SK Hynix debuted on the Nasdaq with a rise of 13% last Friday, after placing $26.5 billion in the largest IPO of a foreign company in the history of the United States.
And that money does not go to dividends or buybacks. He is going to build more factories.
What has happened. The South Korean company has sold 177.9 million ADRs at $149 each, with a 2.9% premium over its price in Seoul, something unusual in this type of placement (they usually come out at a discount to attract buyers). Demand has multiplied the available supply by seven. On the first day of trading, the title reached $170.
With this operation, SK Hynix becomes the second South Korean company to exceed a trillion dollars in capitalization, after Samsung. In the last twelve months its stock has risen more than 630% in its home market, Seoul.
Why is it important. The memory sector has been trapped in the same pattern since the nineties:
- Excess demand.
- Huge investments in capacity.
- Oversupply
- And price collapse.
It has happened in 1997, in 2001 and in 2008. This time SK Hynix defends that AI has broken that pattern forever, that the demand for high-performance memory has no foreseeable ceiling.
And to demonstrate it, it does the only thing that lit the fuse in the three previous crises: pouring billions into expanding factories just when margins are at historic highs.
In figures:
- SK Hynix controls just over half of the global market for HBM memory, which NVIDIA GPUs need to train AI models.
- Its operating margin for the first quarter of 2026 was 72%, with a net margin of 77%.
- Together with Samsung, it has committed more than $500 billion to two new manufacturing complexes.
- The fund for extreme ultraviolet lithography equipment alone amounts to about $7.9 billion.
The context. All this has a technical reason: the 16-layer HBM4. NVIDIA has asked SK Hynix, Samsung and Micron to have it ready before the end of 2026, when the previous 12-layer standard has not even entered mass production. Stacking 16 layers forces each wafer to be thinned to 30 micrometers, a third of a hair, with an almost zero margin of error: a single defect ruins the stack.
SK Hynix has already shown a 48 GB module and plans to mass produce it this quarter. Samsung lags behind, with a yield per wafer of just 10% at the beginning of the year. Micron has preferred not to risk: it has already sold all of its HBM4 production from 2026 and postpones the jump to 16 layers until 2027. Whoever wins this race keeps the orders for the next NVIDIA GPU platform, where SK Hynix would already supply around 70% of the memory.
Between the lines. The market says it believes in the end of the cycle, but it does not act as such. What you say vs what you do: In June, a simple comment from SK Hynix about slowing down part of its AI memory business led to the fifth-worst day in Kospi history. Micron, its reference in the United States, continues to grow a lot, but no longer at the rate of a few quarters ago.
Yes, but. SK Hynix, Samsung and Micron together control more than 90% of the market and face a lawsuit for alleged price pact in the jump from standard DRAM to HBM. And all three are expanding capacity at the same time, with the Micron factory in New York not contributing real volume until 2028.
When all this production arrives together, the sector will return to the usual question: what happens when supply catches up to demand. SK Hynix is betting that, with AI, the answer is no longer the same.
In | 0.3 nm chips have a definitive date: the IMEC laboratory has found a way to extend the life of silicon
Featured image |
