Taiwan's DRAM Failure — shantum's notes

ELI5 / TLDR

Taiwan makes the world’s best logic chips — the brains. But there’s a second kind of chip, memory, and at memory Taiwan is a perpetual also-ran. For two decades, from the late 1980s to 2009, Taiwanese companies poured billions into DRAM (the working memory inside every computer) and lost almost every time. The reason: memory is a brutal commodity business where prices swing wildly, you have to own your own technology, and you need to be huge to survive the down years. Taiwan mostly rented its technology from foreigners, stayed small, and funded everything with risky short-term bank loans. When the cycle crashed — and it always crashed — they got wiped out.

The Full Story

Two kinds of chips, two completely different games

Most people lump all chips together. They shouldn’t. There are logic chips — the brains that compute things — and memory chips — the scratchpad that holds data while the brain works. Taiwan, through TSMC and UMC, conquered logic. At memory, it never came close.

The difference is not technical so much as economic. Logic, as Taiwan ran it, is a service business. A company like TSMC just builds whatever chip the customer designed and hands it back — the customer owns the design, TSMC owns the factory. Memory is the opposite. It’s a commodity, like wheat or copper. Everyone’s DRAM chip is basically interchangeable, so the only thing that matters is who can make it cheapest, and you have to design it yourself.

DRAM is a commodity infamous for its volatility. In the early days, you often had four good years and 1 bad year.

That volatility is the whole story. Think of DRAM makers as farmers who can’t see each other’s fields. When prices are high, everyone plants like crazy — except a memory “field” is a billion-dollar factory that takes two years to build. By the time all the new factories switch on at once, there’s a glut, prices collapse, and everyone bleeds. This is the bullwhip effect: a small wobble in demand at the far end of a supply chain gets amplified into a violent swing by the time it reaches the factory, because each link in the chain over-reacts to the link before it.

Why they kept trying anyway

Taiwan’s PC industry was enormous and hungry for memory, but it had to beg the Japanese and Americans for supply. When the Great DRAM Shortage of 1988 hit, Taiwan’s computer makers literally couldn’t get enough memory to ship their products. The lesson they drew — make our own — was understandable, and wrong.

Even TSMC felt the pressure. Its first CEO, Jim Dykes, was leaned on hard to make memory and refused, sticking to logic. Morris Chang, TSMC’s founder, was a skeptic from the start. He noticed that neither the government research outfit (ITRI) nor the new joint ventures actually owned any memory design know-how — they were borrowing it.

He mused in his memoirs whether TSMC’s success as a spinoff had been “too easy” — adding that the “tragic outcome was almost inevitable”.

That “borrowing” is the recurring sin of this whole saga. Almost every Taiwanese memory company licensed its core technology from a foreigner: Acer from Texas Instruments, Nanya from Japan’s Oki, Powerchip from Mitsubishi, ProMOS from Germany’s Siemens, Winbond from Toshiba. Renting technology means paying a royalty (thinner margins) and — worse — being stranded if your partner stops upgrading. When Oki failed to move to the next generation, it nearly killed Nanya.

The first wave, and the first crash

The late-1980s shortage kicked off a gold rush. Acer started a joint venture with TI called TI-Acer. ITRI ran a government program called the Submicron Project, which got spun off via auction into a company Morris Chang reluctantly agreed to lead: Vanguard. Then a swarm of private startups and “me-too” followers piled in — Nanya, Powerchip, ProMOS, Winbond, Macronix, even a rice-cooker company briefly announced a fab.

For one glorious year it worked. The run-up to Windows 95 (which demanded more memory) sent prices soaring. In August 1995 Vanguard broke even and handed out a full month’s bonus pay.

It lasted about four months. Everyone had built too much.

They built too much. Nine new DRAM fabs in Taiwan came into production in 1995 and 1996… Worldwide, the number of fabs leapt from 73 in 1995 to 170 in 1996.

Here’s the cruel mechanics of the crash. The cost of a memory chip is mostly depreciation — the sunk cost of the billion-dollar factory, spread across every chip. The actual cash cost of running the machines and materials is tiny. So once your factory is built, you keep running it flat-out even as prices collapse, because making chips at a loss still beats making nothing and covering none of the cash. Everyone does this at once, so prices keep falling. Four-megabit chips slid from $12 to $7.50 to $3.40 to $2.30, at which point Vanguard couldn’t even cover its cash costs. The whole market shrank 60%.

The slow bleed and the exits

TI-Acer lost $58 million in 1997. Vanguard lost $155 million in 1998 — a Taiwanese corporate record. Both eventually gave up on memory. TSMC quietly absorbed TI-Acer’s factory (it became Fab 7, and apparently still looks visibly different from TSMC’s other fabs because they didn’t build it). Vanguard was talked into pivoting to logic foundry by Rick Tsai, who flatly told Morris Chang that memory had no future. Chang later called the way it ended “sad.”

The second wave — same movie, bigger budget

You’d think they’d have learned. Instead, in the early 2000s, with China sucking away investment, the Taiwanese government doubled down with the “Two Trillion and Twin Star” policy — pouring tax breaks and special funds into two industries: flat-panel displays and DRAM.

This created a quieter poison: moral hazard. By officially blessing DRAM as a strategic industry, the government signalled it would catch anyone who fell. So companies borrowed recklessly and built recklessly, assuming a bailout if it went wrong. Powerchip teamed with Japan’s Elpida on Rexchip, a $14 billion fab complex — the biggest foreign investment in Taiwan’s history at the time. Nanya teamed with Germany’s Infineon on Inotera.

The 2006 boom (everyone betting Windows Vista would gorge on memory) gave way to the 2007 price crash, which then collided with the 2008 Global Financial Crisis. Memory prices fell so far that a 1-gigabit chip cost 75 cents — “cheaper than water.” Taiwanese firms, with 13–15% of world market share, somehow absorbed 42% of the industry’s losses. The reason, again, was structural: Samsung could lean on profits from selling TVs and phones and cash from its sibling companies; the Taiwanese had financed everything with short-term bank debt that came due at the worst possible moment.

The last roll-up that nobody wanted

In 2009 the government proposed Taiwan Memory Company — merge all the medium-sized DRAM makers into one national champion big enough to fight Samsung. It died for the most human of reasons: the companies didn’t actually want to merge. They wanted bailout cash, not a shotgun wedding. They ran on incompatible technologies (some in Elpida’s camp, some in Micron’s, using physically different capacitor designs), so a quick merger was a fantasy. When prices briefly recovered in April 2009, the urgency evaporated, and the legislature voted the whole thing down.

Where it all landed

The American Micron picked up the wreckage — Elpida, Rexchip, and eventually Inotera — and today makes most of its DRAM in Taiwan with 15,000 employees there. The Taiwanese survivors mostly fled to safer ground: Powerchip became a logic foundry making display-driver chips, Winbond and Macronix went back to flash memory. Only Nanya stayed in DRAM, kept alive by deep-pocketed parent Formosa Plastics, and is now building a modern 10-nanometer fab — with revenues up roughly 600% year-on-year thanks to the current AI-driven memory shortage. A second-tier survivor, finally in the right place at the right time.

Key Takeaways

Logic chips are a service business (customer owns the design); memory is a commodity business (you own the design and compete purely on cost). Conflating the two is the central confusion the video corrects.
Taiwan’s fatal pattern: license technology from foreigners rather than own it — thin margins, and stranded when the partner stops upgrading (e.g. Oki nearly sinking Nanya).
A memory chip’s cost is mostly depreciation (sunk), not cash. So makers keep running factories during price crashes, which makes crashes worse — nobody blinks first.
The bullwhip effect: billion-dollar fabs take two years to build, so capacity arrives all at once after demand has already turned, producing violent boom-bust cycles.
Taiwanese firms held 13–15% of the market but took 42% of the industry’s 2007–2008 losses — they were over-exposed because they funded builds with short-term bank debt rather than diversified cash flow like Samsung.
Government designating DRAM “strategic” created moral hazard — implicit bailout guarantee → reckless borrowing and building.
Morris Chang was a consistent skeptic; his rule was own your core technology, which Taiwan’s DRAM ventures violated by design.
TSMC’s growth got a quiet boost from buying failed memory fabs (TI-Acer → Fab 7), pulling it ahead of rival UMC.
The DRAM industry went from 20–30 companies to effectively three (Samsung, SK Hynix, Micron), with China’s CXMT and Taiwan’s Nanya as smaller players.
Nanya is the lone Taiwanese DRAM survivor, propped up by Formosa Plastics, now riding the AI memory boom (~600% revenue growth).

Claude’s Take

This is Asianometry doing what it does best: taking an industry everyone thinks they understand and showing you the load-bearing wall nobody noticed. The thesis — that logic and memory are economically opposite businesses, and Taiwan’s playbook for winning one was precisely the playbook for losing the other — is genuinely clarifying and holds up.

The history is dense and well-sourced (Morris Chang’s memoirs, Computer History Museum oral histories, a named academic thesis), and the narrator resists the easy nationalist or triumphalist framing. The recurring villains — rented technology, sunk-cost economics, bank-debt fragility, moral hazard — are real structural explanations, not just a list of unlucky breaks. That’s the mark of good business history.

Where I’d push back gently: the video occasionally treats the bullwhip cycle as if Taiwan suffered it uniquely, when in fact everyone got hammered — Samsung, Hynix, the Japanese, Qimonda. Taiwan’s distinctive failure wasn’t the cycles themselves but its inability to absorb them (small scale, no captive cash, rented IP). The video does say this, but the framing sometimes blurs “DRAM is brutal for everyone” with “Taiwan specifically failed.” Minor. The closing irony — that Nanya, the scrappy survivor, is now booming on AI demand — is a nice button that the video earns rather than forces.

An 8. Tight, substantive, genuinely teaches you the economics rather than just narrating dates. Loses a point or two only because the parade of similar-sounding company names and serial crashes can blur together without a visual aid, and the lessons, while sharp, are not wildly novel to anyone who’s read about commodity-cycle industries before.