Here's a number that should stop you mid-scroll: 54 trillion won. That's about $38 billion, and on Thursday SK Hynix's board approved every last won of it to build two enormous new chip factories in South Korea (Korea JoongAng Daily).

The catch — and it's a wild one — is that neither factory will ship a single chip for years. Ground doesn't even break on the first one until early 2027. The first clean room in the bigger fab doesn't open until June 2029 (SK hynix Newsroom).

So the world's most important AI memory maker is spending the GDP of a small country on capacity that arrives closer to the 2030s than to today. The bet underneath it: the AI boom isn't a spike, it's the new baseline — and the company that has the memory ready when everyone needs it wins the decade.

🧠 Why This Matters

You've heard endlessly about Nvidia's GPUs. What gets less airtime is that a top-end AI accelerator is useless without the stacks of ultra-fast memory bolted right next to it. That memory is called HBM — high-bandwidth memory — and SK Hynix makes most of it on Earth.

How much? Roughly 62% of the global HBM market, according to industry trackers, with Samsung and Micron splitting most of the rest (Astute Group). SK Hynix is Nvidia's primary HBM supplier, and it has already sold out its memory production through 2026 (CNBC). When demand outruns what you can physically build, the only lever left is more factories.

That's what $38 billion buys: not a product, but the right to keep saying yes to customers three and four years from now.

📊 Deep Dive

The money splits across two sites, each aimed at a different piece of the AI stack.

  • Yongin Y2 — 35.2 trillion won (~$25B): the crown jewel. It's built for next-generation DRAM and HBM, the memory that feeds AI accelerators. Ground breaks in July 2027; the first clean room opens June 2029; build-out runs through October 2031.
  • Cheongju M17 — 19.1 trillion won (~$13B): a NAND flash fab, aimed at the enterprise SSDs that stuff AI data centers full of storage. Construction starts February 2027; first clean room December 2028.
  • The accelerator: SK Hynix pulled the completion target for its four planned Yongin fabs all the way forward — to 2033, from an original 2045. Twelve years, erased from the timeline.
  • The thesis in one figure: the company projects DRAM and NAND demand will grow at a 19% compound annual rate through 2030, the assumption on which the whole $38 billion rests.

And they're spending from a position of absurd strength. In the second quarter of 2026 SK Hynix posted record results with an operating margin around 76% — for context, that's software-company profitability on a business that pours concrete and buys billion-dollar lithography machines (Investing.com). HBM4, the next memory generation, began mass shipments that same quarter.

"In the AI era, technological competitiveness alone is not enough. The ability to supply customers with the products they need when they need them is itself a source of competitiveness." — SK hynix, on the investment (The Korea Herald)

⚠️ The Catch

Memory is the most brutally cyclical business in tech. It runs hot, everyone floods in with capacity, the glut arrives, prices crater, and the same companies that printed record profits start writing down inventory. SK Hynix has lived through several of these whiplashes — including a bruising downturn only a couple of years ago.

That's the danger of a factory that opens in 2029. You're committing tens of billions today against demand you're guessing at for the back half of the decade. If the AI buildout cools, or if buyers digest the capacity slower than the models multiply, those clean rooms open into a softer market than the one that justified them. Build too late and you miss the wave; build too early and you drown in your own supply.

There's competitive risk too. Samsung has committed to its own massive semiconductor spending, and Micron has been closing the HBM gap. A 62% share is a lead, not a moat — and the next fight is over HBM4, where the rankings could still shuffle.

🎯 What Happens Next

Watch the calendar. Cheongju M17 breaks ground first, in February 2027, followed by Yongin Y2 that July. Before any of that, the near-term signal is HBM4: whoever qualifies fastest with Nvidia and ships in volume sets the pricing for the next cycle. SK Hynix says it's already shipping HBM4 — the question is how much, and how fast rivals answer.

The other thing to watch is whether SK Hynix keeps selling out. As long as its memory is spoken for a year in advance, the $38 billion looks like prudence. The day order books soften is the day this bet starts to look expensive.

🧩 Bigger Picture

Step back and the SK Hynix decision is a tell about where the industry thinks value is migrating. For years the glamour lived in the logic chip — the GPU, the CPU, the thing with the brand name. Increasingly, the bottleneck is memory: the AI models are so hungry for bandwidth that the chips feeding them have become the scarce resource.

SK Hynix put that shift into its own words, calling memory a "structural transformation" where it "transcends its role as a mere component to become core infrastructure determining AI performance itself" (SK hynix Newsroom). Translated: the stuff you used to think of as a spec-sheet afterthought is now the thing that decides how fast the future runs.

Which is why $38 billion, spread across two factories that stay dark until the end of the decade, isn't as reckless as it sounds. It's a company that reads the memory shortage as permanent placing its chips — literally — years before the table is even set.

SK Hynix isn't building for the AI boom you're watching now. It's building for the one it's betting you can't yet imagine.

Sources