A Lapse in Memory

Executive briefing

In April 2026, SK Hynix reported a 72 per cent operating margin — higher than TSMC’s, higher than Nvidia’s, higher than any publicly traded technology manufacturer of comparable scale. Micron, the only American memory manufacturer, posted 67 per cent operating margin in the same quarter, a 45-percentage-point increase from a year earlier. Samsung’s memory division, which is not separately disclosed, is estimated above fifty per cent and ramping. Three companies, in three countries, with different boards and different cost structures, printed the same anomaly in a single earnings season. Memory has been treated as a commodity for thirty years. It is no longer behaving like one.

The mechanism is structural, not cyclical. In October 2025, OpenAI signed Letters of Intent with Samsung and SK Hynix on the same day for the supply of up to 900,000 DRAM wafers per month — roughly forty per cent of global capacity if literally drawn down. Within weeks, every other hyperscaler followed. By January, all 2026 production at Samsung, Hynix and Micron — DRAM, NAND and HBM — was committed. The contracts reserve undiced wafers, leaving dicing, sorting and packaging to the manufacturer’s discretion. The instrument captures the silicon at the point of highest optionality and leaves what it becomes, or whether it ever enters the line, to be decided later.

Capacity allocation has shifted accordingly. HBM has roughly doubled its share of DRAM wafer-starts in two years and each HBM wafer removes roughly three wafers’ worth of usable commodity bits, because HBM dies are committed at the front end through irreversible processing. DDR4 has collapsed from a third of output to a rounding error. Bit shipments (the count of memory cells actually delivered) barely grew in Q4 2025, while quarterly DRAM revenue rose nearly thirty per cent on essentially flat volume. Customers are receiving roughly the same memory at much higher prices.

The market structure is not three manufacturers extracting rent from a captive hyperscaler. Both sides are aligned on the same incentive: minimise physical realisation. Fab-time is cheap to hold, and capacity reserved by one buyer cannot be redirected to another. For the hyperscaler, every undrawn wafer is a warm-socket commitment — data-centre footprint, substations, switchgear, water rights, grid interconnect — not yet made, and a wafer no competitor can access. The Letter of Intent is therefore not a procurement document but an option contract on warm-socket build-out. The reservation itself is the moat.

Antitrust law cannot reach this. In March 2022, the U.S. Court of Appeals for the Ninth Circuit affirmed dismissal of an indirect-purchaser class action against the same three companies, holding that conscious parallelism in a concentrated, interdependent market is lawful absent direct evidence of agreement — even with prior 2005 criminal convictions for explicit coordination weighing in plaintiffs’ favour as a plus factor. Each move in the current arrangement, taken alone, is defensible: an HBM allocation decision because HBM is more profitable per wafer, a commodity line run below maximum because spot prices are extreme, a consumer brand exit because enterprise demand is unmet. None of these decisions requires coordination. The aggregate is the supply discipline that the 2005 Department of Justice prosecutions described.

The deterrent math has broken down. Cumulative historical fines for explicit cartel conduct (the 2005 DOJ pleas plus the 2010 European Commission action) total roughly $1 billion in current dollars. The incremental quarterly operating profit produced by the three companies between Q4 2024 and Q4 2025 totalled approximately $22 billion. The marginal rent extracted in a single quarter exceeds the entire historical enforcement record by more than twentyfold. Any future fines would be absorbed as a cost-of-doing-business line item.

The tax is now visible. Retail DDR5 kits rose 242 per cent between May and December 2025. Server DRAM contract prices for Q1 2026 came in 60 to 70 per cent above Q4 2025, with another 30 per cent uplift queued for Q2. DDR4 chip prices briefly traded above DDR5 in mid-2025 — older, slower memory at a premium per gigabyte over newer product, unprecedented in DRAM’s history. Hyperscalers absorb the cost or pass it through to cloud customers. Networking, automotive, industrial and embedded buyers (and consumers) sit third-priority and pay it directly. The 2025-2026 retail spike is the visible tip of a capital reallocation in which Big Five hyperscaler 2026 capex, projected at $602 billion or above, flows in part through Samsung, Hynix and Micron as committed forward revenue. The architecture funds itself.

In March 2026, Iranian drones struck QatarEnergy’s Ras Laffan complex, the world’s largest LNG hub and a major helium byproduct source. South Korea sources nearly two-thirds of its helium from Qatar; Samsung and Hynix were the most exposed memory manufacturers in the world to a Ras Laffan event. Spot helium prices rose 40 per cent within weeks. None of this caused the supply discipline. It was in place by late 2024, before Ras Laffan was a known risk and before the Stargate Letters of Intent were signed. A fab running at reduced utilisation by choice looks identical, from the outside, to a fab running at reduced utilisation by force majeure. The discipline did not need helium. It needed legal cover, and helium provided it.