SK Hynix's 72 percent operating margin is a semiconductor-industry record, and DRAM pricing is why

For every 100 won of revenue SK Hynix booked in the first quarter of 2026, 72 stayed behind as operating profit.

The full numbers, from the company's Q1 earnings, run like this. Revenue of 52.58 trillion won, up 198 percent year over year. Operating profit of 37.61 trillion won, up 405 percent year over year. CNBC, reporting the results on April 23, 2026, described the resulting 72 percent operating margin as a new all-time record for the semiconductor manufacturing industry. Not a record for memory companies. A record for semiconductor manufacturing, full stop, an industry that includes some of the most profitable enterprises on earth.

This post is an autopsy of that number, because a 72 percent margin at a memory maker is a corpse on the table of everything the industry believed about its own economics.

Memory was supposed to be a bad business

DRAM has spent four decades as the textbook commodity semiconductor. The product is standardized, the buyers are enormous, capacity arrives in billion-dollar lumps on multi-year lead times, and every boom in living memory has ended with too many wafers chasing too few sockets and margins competed down to the floor or through it. The polite word for this is cyclical. The impolite version is that memory makers have historically destroyed capital in the troughs at roughly the rate they earned it at the peaks, and everyone who underwrote the sector knew it.

A 72 percent operating margin does not belong to that industry. It is the kind of margin a monopoly software franchise prints, a business with no factories and no marginal cost. A manufacturer that pours capital into fabs, runs them around the clock, and still keeps 72 of every 100 won is telling you something, and the something is not about its costs.

It is about its customers.

The margin describes the buyer, not the seller

Work backward from the print. Profit grew 405 percent while revenue grew 198 percent, which means the margin expanded violently, which means price did the work rather than volume. No fab triples its output in a year. TrendForce puts annual DRAM capacity growth at roughly 10 to 15 percent, so a near-tripling of revenue is overwhelmingly a price event, and the first-quarter contract data agrees. Conventional DRAM contract prices rose 93 to 98 percent quarter over quarter in 1Q26, per the wafer-math arithmetic covered in the sister post.

A commodity maker only gets pricing like this when buyers hold no alternative and no timing option. They cannot switch suppliers, and they cannot wait.

Look at the supplier table and the first half becomes obvious. In HBM, the memory at the center of the AI buildout, SK Hynix holds roughly 57 percent market share, Samsung 35 to 40 percent, Micron 5 to 10 percent. Three sellers on earth, one holding a clear majority, none with idle capacity to dangle. The second half, the timing option, died with the buyers' own arms race. When the customer is a hyperscaler measuring delay against a rival's cluster count, waiting costs more than paying, at almost any price the seller cares to name. The 72 percent margin is the exact measure of that asymmetry. It is not a management achievement. It is a queue, expressed in won.

What a record margin does to a hardware underwrite

If you touch data-center economics for a living, two consequences follow, and they point in opposite directions.

Near term, the sign on your hardware assumption flipped. Every pro forma built in the last decade leaned on the same quiet premise, that compute gets cheaper over time, so a tenant's cost stack improves as the building fills. On the memory line, the largest single input by some measures, that premise is running in reverse this year, and the seller has both the market power and the sold-out order book to keep it there. Anyone modeling tenant health or expansion probability off 2025 hardware costs is holding a stale number and, worse, a stale direction.

Longer term, margins like this summon their own competition. The classic memory-cycle playbook says a 72 percent margin is an engraved invitation to build capacity, and reversion follows as surely as the trough follows the peak. The complication is the 4x wafer penalty. Each gigabyte of new HBM capacity consumes roughly four gigabytes' worth of conventional wafer output, so effective supply responds slower and at higher cost than the playbook assumes, and every wafer of the response cannibalizes conventional DRAM that is itself repricing by near-triple-digits.

The cycle is not dead. It is just slower than the people paying these prices need it to be.

My call, in a form that can embarrass me. I expect 72 percent to stand as this cycle's high-water mark for the industry. Pricing power this extreme pulls in supply eventually, even with the wafer penalty slowing the response, and record margins have a way of marking the top rather than the middle. If SK Hynix or either of the other two memory makers prints an operating margin above 72 percent within the next four quarters, then supply is even tighter than the bulls claim, the record was a waypoint rather than a peak, and every hardware-cost line in a 2026 data-center underwrite, including the ones I read, is too low.

This analysis is a source-cited research summary drawn from public reporting and market research, not investment advice. It can contain errors and should be verified independently before any investment decision.