Despite billions poured into generative AI, most companies still aren’t seeing real returns. Analysts say the problem isn’t the algorithms — it’s the plumbing underneath. Inadequate data infrastructure, poor scalability, and storage bottlenecks are preventing enterprise AI from moving beyond flashy pilot projects into production systems that actually generate profit.

A recent MIT study found that U.S. companies have collectively invested between $30 and $40 billion in generative AI, yet 95% report no measurable ROI. Only 5% of organizations have managed to deploy GenAI tools successfully at scale.
According to researchers, the issue lies not in human talent or corporate enthusiasm but in the technology’s limitations — especially its inability to learn continuously, adapt dynamically, and integrate into mission-critical workflows. (The study has not been published directly by MIT but was widely reported by other outlets.)

The research also revealed a sharp divide: a tiny fraction of companies are extracting millions in value from AI, while the vast majority see no measurable impact on their bottom line. Experts point to four recurring pitfalls that explain why.

Why GenAI Projects Keep Failing

Most corporate AI projects collapse before they leave the pilot stage. Out of more than 300 public implementations, only one in twenty reaches production with measurable results.
The cause? Poor enterprise integration. Many AI tools cannot learn from feedback or plug seamlessly into daily operations, leaving them stuck in experimental mode.

Another major issue is the rise of a “shadow AI economy,” where employees independently use tools like ChatGPT without alignment to company objectives. At the same time, AI budgets are often misallocated — roughly half the funding goes to sales and marketing, even though back-office automation typically delivers higher, more consistent returns.

According to Björn Kolbeck, CEO and co-founder of Quobyte, the lack of ROI cuts across every organizational layer. “Some products are pushed by executives or contain half-baked ‘AI features,’ while others simply can’t scale due to weak infrastructure,” he told TechNewsWorld. “All suffer if you can’t feed GPUs at scale — in terms of memory, adaptability, and integration.”

The Hidden Cost of Weak Storage

Kolbeck argues that storage is the most overlooked component of the AI stack. Companies spend billions on GPUs and models but neglect the data backbone required to keep them running efficiently. This oversight leads to three recurring pain points: data silos, performance lag, and uptime instability.

“Storage systems must scale and provide unified access to create an AI data lake — a single, centralized, and efficient storage foundation for the entire company,” Kolbeck said.
When data is fragmented across silos, scientists can’t access the full picture. And when storage can’t keep up with GPU throughput, expensive hardware sits idle while projects stall.

Similarly, many traditional high-performance computing (HPC) storage systems aren’t built for continuous availability, resulting in delays that ripple through AI workflows.

Why Traditional Storage Breaks at Scale

The MIT analysis found that successful AI deployments rely on fault-tolerant, scalable storage. But most enterprises still use legacy “scale-up” systems designed for smaller, static workloads.

“Early AI projects may run fine, but as soon as they scale up and add GPUs, traditional storage arrays tip over — and that’s when mission-critical workflows grind to a halt,” Kolbeck warned.

The solution, he says, is scale-out architecture — a design that grows horizontally with each new workload. Quobyte’s own platform transforms commodity servers into a parallel, high-performance storage system that can grow from a few nodes to entire data centers.
“In every domain — from HPC to CPUs to the cloud — scale-out has won,” Kolbeck noted. “AI is no different. If you can’t scale storage alongside GPUs, you’ll hit a wall.”

Managing the Data Chaos

AI development requires handling both massive datasets and countless small files, often accessed by many GPUs at once. That combination pushes even advanced systems to their limits.

“Developing and training AI models is still a highly experimental process,” Kolbeck explained. “The infrastructure — especially storage — must adapt quickly as data scientists test new ideas.”

Real-time analytics are essential. Storage administrators need visibility into how training pipelines affect performance so they can tune or expand the system intelligently. Quobyte’s policy-based management engine addresses this by allowing administrators to reconfigure storage, reassign data, or change access policies instantly — without downtime.

Outdated Tech Can’t Support Modern AI

Kolbeck points out that much of today’s enterprise storage still relies on decades-old protocols like NFS, first created by Sun Microsystems in 1984. “Believing that this recycled technology will suddenly power successful AI is wishful thinking,” he said.

He contrasts Yahoo, which relied on NFS-based appliances, with Google, which built its infrastructure on distributed software-defined storage running on inexpensive servers — a philosophy that still powers hyperscalers today.

To truly unlock AI’s potential, Kolbeck argues, companies must think like hyperscalers. Quobyte’s approach applies distributed systems algorithms to deliver reliability and performance on commodity hardware, scaling seamlessly from a handful of servers to entire data centers.

“AI success starts where most people never look — the data layer,” he said. “If your storage can’t scale, your AI never will.”

Source