SK hynix Inc. SKHY

Key data

The business

SK Hynix makes memory chips. Two kinds: DRAM, the fast working memory a computer thinks in, and NAND flash, the storage that keeps data when the power is off. Three companies make almost all the world's DRAM, Samsung, SK Hynix, and Micron, because a new memory fab costs tens of billions and takes years, so no fourth player walks in. Inside DRAM sits the prize: HBM, high-bandwidth memory, eight to twelve DRAM dies stacked and wired together so they can sit inches from an AI processor and feed it fast enough. A single HBM stack sells for several times a plain DRAM chip, and SK Hynix got there first and best, which is why it is the anchor HBM supplier into Nvidia's accelerators. A small foundry business rounds it out. The moat is the oligopoly plus a yield-and-packaging lead on HBM that rivals have spent two years trying to close.

The numbers

The earnings line went vertical the moment the AI memory cycle turned. Four straight quarters, each further above what analysts penciled in.

The inflection is Q1 2026, when HBM went from a promising line item to the thing driving the whole result. Note the shape: not just growth, but growth the sell side kept under-guessing by a wider margin each quarter, which is what a genuine supply shortage looks like from the outside. Gross margin sits near 76%, the kind of number a memory maker only prints when every wafer it can build is spoken for. That is the tell, and the trap.

Because this is a commodity cycle, and Korea's own export data shows how violent it is. SK Hynix is the largest piece of it.

That 24% drop into 2023, then the 76% climb out, is the same swing running through this company's earnings, one cycle earlier. Compound it forward and the math is simple and dangerous: at the ₩8,760 quarterly run rate the stock is near 5 times earnings, and at the 2023 trough it lost money outright. The 12-year record shows a year-end P/E between 39 and 661, and a loss year in the middle. Today's ≈8x trailing is the lowest the stock has ever carried, not because it is cheap but because the E is at a top. What this memo believes that the tape does not is narrower: the shortage has one more year in it than a normal cycle, because HBM is sold on multi-year commitments, not spot. The print that settles it is the DRAM contract price, monthly.

Management

There is no insider read here, and no one should manufacture one. SK Hynix is a Korean issuer that files a 6-K, not the Form 4 that captures open-market buys and sells, so the zero in the feed means "not reported," not "nobody traded." What the record does show is capital: capex runs near a fifth of revenue and is climbing into the super-cycle, into HBM lines and new fabs at home and in the United States. That is the one thing to watch on this team, because the memory industry's signature mistake is building its biggest capacity right as demand rolls, and every operator swears this time is disciplined. Pay and proxy detail are not in this run.

How it fails or surprises you

Memory prices roll over. The whole thesis is a price. When all three makers finish building and HBM plus standard DRAM tip into oversupply, contract prices fall and the 76% margin unwinds fast. Watch DRAM contract price and inventory days: two consecutive down months turns a 5x run-rate multiple into a 30x multiple on shrinking earnings, and the stock does not wait for the second month.

HBM concentration cuts both ways. A large share of HBM revenue leans on a short list of AI-chip customers, Nvidia above all. If Nvidia fully qualifies Samsung or Micron on HBM4 and dual-sources, SK Hynix gives back both price and share at the exact top. The qualification announcements are public and are the thing to watch, not the quarterly print.

The cycle runs a year long (right tail). If hyperscaler capex stays vertical through 2027 and HBM stays sold out on locked commitments, earnings compound again and the trough multiple re-rates hard. The market pays a peak-cycle discount today and prices zero second up-year. Watch the hyperscaler capex guides and any "sold out through next year" line in the results call.

Closing thoughts

Nothing in the next quarter resolves this, and the thing that matters is not the earnings print but the price of a DRAM chip, so that is what to watch rather than the headline. My read, offered as judgment and not a number: the near-term momentum genuinely favors the upside, because the commitments are signed and the shortage is real today, but the fatter tail over a full hold is the left one, since the permanent damage in memory comes from the down-leg, not the up. What is at risk if prices roll is most of the margin; what the extra year is worth if it lands is a re-rating off a trough multiple. Both are large, which is the nature of the name.

The bet is still that the world keeps building AI data centers and keeps buying SK Hynix's memory chips to feed the processors that run those models, and that memory prices stay high instead of collapsing. What breaks it is two numbers moving together, the DRAM contract price turning down while inventory days turn up. Watch that pair and you will know before the income statement does. This memo is wrong if HBM contract pricing softens in 2026 while SK Hynix is still expanding capacity into it.