: UK businesses are investing heavily in AI hardware, but inefficient data access means many are not getting the performance they paid for. Fixing data fragmentation could unlock faster AI training without additional hardware spend.

What this means for you: If your company uses AI, your expensive GPU hardware may be underperforming because data is scattered across old storage systems — fixing data access could save money and speed up results.

Related Articles

Get the news that matters.

Join thousands of readers getting the best of British news straight to their inbox.