# The Company That Wouldn't Sell, and a Business That Lost Money for Ten Years

> How SK hynix went from a debt-laden merger nobody wanted to the heart of AI memory. Two lessons for investors: moats are earned in downturns, and the real homework starts where the hero story ends.

Published: 2026-08-15
Locale: en
Tags: semiconductor, memory, hbm, moat, education

TL;DR: In 2002 Hynix's board voted down Micron's buyout offer, unanimously. Twenty-four years later the company sits at the heart of AI memory. HBM lost money for a decade and they kept at it, not because the spreadsheet worked but because AMD wanted it. The episode leaves two questions unanswered, though: what usually follows when a memory maker announces record capex into sold-out supply? And if packaging decided this race, who controls packaging's upstream?


![Oil painting cover: a craftsman stacking glowing wafers by lamplight while a storm rages over a shipyard outside](/covers/sk-hynix-hbm-moat-cover.png)

In April 2002, Hynix's ten-member board voted on Micron's buyout offer. They rejected it, unanimously. This was a company stitched together from the wreckage of the Asian financial crisis, carrying over ten billion dollars of debt, fresh out of creditor management. By any sober financial logic, selling was the right answer.

Twenty-four years later, SK hynix is the heart of the world's AI memory supply. That vote may have been the most expensive "no" in Korean semiconductor history.

Asianometry covered this story in a recent podcast episode ("SK hynix and the HBM Revolution," August 13, 2026), and it's told well. Here's the chain of reasoning I took from it, plus two questions the episode never asks — which, for investors, matter more than the story itself.

## Moats aren't calculated. They're earned in bad years.

What stayed with me wasn't the comeback. It was the ugly decade in the middle.

HBM — the high-bandwidth memory that every AI server now fights over — spent roughly ten years losing money, by the episode's account. The first generation didn't sell. The second generation bet on the wrong process node; yields fell to about half, the design team went through three leaders in eighteen months, and the CEO got an earful from customers. The packaging plant they finally built opened straight into the crypto winter of 2018.

So why keep going? The episode's answer is almost embarrassingly simple: the customer wanted it. AMD had come knocking to co-develop the technology. Lisa Su wanted this thing to exist. Chairman Chey Tae-won kept the line alive not because the model penciled out, but because a customer was standing there. The episode quotes advice Morris Chang once gave Chey: "The deeper the downturn, the better you must connect with your customers."

Then came November 30, 2022. The day ChatGPT launched, exactly one company in the world could supply HBM3 at volume. That wasn't luck. Samsung had spent those years betting on a different memory architecture designed for edge inference — a road that happened to point away from what large language models would need. Strategic forks get settled years after they're taken.

One more detail that's easy to miss: the decisive battleground wasn't process geometry. It was packaging. Stack eight or nine dies, and a single failed vertical interconnect scraps the whole stack. What broke that ceiling was a new molding method co-developed from scratch with Japanese equipment and materials suppliers, iterated dozens of times. Memory coverage usually obsesses over nanometers. This episode says the race was won at the unglamorous step of gluing chips together.

## The two questions the episode doesn't ask

First: this is a hero narrative, and hero narratives don't do valuation. Yet the episode itself hands you two facts — memory supply is booked out far into the future, and SK hynix has announced a long-term capex program of astronomical size. If you know memory-industry history, that combination should make your scalp prickle. "Sold out, expanding aggressively, vowing to be number one" is a sentence pattern with a track record, and the last company to talk this way ran straight into the DDR2 crash and the financial crisis. Maybe this time is different; AI demand really is structured differently from the handset era. But "this time is different" remains the most expensive phrase in investing. At minimum, it's worth pulling up what memory prices did in the quarters after past capex announcements of this scale before taking the story at face value.

Second: the episode tells you who won. It never asks who holds the winner by the throat. If packaging decided this race, then packaging's upstream — the molding equipment, the underfill materials, the interconnect tooling — is somebody's business. While everyone crowds onto the winner, the unglamorous, hard-to-replace links in its supply chain tend to be where the next layer of answers sits. That's the question we ask about every supply chain we study: when demand doubles, which layer breaks first?

Hero stories are fun to listen to. An investor's homework usually starts where the story ends.

## One thing to take with you

If this piece leaves you with a single idea, make it this one: moats aren't computed in good times. They're earned in the worst years, alongside your customer. HBM survived a money-losing decade because AMD was standing there.

And a practical exercise: think of the most important "customer" in your own work — an actual client, or your boss, or a partner. Recall the last time they were overwhelmed and short of help. Were you there? This week, find one chance to show up in someone's downturn — no pitch, no credit-taking, just catch what's falling. Twenty years on, you'll find your most valuable relationships grew exactly this way.