The Memory Chip Squeeze: Why AI Demand Is Rewriting the Semiconductor Cycle

Semiconductor cycles have a well-worn reputation for running on sentiment as much as substance — a wave of enthusiasm pushes chip stocks up, capacity eventually catches up with demand, and the cycle turns hard in the other direction. What's notable about the current run in memory chip stocks is how consistently analysts covering the sector are describing it as something different: a genuine, persistent order backlog rather than a story getting ahead of itself.

The demand side of that story is fairly easy to state plainly. Building AI infrastructure at the scale current data centre construction implies requires enormous quantities of high-bandwidth memory sitting alongside the logic chips that do the actual computing — GPUs and AI accelerators need memory bandwidth to match their processing power, or the extra compute simply sits idle waiting for data. As AI training and inference workloads have scaled up faster than almost anyone forecast even two years ago, the memory component of that build-out has scaled with it, and manufacturers have struggled to add capacity at anything like the same pace.

Capacity is slow, and deliberately so

That capacity constraint isn't really a mystery, even if it's frustrating for anyone hoping for a quick supply-side fix. Building a new memory fabrication facility is a multi-year, multibillion-dollar undertaking, not a switch that flips when demand justifies it. And memory manufacturers, more than most parts of the technology supply chain, carry genuine institutional scar tissue from previous oversupply cycles — the industry has a long, painful history of adding capacity aggressively during a boom, only to watch prices collapse when that capacity finally comes online just as demand cools. That memory of past cycles, no pun particularly intended, has made manufacturers considerably more cautious about over-committing capital this time around, even with demand signals as strong as they currently are.

The earnings calendar has been reinforcing the structural story with dated, specific evidence rather than leaving it as a purely theoretical supply-demand argument. Micron's recent results sit squarely inside this narrative, alongside continued strong demand signals from Nvidia and the broader ecosystem of companies supplying the current AI infrastructure build-out. When a company reporting quarterly numbers describes an order backlog as remaining massive rather than showing early signs of clearing, that's meaningfully different information from an analyst's forward-looking projection — it's a company with actual customer orders on its books describing what it can see, not what it expects.

Why this cycle looks different — and where it might not

It's worth being specific about why this particular AI-driven demand cycle looks structurally different from prior semiconductor booms, rather than assuming the label "AI boom" alone explains the pricing action. Previous cycles were often driven by a single product category reaching mass adoption — smartphones, for instance, drove a huge memory demand wave in the early 2010s, but that demand had a reasonably predictable ceiling tied to global handset sales. AI infrastructure demand doesn't have an obviously comparable ceiling yet; it's tied to how much compute the world's largest technology companies are willing to keep building, which has shown no clear sign of plateauing so far this year.

None of this is a guarantee the cycle stays this way indefinitely, and it's worth saying so plainly rather than treating the current run as a one-way bet. Semiconductor history is genuinely littered with cycles that looked structurally different right up until the moment capacity additions finally caught up with demand and pricing inverted sharply. The specific risk here is less about whether AI infrastructure demand continues — most forecasters expect it to for some time — and more about the lag between capacity decisions made now and capacity actually coming online two or three years out, a gap wide enough that a sudden supply glut showing up in 2028 or 2029 wouldn't be a historical first.

Pricing behaviour across the memory market has reflected this backdrop fairly directly. Where previous cycles have often seen chip prices spike sharply and then correct just as sharply within a year or two, the current pattern looks more like a sustained repricing — memory contract prices moving up in a way that's held rather than immediately given back, which is generally a better signal of genuine structural tightness than a short, volatile spike would be. Buyers of memory — the hyperscale cloud providers and AI-infrastructure builders placing the actual orders — have reportedly been willing to sign longer-dated supply agreements at these higher prices, itself a signal that the buy-side doesn't expect the tightness to resolve quickly either.

The other end of the same supply chain

There's a genuine European angle to this story too, beyond the US-listed names most coverage focuses on. The infrastructure being built to run all this AI compute needs to physically exist somewhere, and Central and Eastern Europe has become an active part of that build-out — sister publication Czech Business Review has covered the data centre and energy investment flowing into the region specifically to power AI compute, a demand-side story that connects directly to the memory and chip supply picture covered here, even though the two pieces approach the same underlying AI-infrastructure boom from opposite ends of the supply chain.