There is a repricing underway that most buyers experience only as a vague sense that hardware “got expensive lately,” without seeing the machinery behind it. A drive that should cost sixty dollars costs a hundred and forty. Memory that was cheap a year ago is scarce. Graphics cards that were briefly affordable are gone again. Each of these feels like an isolated annoyance. They are not isolated. They are the same event, viewed from different aisles of the same store.
The event is the buildout of artificial intelligence compute infrastructure, and its effects do not stay inside the datacenter. They propagate down the supply chain into the components that everyone else — gamers, small businesses, researchers, tinkerers — depends on. Understanding the mechanism does not make the drives cheaper, but it does something more useful: it tells you how to buy through the distortion, when to wait, and when scarcity of price is about to become scarcity of availability.
This is a structural story, and structural stories are the ones worth learning, because they repeat.
Shared fabs, shared fate
Start with the physical reality that ties everything together: the components AI infrastructure consumes and the components consumers buy come off the same manufacturing base.
- Memory (DRAM) for AI accelerators and memory for your desktop are made by the same handful of manufacturers on the same fabrication lines. High-bandwidth memory for datacenter accelerators competes for the same fab capacity, the same wafers, and the same engineering priority as ordinary DDR.
- Storage (NAND flash and hard drives) for AI training datasets, checkpoints, and the enormous data lakes that feed models is produced by the same manufacturers who make the SSDs and hard drives on retail shelves.
- Silicon (GPUs and accelerators) for AI compute shares foundry capacity, advanced packaging, and the same constrained supply of leading-edge manufacturing with the broader chip market.
When a wave of demand hits any one of these shared bases, it does not politely confine itself to the datacenter SKUs. Manufacturers have finite capacity. Every wafer allocated to a high-margin datacenter product is a wafer not allocated to a consumer product. The supply chain is not a set of separate pipes; it is a shared reservoir, and when the biggest buyers drink deeply, the level drops for everyone.
This is why a surge in AI infrastructure spending shows up, months later, as a price spike on a four-terabyte hard drive that has nothing to do with AI. Same reservoir.
The allocation cascade
The second mechanism is prioritization, and it explains why the pain is not evenly distributed — why small, cheap components often get hit hardest.
When supply is constrained, manufacturers rationally allocate scarce capacity to their most profitable products and their largest, most strategic customers. In a shortage, this produces a cascade:
- Highest-margin datacenter products get first claim on capacity. These are the enterprise memory, enterprise storage, and accelerator products that AI buyers purchase by the pallet.
- Retail supply of high-capacity components thins, because those high-capacity units are the ones closest to the enterprise products in the production queue, and some are literally the same parts.
- Low-margin commodity products get deprioritized. The small, cheap components — the entry-capacity drives, the mainstream memory kits — were already the thin-margin, low-priority end of the catalog. In a squeeze, they are the first to be under-produced.
The counterintuitive result: the cheapest, most commodity products can spike the hardest in percentage terms. A component that was already priced near cost has nowhere to absorb a supply shock except in price. This is why, during a genuine crunch, you see the value curve invert — the small drive that used to be a modest premium becomes a punishing penalty, while the large drive, closer to the enterprise product it shares a line with, sometimes holds a better cost-per-unit even as its sticker climbs.
If you have ever wondered why, in a shortage, the budget option is the one that becomes absurdly overpriced while the big-ticket item merely gets expensive, this is the answer. The budget option never had margin to give.
Why the capacity cycle makes it worse
Layer the ordinary semiconductor capacity cycle on top of the allocation cascade and you get amplification.
Fabrication capacity cannot respond quickly to demand. A new plant costs billions and takes years to build and qualify. So when AI demand surges, supply is fixed in the near term, and the entire adjustment happens through price. Manufacturers, seeing high prices and strong demand, commit to new capacity — but that capacity arrives years later, frequently after the initial demand spike has matured, and often all at once.
This is the same lagged capacity cycle that governs oil, shipping, and semiconductors generally: supply overshoots in both directions because it cannot see the future and cannot move quickly. In the near term, a demand shock like the AI buildout produces shortage and price spikes. In the longer term, the induced capacity can produce glut and collapse — which is why commodity hardware has always whipsawed, and why the current tightness is not necessarily permanent.
For a buyer, the cycle carries a clear message: the distortion is real, but it is also temporal. Prices this far above the historical trend tend to mean-revert as capacity catches up — unless demand keeps growing fast enough to outrun the new supply, which is the open question that determines whether current pricing is a spike or a new floor.
How to read whether you are in a “wait” or a “buy now” regime
The most valuable skill during a supply distortion is distinguishing two very different situations that feel superficially similar.
Regime one: expensive but available. Prices are elevated, sometimes badly, but the product is in stock at multiple sellers. This is a price problem, and price problems are usually temporary. The correct response, if your need is not urgent, is to wait — set an alert at a sane unit price and let capacity catch up to demand. Buying into an expensive-but-available market is paying a premium to avoid a delay you can afford.
Regime two: vanishing availability. Prices are elevated and stock is thinning — fewer sellers, “out of stock” spreading, backorders lengthening, sellers steering you to “request a quote.” This is a supply problem, and supply problems can end in the worst state of all: unavailable at any price. When you see availability itself eroding, the calculus flips. Overpaying is bad; being unable to buy the thing you need to operate is worse. In this regime, buy what you genuinely need before it disappears, even at an unpleasant price.
The mistake buyers make is treating both regimes identically — either panic-buying into an expensive-but-available market (overpaying needlessly) or stubbornly waiting through a vanishing-availability market (and getting shut out). Watch two signals, not one: the unit price and the breadth of availability. Price tells you the cost of buying now. Availability tells you the risk of not buying now.
The through-line: normalize, then decide
Every practical response to this distortion runs through the same discipline that governs commodity buying in calm times: reduce the market to a comparable unit, then decide.
- For storage, the unit is dollars per terabyte. The crunch inverted the curve, so the best terabytes moved to larger and external drives. You find them by ranking $/TB, not by trusting the old “buy small” rule — which is what our hourly $/TB hard drive board does automatically.
- For memory, the unit is dollars per gigabyte. The crunch amplified an already-brutal cycle, so timing against the $/GB history matters more than ever; our DDR5 price-per-gigabyte board tracks exactly where today sits on that curve.
- For GPUs, the units are dollars per VRAM gigabyte and dollars per teraflop. The crunch pushed the best unit value into used and prior-generation cards as datacenters absorbed the new supply — visible at a glance on our GPU $/VRAM-GB rankings.
In each case the distortion is different in detail but identical in structure, and the tool is the same. The unit price cuts through the noise the market generates precisely to prevent you from seeing clearly. And the unit price history — where the number sits relative to its own past — tells you whether today is a peak to wait out or a floor to buy on.
Mark’s Take: The AI buildout is the largest demand shock the computing supply chain has absorbed in a generation, and its most underappreciated feature is reach. It does not stay in the datacenter. It reaches into the aisle where a researcher buys an archive drive and a small business buys memory, and it reprices both through shared fabs and ruthless allocation. You cannot fight it. But you can read it. Normalize every purchase to its unit price, watch that price against its own history, and watch availability alongside it. When the market is expensive but stocked, wait. When it is expensive and emptying, move. The buyers who get hurt are the ones who see a sticker price and react to it. The buyers who do fine are the ones who see the plumbing behind it.
The bottom line
The price of a hard drive, a memory kit, and a graphics card are not three separate stories. They are one story — the repricing of a shared manufacturing base under the largest compute demand shock in decades — told in three aisles. The mechanism is shared fabs, ruthless allocation to the highest-margin buyers, and a capacity cycle that cannot respond in time.
You do not need to predict where it goes to buy well through it. You need to normalize each purchase to the unit that matters, read that unit against its own history to judge peak versus floor, and watch availability to know whether you are in a wait regime or a buy-now regime. Do that consistently and the distortion becomes navigable — expensive, sometimes infuriating, but navigable.
We do not predict. We follow the plumbing. And right now the plumbing of the entire computing supply chain is being re-pressurized by a demand it was never built to serve.
MarketCrystal provides trend analysis and market commentary for informational purposes only. Nothing in this publication constitutes financial advice or purchasing recommendations. Hardware markets are volatile and change constantly; always verify current pricing and availability before buying. Past trends do not guarantee future results.
Follow the plumbing.