Analysis
The Register's July 12 analysis makes a case that's easy to forget amid the current AI infrastructure euphoria: memory chipmakers have always lived on a structural boom-bust cycle, and the AI-driven high-bandwidth memory supercycle now underway is producing the most extreme version of that swing the industry has seen.
The underlying mechanic is straightforward but consequential. Memory fabs take years to plan, permit and build, meaning capacity decisions made today are actually bets on AI compute demand holding steady several years into the future -- a fundamentally different risk profile than software companies scaling cloud capacity on-demand within weeks. SK Hynix, Samsung and Micron are all currently expanding HBM capacity aggressively, and each is implicitly betting the current AI capex supercycle continues on roughly its current trajectory long enough to absorb that new supply profitably.
The timing makes the tension unusually visible. SK Hynix just completed a $26.5 billion US IPO explicitly to fund new HBM production, arriving the same week VentureBeat's enterprise survey found 86% of companies running their own GPUs report utilization of half capacity or less -- meaning the capital markets are financing more memory supply at the exact moment buy-side data suggests existing AI compute isn't being fully utilized.
“The underlying mechanic is straightforward but consequential.”
History offers real precedent for how fast this can turn. Memory markets have swung from severe shortage, with chipmakers unable to meet demand and prices spiking, to oversupply gluts that crushed margins industry-wide, within two-to-three-year windows multiple times over the past two decades -- well before AI created a new demand category layered on top of the industry's existing cyclicality.
For infrastructure investors, the boom-bust framing is a useful corrective to the assumption that AI-driven memory demand is a permanent, structural step-change rather than the latest iteration of a cycle the industry has lived through many times before -- SK Hynix's oversubscribed IPO doesn't make the underlying cyclicality disappear, it just means more public shareholders are now exposed to it. For founders building AI infrastructure or hardware-adjacent businesses, memory pricing volatility remains a real input-cost risk worth actively hedging or contracting around, not a background assumption to ignore.
The bear case: even chipmakers themselves would argue AI-driven HBM demand is different in kind from prior cycles, given genuinely novel end-use demand from a fast-growing new compute category rather than a replacement cycle for existing hardware. What to watch next: whether HBM pricing shows early signs of softening as new capacity from SK Hynix, Samsung and Micron all comes online roughly simultaneously over the next two to three years, and whether enterprise GPU utilization improves enough to validate the current capacity buildout.
