Source: Techrevolt News, Oct 19, 2026 – When AI factories scale, storage architecture becomes a critical bottleneck. Matt Swalley says AI clouds need intelligent, software-defined storage that can balance high-performance flash, lower-cost capacity, multi-tenant isolation and real-time observability.

As artificial intelligence scales and AI factories multiply, compute capacity has dominated headline investments. Thousands of high-performance GPUs are being deployed across data centers globally, with vast amounts of capital.

However, a growing operational reality comes to the surface – GPUs cannot perform efficiently if the underlying data layer fails to keep pace.

Without continuous, ultra-fast data delivery, high-cost GPUs either spend valuable processing cycles sitting idle during training checkpoints or struggle to maintain context during long inference tasks. For data centers and emerging neocloud providers, these idle cycles directly impair unit economics, inflate total cost of ownership (TCO), and restrict operational scalability.

Matt Swalley, Senior Director at VDURA, works closely with neoclouds, enterprise data centers, and AI infrastructure operators to resolve these bottlenecks.

At AI Everything in Abu Dhabi, Swalley met with Tech Revolt to discuss why storage architecture is becoming a central economic driver for GPU deployments and how software-defined parallel file systems optimize multi-tenant environments for AI clouds.

1: GPU infrastructure gets most of the attention in AI infrastructure, but VDURA argues storage can be the hidden bottleneck. How does that affect the GPU economy?

Across the AI pipeline, distinct operational challenges exist in both training and inference.

In training, GPUs perform heavy computation and write sequential bursts to data storage during checkpointing. Every time a GPU is waiting for a checkpoint to write, it remains idle.

When a GPU sits idle, the cloud provider or enterprise isn’t generating value. Their core objective is to keep GPUs running at maximum utilization and efficiency.

The economics shift even more as workloads transition toward inference, which will account for roughly 90% of AI workloads by 2030.

In inference, where tokens directly drive revenue, extending local memory into storage without latency spikes is critical. If an agentic AI system or long-context model runs out of local GPU memory, it is forced to clear data and recompute, which costs extra tokens and inflates operational expenses.

To solve this economically, VDURA developed the V12 parallel file system, which combines high flash throughput, reaching up to 75 GB/s per node with intelligent, context-aware tiering.

By holding roughly 90% of hot operational data on flash while moving 90% of total long-term volume onto disk, operators maintain high performance while reducing total cost of ownership by up to 60%.

That capital efficiency feeds directly back into expanding active GPU deployments rather than overspending on storage commodities.

2: V12 is designed for multi-tenant AI environments. As neoclouds increasingly serve many customers on shared infrastructure, what becomes harder: maintaining performance, guaranteeing isolation, or managing the operational complexity behind both?

You have to evaluate the entire ecosystem, from investors and infrastructure landlords to neoclouds selling to enterprise off-takers across specialized industries like defense or foundational AI research.

Each end customer brings vastly different security and performance requirements.

Data isolation is a non-negotiable. V12 addresses multi-tenancy by establishing isolated environments where tenant data cannot cross boundaries, protected by end-to-end encryption from ingest to storage.

From a performance standpoint, we have all-flash memories and a mixed fleet. It can be 100% flash or 10% flash. It depends on the requirements.

Modern GPU nodes require significant data throughput, around 4 GB/s read and 2 GB/s write per GPU, with capacity requirements ranging from 2 TB to 25 TB per GPU depending on modern multimodal models.

Neoclouds need the operational flexibility to offer tiered service levels. That’s where VDURA understands their operational load and offers 100% all-flash tiers down to mixed-fleet options.

And V12 also provides control APIs that allow neocloud operators to manage these multi-tenant environments dynamically using their own orchestration systems, while keeping performance scalable across thousands of nodes.

3: AI workloads are becoming less predictable, with data moving between training, inference, checkpoints, and long-term storage. How will making storage intelligent to respond dynamically help?

Making storage intelligent and observable is central to handling shifting workloads.In the past, storage was largely treated as a static repository. Today, as workloads swing unpredictably between training runs, continuous checkpointing, and agentic inference, storage needs active, embedded intelligence.

We address this with VDURA Sentinel, an observability and management framework integrated within V12 for full visibility.

Sentinel runs intelligent telemetry within the data layer to proactively monitor system health, automatically log operational tickets, and provide visibility into workload behavior.

By intelligently managing how context is preserved and extended from GPU memory down into high-performance parallel storage, the storage layer actively prevents unnecessary re-computation.

The result is consistent throughput, lower token generation costs, and improved response times for end users.

4: There is a huge amount of capital going into AI infrastructure across the Gulf. But are operators thinking enough about the data layer or is the attention still concentrated on acquiring GPUs and compute capacity?

The speed of digital transformation across the Gulf and broader Middle East region is remarkable. It’s one of our top markets.

Initially, the global focus was almost exclusively on compute accumulation, acquiring as many GPUs as possible as quickly as possible.

But now, the approach is more holistic. With the massive surge in price, supply chain dynamics, hardware commodity shifts, and operational power constraints, the conversation is maturing rapidly.

It’s now turning to efficiency, which is great, since it’s a more holistic view of infrastructure design.

Industry leaders are recognizing that reaching leadership in AI and digital transformation requires balancing compute procurement with software-defined data storage that protects unit economics and accelerates execution.

5: As GPUs become increasingly accessible, how does storage architecture and data movement become a real competitive differentiator between AI clouds?

Storage architecture and data movement are rapidly becoming core differentiators for AI clouds. Accumulating compute is only the first step. But running GPUs above an 85% utilization threshold is what determines whether an AI cloud is profitable.

That is central to VDURA. It’s a software-defined architecture which offers speed and accuracy. Governments and operators can now build on their preferred hardware while retaining the flexibility to scale nodes independently as demand grows.

When Storage Unlocks Compute Scale

The race to build global AI capacity is undergoing an operational shift. While GPU procurement remains a headline metric, the underlying data architecture dictates whether that compute operates efficiently or sits idle.

For AI clouds and enterprise data centers, managing high-throughput training bursts and low-latency inference with context requires an intelligent, multi-tenant data layer.

By combining context-aware tiering, open hardware flexibility, and software-defined parallel storage, total cost of ownership is lowered while maximizing GPU utilization.