Feeding the Beast: Why Disaggregated Storage Is Essential for AI Infrastructure and Sustainability

AI adoption across sectors is at an all-time high and continues to accelerate at an unprecedented pace. The global AI infrastructure market is on track for unprecedented growth, poised to reach USD758 billion in spending by 2029, according to IDC’s Worldwide Quarterly Artificial Intelligence Infrastructure Tracker.

In Hong Kong, a similar momentum is taking shape as enterprises increasingly embed AI tools into their daily operations. According to the Hong Kong Productivity Council’s AI Readiness in Workplace Survey 2025, 88% of employees in local enterprises already use AI tools in their daily work, and 92% of companies plan to formally integrate AI into workflows. While much attention is given to compute performance and GPU advancements, one of the most critical and often overlooked components of scalable AI infrastructure is storage.  

AI is only as valuable as the data it can process. Organizations increasingly work with massive volumes of unstructured data across AI pipelines. Yet from ingestion and preparation to training and inference, feeding data-hungry GPUs efficiently remains a serious bottleneck. With Hong Kong enterprises rapidly embedding AI into customer service, data analysis and marketing, demand for faster data ingestion, preparation and high-throughput storage is rising sharply. Storage systems must now deliver extreme throughput, low latency, and operational agility, all while aligning with aggressive energy-efficiency and sustainability goals.

Traditional, tightly coupled architectures that bundle storage with compute fall short in this environment. In server- or HCI-based deployments, over-subscription becomes more of an issue at large scale, leading to overprovisioning, stranded capacity, unnecessary hardware purchases, and increased power draw. These architectures not only hinder performance but also complicate sustainability initiatives and inflate the total cost of ownership (TCO). CBRE’s research further shows that Hong Kong faces tightening data-centre space and rising power and colocation costs, making any structural inefficiency in infrastructure deployments far less affordable for enterprises operating in the city. space and rising power and colocation costs, making any structural inefficiency in infrastructure deployments far less affordable for enterprises operating in the city.

Disaggregated storage offers a modern solution to these challenges. By decoupling compute from storage, infrastructure teams gain the flexibility to scale resources independently, enabling far greater efficiency, performance and sustainability. With disaggregated architectures, organizations can build shared storage pools accessible across multiple compute nodes via high-speed, low-latency interconnects like NVMe-over-Fabrics (NVMe-oF™) using RDMA (Remote Direct Memory Access) over converged Ethernet (RoCE), TCP/IP (Transmission Control Protocol/Internet Protocol) or fibre channel. This shift not only improves infrastructure performance and flexibility to scale but also directly supports sustainability initiatives by eliminating waste and maximizing resource utilization.

Here are four reasons disaggregated storage is essential for scaling AI and doing it sustainably:

1. Independent scaling reduces overhead and environmental impact

According to Nature.com, generative e-waste streams could reach a total accumulation of 1.2–5.0 million tons during 2020–2030. Disaggregated storage enables IT teams to scale storage and compute independently, avoiding costly, inefficient overprovisioning. Whether you're training a new model that requires additional GPU power or storing vast archives of training data, disaggregated infrastructure lets you add only what’s needed — no more, no less. This modular approach helps reduce hardware lifecycle costs and e-waste, as well as infrastructure bloat and underutilized capacity. It also aligns well with circular economy principles, where compute and storage components can be refreshed independently, contributing to a more circular, low-waste model of IT infrastructure.

2. High throughput for performance-hungry AI workloads

Modern AI applications, from LLMs to AI-enhanced analytics, demand sustained, high-speed access to large datasets. Disaggregated architectures meet this demand by enabling direct, low-latency communication between GPU memory and storage, bypassing CPU bottlenecks and increasing end-to-end data pipeline efficiency.

Western Digital’s OpenFlex™ Data24 platform, paired with NVIDIA GPUDirect® Storage (GDS), is one example of this model in action. In benchmark testing using GDSIO, the system achieved a 460% increase in throughput with GDS enabled, compared to traditional configurations. By improving GPU efficiency, these high-throughput systems help avoid wasted compute cycles, reduce idle time, accelerate model completion, and lower overall energy use per workload. And better throughput isn’t just about speed; It also means fewer total resources are consumed to complete a task, resulting in lower energy per unit of computation. Given that power demand from generative AI will increase at an average annual rate of 70% through 2027, largely driven by data center growth, optimizing infrastructure will be critical.

3. Smarter resource utilization improves efficiency and TCO

Disaggregated storage unlocks dynamic provisioning, allowing IT to direct storage capacity where it’s needed most without rearchitecting the entire environment. This flexibility is especially useful for AI pipelines with variable workloads and unpredictable demands.

By limiting underutilized nodes and consolidating storage for higher density, organizations can reduce power consumption and cooling overhead per terabyte of data. This, in turn, extends asset lifecycles and delays costly infrastructure refreshes, improving both environmental impact and return on investment

4. Right-sized infrastructure supports dynamic AI pipelines

AI workloads are not one-size-fits-all. Training jobs may require petabytes of archived data and massive parallel access, while inference might rely on low-latency, high-frequency transactions. Disaggregated storage supports both ends of this spectrum and everything in between.

Whether deploying cold data lakes, hot-edge inference platforms, or high-throughput training farms, disaggregated systems allow infrastructure teams to optimize rack space, thermal output, and performance without locking themselves into rigid configurations that may become obsolete within a year.

Rethinking storage for the AI era

As AI continues to reshape industries, it’s clear that storage infrastructure must evolve to keep pace. Disaggregated storage offers a practical, proven path forward, providing the agility, performance and efficiency required to keep pace with AI’s explosive growth.

More importantly, it gives CIOs and infrastructure leaders a framework to scale without compromise: on budget, on energy use, and on long-term sustainability goals.

As you plan your next-generation AI infrastructure, consider not just how much data you’ll generate, but how efficiently you can move, store, and extract value from it. Disaggregated storage isn’t just a technology choice. It’s an architectural shift toward a smarter, more sustainable future. As Hong Kong positions AI as a pillar industry in the latest Policy Address, and as enterprises advance from experimentation to a fully integrated, cost-efficient, sustainable data infrastructure at scale will become a prerequisite for competitive advantage.

The views and opinions expressed in this article are those of the author and do not necessarily reflect those of CDOTrends. Image credit: iStockphoto/rogertrentham