Your Hard Drives Are Stealing Your AI’s Lunch Money

In 2025, the word “power” was on the lips of most tech leaders. With AI guzzling electricity, it’s easy to see why. But they may have gotten the source totally wrong. Yes, it’s not those snazzy GPUs that’ll kill your AI growth plans; it’ll be your building blocks, particularly storage.

So while everyone obsesses over NVIDIA’s latest accelerators, the real issue lies in spinning hard drives sitting in racks, burning watts like it’s 1999. Matthew Oostveen chief technology officer and vice president for Asia Pacific and Japan at Pure Storage frames it: “In a power-constrained world for data centers, there’s no space for hard drives. Not just in the physical sense that they take up a lot of rackspace, but in energy terms.”

Those “terms” mean that every watt your storage wastes is a watt stolen from the GPUs that actually drive revenue. In other words, we need to see inefficient storage as more than just higher electricity bills; it also means slower AI models, less accurate outputs and competitors eating your lunch while you’re busy cooling down spinning platters.

The hidden tax on AI

The data center power crisis is not some near-future event. We’re already living through it with governments reacting.

For example, Ireland’s EirGrid stopped new data center grid connections until 2028 due to capacity concerns. After a moratorium, Singapore restarted data centre approvals through a pilot that limited new capacity to around 60MW, before shifting to a larger, sustainability‑linked expansion program like the Green Data Centre Roadmap. Northern Virginia, home to the world’s densest concentration of data centers, hit the brakes on new construction due to power grid limitations.

Yet people overlook the fact that storage has a significant impact on total data center power consumption globally. Let’s do the math: in 2022, datacenters consumed roughly 240–340 TWh. Often we see that this power draw comes from servers. But storage adds to the equation too, which is around 5%, according to an International Energy Agency report. That’s significant when you consider that AI workloads that are projected to double data center power demands by end of next year from 2022.

What’s hidden behind those storage energy numbers on spec sheets: the indirect energy load from storage can exceed the direct power draw — especially as data centers shift toward high-density design.

“Higher density equals fewer racks equals less cooling plus less facility overhead,” Oostveen explains. Flash storage systems, like those from Pure, deliver significantly more terabytes per watt than HDD-based arrays, so they generate less heat per terabyte and reduce cooling and rack-space requirements. Traditional HDD-heavy architectures require many more drives and enclosures to reach the same capacity and throughput.

The problem becomes exponential: a petabyte of hard drives requires 50-plus drives. Every additional drive draws idle power, adds cooling requirements, increases airflow obstruction, and requires controllers to manage them. The cooling load scales superlinearly with heat density, which means the more heat-dense your clusters are, the more aggressive (and power-hungry) your airflow needs to be.

Interested in knowing more about architecting your data infrastructure? Join Matthew Oostveen and leading practitioners as they discuss the same article topic online on 15 January 2026. To register or find more information about this webinar, which is part of the Pure Leadership Series, click here. 

Capacity isn’t what you think anymore

As a result, Oostveen sees a fundamental shift among Asia-Pacific hyperscalers: the introduction of Energy Tiered Storage. They’re asking existential questions: How much data can remain in high-energy tiers versus how much must be forced into cold storage to stay within operational constraints?

This introduces a metric that deserves more attention: data gravity per watt. “Storing data is one dimension; moving it is often more energy-intensive,” Oostveen notes. As energy constraints tighten, architectures that optimize data locality and minimize long-haul data movement produce higher ROI.

Pure has already baked energy awareness into Pure1 and Fusion, allowing customers to monitor energy consumption per tier and set automated policies. But the real design implications point toward three focal areas: more in-rack compute for data enrichment, software that minimizes shuffle and exchange in distributed systems, and intelligent placement systems optimizing for energy cost rather than just compute cost.

The agentic AI architecture inversion

AI’s evolution from centralized training to distributed inference and agentic loops changes everything about storage I/O patterns. At least AI training was predictable, sequential, and throughput-heavy. With agentic AI, it becomes chaotic, latency-sensitive, and concurrency-heavy.

“Agents must retrieve embeddings, features, context windows, and state data in milliseconds or microseconds,” Oostveen says. “They request millions of random reads and have explosive growth in IOPS.” When data fragments across NAS, SAN, object clusters, and cloud buckets, agents choke.

Pure addresses this through Fusion’s unified fast data plane, which simplifies data placement and delivers consistent performance. It reduces duplicate copies, enables better global caching, and minimizes excess infrastructure. But the bigger architectural shift demands dynamic tiering — moving away from static hot/warm/cold classifications toward automated, real-time optimization.

Oostveen expects agent-aware data placement to become critical: “The location of data should be as local as possible to inference compute. You want to minimize cross-cluster hops and ensure the context lives near the agents that use it.”

This resurrects the edge-compute conversation with renewed urgency. Inference is moving from core datacenters to factories, stores, branches, and vehicles. That means storage architecture must prioritize locality and minimize data movement at every level.

The 600TB revolution

Pure Storage’s DirectFlash Module roadmap tells the density story: 150TB in 2024, 300TB this year, and 600TB next year. That many terabytes in the palm of your hand sounds like science fiction, but the physics are sound. Combined with enhanced data reduction capabilities, like compression and deduplication advances that keep pushing effective capacity higher, the gap between flash and disk architectures widens every quarter.

Hard drives still dominate hyperscale cold storage, but that transition is accelerating. Yet, simply replacing hard drives with flash is not so simple. The inefficiency of managing flash as if it were a spinning disk — with controller chips, translation layers, and redundant firmware — adds unnecessary power draw.

Enter DirectFlash, which takes a system-level approach to flash management to eliminate those inefficiencies.

With FlashArray™ and FlashBlade®, the partnership allowed RSC to rapidly analyze both structured and unstructured data. One major reason was that Pure’s solution delivered an 80% lower total cost of ownership than alternatives. That TCO advantage came primarily from energy efficiency, resulting in fewer hardware components that consume significantly less power and floor space.

What this means for architecture decisions

Storage architecture choices now cascade through every layer of datacenter design. Choose poorly, and you’re buying inefficient infrastructure and knee capping your AI ambitions from the start.

The trade-offs are stark: a legacy storage architecture based on spinning disks forces you to run fewer units per rack due to power and cooling constraints. That power could feed GPUs, and that cooling capacity could support higher-density compute. Every watt wasted on storage is a watt unavailable for the workloads that actually generate value.

It’s why Oostveen thinks an all-flash datacenter has become a prerequisite for competing in AI. And not just any flash—architectures optimized at the system level, but with global flash management, data reduction built-in, and energy awareness embedded throughout the stack.

As Oostveen notes, this isn’t really about replacing one technology with another. It’s about recognizing that in a power-starved world, efficiency becomes the foundation. Storage architecture decisions made today will determine which organizations can scale AI workloads tomorrow and which will find themselves stuck, watching competitors overtake because they have watts to spare.

Image credit: iStockphoto/Gearstd