The AI boom’s ROI solution

A few years into the AI boom, and the math has exploded. Between 2021 and 2023, the scale of AI training datasets skyrocketed. The total volume of training data for notable large language models (LLMs) grew from hundreds of millions to over a trillion data points (tokens), representing an increase of roughly 1,000%–3,000% depending on the specific model, and it continues to grow.¹

So what’s one of the most important things to think about when you’re talking about AI infrastructure? Storage—managing AI token consumption, aka tokenomics—and ROI should be top of mind.

It’s all about that storage and the ROI it brings

Storage has graduated from warehouse duty; it’s an extension of accelerator memory itself. Every AI token’s economics get decided below the GPU, and yet, astonishingly, most infrastructures still bolt storage on only after sizing compute. That might be backward these days, especially if you consider that storage has become a day-zero design decision that shapes throughput, latency, and the cost of running AI at scale.

High-density flash solutions now pack more than 5x the capacity of traditional hard drives² into a deck-of-cards form factor. It’s like having a smaller boat that holds more people than your friend’s bigger boat. It’s (weird) science.

The 10 petabytes you need for your next model could cost you 417 hard drives (HDs). Ouch. But with higher-density solid-state drives (SSDs), that could drop to roughly 84. Less ouch.

And less everything. High-capacity SSDs shrink your storage footprint and power draw per terabyte. That freed-up power and rack space doesn’t vanish; it becomes budget for more GPUs. Better storage = more compute headroom.

Now that should have you thinking about ROI in 3D (you’ll see). SSDs trade higher up-front hardware costs for massive long-term operational savings. They maximize profitability by eliminating costly downtime and accelerating system productivity.

Tokenomics and why the C-suite cares

A thing or two about token pricing (input/output cost per million). Storage lowers both fixed and variable AI costs:

• Fixed: Reduces training costs (checkpointing, data I/O across GPU clusters).
• Variable: Reduces per-inference cost at serve time.

Density is a primary lever: Replacing multiple HDs with high-capacity SSDs cuts power, space, and materials, and that can’t be overstated.

ROI has another dimension: What drives tokenomics is output value versus cost.

• Output value: The services your AI infrastructure delivers—real-time inference, model serving, data retrieval, streaming, rendering. SSDs provide the low latency needed to deliver these services reliably.
• Production cost: The full stack of keeping models running—compute, storage, power, cooling, and per-token inference across your infrastructure.

Translation: If output value doesn’t beat production cost, you don’t have ROI—you have a token price with its fingers crossed.

Over the last few years, storage tech has evolved from consumer-focused SSDs to enterprise-grade solutions built for data centers and edge environments alike. What started with quad-level cell NAND in 2018 (storing 4 bits per cell instead of 3) has matured through multiple generations. Today’s fourth-gen QLC-based drives cross the 100TB threshold per device, with breakthroughs in endurance and reliability that make them viable for 24/7 workloads.

And speaking of crossing the 100TB threshold, the Solidigm 122TB D5-P5336, the first-to-market SSD to break the 100TB barrier, put the “e” in eSSD. That’s enterprise solid-state drive. In fact, they coined the initialism eSSD. But more than that, when you consider how much storage 122TB holds (the entire collected works of William Shakespeare over 17 million times³), it’s proof that density and efficiency will define who scales profitably in a resource-constrained, heavily regulated world.

Fun and games

At AI Field Day 8 in May, their engineers walked through the anatomy of a single AI prompt to reveal how much storage sits behind responses that feel instantaneous. A single user query balloons into tens of thousands of tokens as context, domain rules, and session history get layered in. Storage assembles this comprehensive prompt before it hits the GPU. By caching stable portions of context, systems convert expensive recomputation into faster reads, freeing GPU cycles and cutting power draw. Like a ripple effect.

And the implications continue to ripple. In one benchmark, extending key-value (KV) cache to non-volatile memory express (NVMe) delivered up to 4x throughput and dramatically improved latency. And on a reasoning-heavy math test, additional output token capacity let models deliberate longer and score higher. Capacity being not merely background infrastructure but part of the answer is what is known in some circles as an attractive return on investment.

The story gets better and broader

Storage tiers now span from fast system memory through local NVMe, in-rack solutions, and network-attached drives. Each layer serves different workloads, from random reads demanding tight latency to bandwidth-heavy ops processing large blocks. The right architecture depends on balancing performance against cost for multi-agent, long-context workloads.

Meanwhile, the physical constraints defining modern AI buildouts are getting stricter. Power availability, sustainability mandates, and footprint limits are less minor concerns and more design challenges. Liquid-cooled form factors, like the world’s first cold-plate-cooled eSSD, and power-sipping optimizations help dense deployments stay viable in regulated environments. Reliability at the edge matters, too, since data doesn’t always stay in a central data center.

The companies pulling ahead are optimizing around their whole systems. And that’s where decades of focused engineering on storage matter.

Smaller is the new bigger

You can think of it as paying for more storage, but more accurately, you’re paying less for everything storage touches, and storage touches everything.

If we do the math, the answer will always be the same. High-capacity SSDs deliver staggering storage density while consuming less space (freeing more space for other stuff) and less energy. The return on investment could trickle down to every other cost factor in the long run. This is the moment for leaders banking that AI’s next chapter doubles down on storage, quietly carrying the load while delivering solid ROI. Decades of innovation, pure-play focus, and end-to-end portfolio—from core data centers to the edge—could make for an obvious choice.

1. ^{As tracked by epoch.ai}
2. ^{enterprisestorageforum.com/hardware/types-of-computer-memory/}
3. ^{nlp.stanford.edu/IR-book/html/htmledition/an-example-information-retrieval-problem-1.html}