Memory Is the New Oil

Every commodity cycle produces one input that quietly becomes the choke point while attention is focused elsewhere. In the AI buildout, that input is not the GPU. It is the stack of DRAM dies sitting next to it.

High-bandwidth memory — HBM — is the interposer-mounted, vertically stacked DRAM that feeds data to AI accelerators fast enough to keep their arithmetic units busy. No HBM, no accelerator. And unlike logic, which TSMC can (slowly, expensively) scale, HBM is produced at volume by exactly three companies: SK hynix, Samsung, and Micron.

An OPEC of three

The structure of this market would make a 1970s oil minister blush:

• Demand is inelastic. An accelerator without memory is a very expensive space heater. Buyers cannot substitute, defer, or design around it on any timescale that matters.
• Supply is quota-like. HBM capacity is sold out via prepaid, multi-year agreements. Industry trackers put HBM above 30% of total DRAM revenue in 2026, on wafer volumes that are a fraction of commodity DRAM — the margin structure of a cartel, achieved without a cartel.
• Conversion is zero-sum. Every wafer converted to HBM production is a wafer taken from conventional DRAM, which is why laptop, phone, and server memory prices spiked through late 2025. The AI buildout is being partially financed by everyone who buys a consumer electronic device.

The oil analogy is imperfect in one important way: oil producers can pump more. An HBM line takes years and billions to stand up, and the yield curve on 12-high stacks is brutal.

The geopolitics arrive on schedule

Where there is a chokepoint, there is statecraft. Export-control regimes that began with logic have extended to advanced memory. Korea's two champions sit inside the U.S. alliance system but operate fabs and customer relationships that span the divide; Micron carries the American flag with Idaho and New York fabs that are themselves years from full HBM scale. China's CXMT is racing up the older-generation curve precisely because Beijing read this map early.

Capital markets have started to price memory as a strategic asset rather than a cyclical commodity — the "memory supercycle" thesis. The skeptic's case is that every memory supercycle in history ended in oversupply, and that the industry is currently making the largest coordinated capacity bet it has ever made. The bull's case is that this time the demand curve is set by a compute buildout with sovereign backing on three continents.

Both can be true in sequence. Oil, after all, has crashed many times. It never stopped being strategic.