First Through the SiliconNarrow moat

SK hynix (SKHY) — moat facet

The advantage was never the idea of stacking DRAM; it was doing it profitably at volume before anyone else had finished arguing about whether it mattered.

Being first is usually worth less than it sounds. In HBM it was worth a great deal, because SK hynix was not first to a product so much as first to a manufacturing method that the rest of the industry then had to adopt.

Why stacking is hard, and why being first matteredThin the wafersDrill throughthe siliconBond 12dies, alignedTest the stack,not the dieFirst to mass-produce multiple HBM generations, including HBM3 and HBM3E.
The idea was public for years. Doing it profitably at volume was not.

Through-silicon vias — drilling connections straight down through a stack of dies rather than routing them around the edge — are what make high bandwidth physically possible in a package small enough to sit beside a processor. SK hynix built the first HBM using TSV packaging1, which meant it accumulated the practical knowledge of how such a stack fails several years before anyone needed to compete with it.

That head start compounded in an unobvious way. The hard part of HBM is not the design, which is a published standard, but bonding a dozen thin dies without breaking any of the thousands of connections. Every wafer run teaches you something about that, and the company with the most runs learns fastest. By the time HBM mattered commercially, SK hynix had been failing at it, productively, for longer than anybody else.

What it does not buy is permanence. A standard that everyone implements is a standard everyone eventually implements adequately.

The measure is generational: whether SK hynix is first to volume on each new HBM standard, or merely present at it.

Moat trajectory: Narrowing

A head start is consumed by definition. What is left of it is measured in generations, not years, and each generation is re-tendered.

The number that tests this moat
Reported
Operating profit growth, quarter on quarter
+61% in Q2 2026

Being first to each HBM generation has let SK hynix sell the newest parts before rivals qualify. Profit still rising quarter on quarter as HBM4 ships says the lead continues.

Source: SK hynix Q2 2026 results ↗
⚠ Threats to the moat
⚠ THREAT 01
Being First Stops Mattering Once Everyone Is Adequate
A buyer does not need a better stack from a rival -- only an adequate one, at which point the premium is negotiated rather than earned.
References
  1. ReportedSK hynix built the first HBM using through-silicon-via packaging.
    SK hynix Inc., Form 424B4 prospectus for its Nasdaq offering (SEC, CIK 2120882) — FY2025 revenue W 97,147bn, cost of sales W 38,456bn, gross profit W 58,691bn, S&A W 5,019bn, R&D W 6,466bn and profit for the year W 42,948bn, against FY2024 revenue W 66,193bn / profit W 19,797bn and FY2023 revenue W 32,766bn / cost of sales W 33,299bn / gross loss W 533bn / loss for the year W 9,138bn; Q1 2026 revenue W 52,576bn, gross profit W 41,679bn, profit W 40,346bn. IDC market shares for Q1 2026: second in DRAM at 29.1%, first in HBM at 56.4%, second in NAND at 18.5%, with the three DRAM producers together above 90% of revenue. DRAM products were 77.1% of total sales in 2025 and 77.3% in Q1 2026; NAND 21.3% and 22.0%. Competitors named: Samsung Electronics, Micron Technology and CXMT in DRAM; Samsung Electronics, Kioxia, Micron Technology and Sandisk in NAND; competitive factors listed as 'pricing; manufacturing costs, yields and product availability; product performance, quality and reliability'. First to develop HBM using TSV packaging, commercialised HBM3E in 2024 and developed HBM4 in 2025; HBM carried a per-gigabyte price premium of more than five times traditional DRAM in 2025. M15X cleanroom opened October 2025 with wafer input from Q1 2026; Yongin construction began February 2025 with the first fab's phase-one cleanroom expected Q1 2027; an advanced packaging plant (P&T7) is under construction in Cheongju and one is planned in Indiana for the second half of 2028. Intel's NAND business cost US$6.6bn in December 2021 plus US$2.2bn in March 2025, operated as Solidigm; NAND mass production is transitioning from 176-layer to 238- and 321-layer technologies. Gartner forecasts DRAM revenue of US$143bn in 2025 rising to US$401bn in 2027, HBM US$33bn to US$86bn and NAND US$68bn to US$341bn. A substantial portion of sales is attributable to a limited number of customers located in the United States and China; the two largest were 14.8% and 12.4% of revenue in Q1 2026 and the largest was 23.9% of revenue in 2025. The offering was 177,900,000 ADSs at US$149.00, each ADS one-tenth of a common share, representing 17,790,000 shares against 712,702,365 outstanding; Baillie Gifford, Coatue and Situational Awareness Partners indicated non-binding interest of up to US$7bn. — FY2023-FY2025 and Q1 2026 · publ. 2026-07-10 · source ↗
Sources
Generated September 23, 2026