Enterprise AI has managed to make GPUs glamorous. Storage hasn’t had quite the same makeover, but that may be a problem for companies spending heavily on AI infrastructure without asking enough questions about the data those systems will actually use.
The assumption that AI infrastructure begins with compute is understandable. GPUs are expensive, supply has been constrained and every major infrastructure vendor has spent the past few years talking about accelerated computing. Yet an AI system that can’t get reliable access to the right data quickly enough is still an expensive system waiting around for work. We recently looked at precisely that problem in the context of GPUs sitting idle despite continued demand for AI compute. Storage is another part of the same infrastructure problem.
“Many organisations think their AI journey starts with models or GPUs,” says Marc Layne, sales director at Dell Technologies South Africa. “In reality, it starts with understanding whether your infrastructure is ready to support AI at scale.”
It’s a useful argument, even allowing for the fact that Layne works for one of the world’s largest infrastructure vendors. Dell itself is making storage a significant part of its AI pitch. In April, the company identified AI data pipelines and cyber resilience among the forces changing enterprise storage, while its May infrastructure announcements tied new storage, servers and data-protection products explicitly to AI workloads.
The more interesting question for businesses isn’t whether they need newer storage because AI has arrived. It’s whether AI is exposing infrastructure decisions they’ve been able to get away with until now.
AI has very little patience for bad data architecture
A company can accumulate messy infrastructure for years without anything visibly breaking. Data ends up spread across different systems. Older applications remain because replacing them is difficult. Departments build their own repositories. Capacity gets added when necessary. None of this is ideal, but an organisation can continue operating around it.
AI changes the tolerance for that mess because data stops being something applications merely save and retrieve. It increasingly becomes an input into systems expected to search, analyse, infer and generate answers from enormous amounts of information.
Layne points to data silos, limited scalability and complex storage environments as weaknesses that become harder to ignore once AI workloads arrive. The argument isn’t that storage magically determines whether an AI model is good. Data quality, governance, model selection, skills and the problem being solved all still matter. Infrastructure can, however, constrain what organisations are able to do with the data they already possess.
That’s particularly relevant as more enterprises consider keeping some AI workloads and sensitive information within infrastructure they control. Reframed has previously looked at Huawei’s pitch for AI-ready data infrastructure in South Africa, and the notable part wasn’t Huawei making the argument. Every large enterprise infrastructure company has some version of it. The repetition tells us something about where vendors believe the bottlenecks are moving.
South Africa’s own AI ambitions make that question harder to dismiss. The country’s draft national AI policy includes proposed investment in supercomputing and digital infrastructure while simultaneously raising concerns about sensitive South African data being dependent on foreign infrastructure. Those are partly questions of national policy and sovereignty, but businesses encounter smaller versions of the same tension: where does their data live, how easily can it move, who controls it and what happens when something fails?
Storage now has to survive the attack too
The other change is cybersecurity. Storage used to be discussed primarily in terms of capacity, availability and performance. Ransomware has made recovery just as important.
“Cyber resilience has become one of the defining requirements of modern infrastructure,” says Layne. “Organisations want confidence that their most valuable asset, their data, remains protected and recoverable regardless of what happens.”
There is some vendor language buried in that statement, but the underlying shift is difficult to argue with. A company’s cybersecurity strategy can successfully detect an intrusion and still leave the business in serious trouble if critical data has been encrypted, corrupted or destroyed and recovery is slow or unreliable.
That changes the economics of a storage decision. The cheapest way to accommodate another few hundred terabytes of information may not be the cheapest option once downtime, recovery requirements, data protection and the cost of moving between systems are considered.
It also complicates the familiar infrastructure-refresh cycle. Layne argues that organisations should be able to grow capacity and performance without repeatedly replacing large parts of their environment. “No business plans to remain exactly the same size over the next five years,” he says. “Technology investments should reflect that reality.”
There is an obvious commercial incentive behind that argument too. Infrastructure vendors would much rather sell organisations platforms and long-term relationships than boxes of storage every few years. Businesses should interrogate claims about lower disruption, guaranteed data reduction, simplified management and investment protection with the same scepticism they’d apply to any other technology purchase.
But the buying question itself has genuinely changed.
The person choosing the storage may have changed too
When storage existed mainly to give applications somewhere dependable to put information, it made sense for capacity and performance requirements to remain largely inside IT. Once the same systems affect AI projects, ransomware recovery, cloud strategy, regulatory obligations and how quickly a business can put its own data to work, the consequences travel much further through the organisation.
That doesn’t suddenly mean chief executives should start comparing storage architectures. It means infrastructure decisions need to begin with business questions that are much harder than asking how many terabytes the company expects to need.
What data will the organisation need to use differently in three years? How quickly would the business have to recover it after an attack? Which workloads can live in public cloud infrastructure and which cannot? What happens to cost as AI creates new copies, embeddings and derivatives of existing information? How difficult will it be to move that data if the technology strategy changes?
These aren’t particularly exciting questions compared with the possibilities being promised around generative and agentic AI. They may determine how many of those possibilities survive contact with the company’s actual infrastructure.
The AI spending cycle has encouraged businesses to think about what their next generation of computing can do. It may also force them to reckon with years of infrastructure decisions made when the data underneath that computing attracted far less attention.