THE CULL MOVED TO THE BOARDROOM
How the semiconductor entrant cycle died, and why shortages now last longer
Steve Williams · babnews.org
There's a mechanism the semiconductor industry ran on for three decades that has quietly stopped working, and its failure explains something about the current moment that the shortage headlines miss. For thirty years, a shortage built companies. A price spike pulled new entrants into the market, they overbuilt, the glut that followed destroyed the weak ones, and the survivors emerged leaner. The market healed itself. That self-healing is gone now, and understanding why it's gone tells you why this constraint is lasting longer than the ones that came before it.
How the old cycle worked
Go back to the memory market of the mid-1990s and you find something that looks almost unrecognizable today: a crowd. More than 20 meaningful DRAM manufacturers were in the market, with the top ten sharing roughly 80 percent of it. That crowd existed for a specific and now-vanished reason. A leading-edge memory fab cost around $300 million in the early 1990s, and the manufacturing process, while hard, was simple enough that a well-capitalized newcomer could cross the gap to the frontier in a single generation of fabs. The barrier to trying was low enough that the market could afford to be wrong many times.
So the cycle turned like clockwork. A shortage sent prices and margins soaring, gross margins for the leaders ran well above 50 percent in the 1993-95 boom, and that pulled capital in. Nearly 50 fab construction projects were announced in 1995 and 1996 alone, and semiconductor capex climbed past 30 percent of production. Everyone built at once, against the same rosy forecasts. By 1996 the market flipped from shortage to glut, prices fell more than 60 percent, and the weak began to die.
That death was the other half of the mechanism, and it did real work. Through the late 1990s and 2000s, wave after wave of consolidation cleared the field. The 1997 Asian financial crisis pushed LG Semiconductor and Hyundai together into what became Hynix. Micron absorbed Texas Instruments' memory business in 1998. Elpida formed in 1999 from the merged DRAM operations of NEC and Hitachi. The dotcom crash of 2000-02 took down more. Qimonda went bankrupt in 2009, and Elpida followed, ending up inside Micron by 2013. Twenty-plus makers became, eventually, three: Samsung, SK Hynix, and Micron, together holding north of 90 percent.
Read that as a tragedy and you miss the point. It was a mechanism. The shortage summoned supply by pulling in entrants, and the glut destroyed the excess by culling the weak. Brutal, cyclical, wasteful, and genuinely self-correcting. When the market got tight, the market fixed it, and then the market punished everyone for fixing it too much.
The two curves that killed it
That entire loop rested on one precondition: the barrier to entry had to be low enough that new players could form on the upswing. Two curves, rising relentlessly for thirty years, finally pushed that precondition out of reach.
The first is capital cost. A leading-edge fab ran about $300 million in the early 1990s. By 2016 it was $10 billion. Today a new DRAM fab costs $15 to $20 billion, and a 2-nanometer logic fab with EUV lithography costs more still. That's on the order of a 60-fold increase in the price of admission, in an industry where the entrant's whole historical role was to place a speculative bet.
The second is the recipe. The process a newcomer had to master went from something crossable in a generation to something that now encodes a decade of accumulated yield learning, hundreds of interdependent process steps, and know-how that exists partly as tacit skill inside an experienced organization. Money can buy the tools. It cannot buy the ten years of learning that separates a working recipe from a pile of expensive equipment.
Put the two curves together and the speculative entrant, the company that builds on the boom hoping to get acquired in the bust, simply cannot exist anymore. You do not raise $20 billion to build a fab on spec, against incumbents with a ten-year process lead, in the hope that a downturn lets you sell the wreckage at a discount. The math that made the entrant cycle possible inverted.
The cull didn't disappear. It relocated.
Here is the part that reframes the whole thing, and it's the observation worth keeping. The Darwinian cull that used to run in the downturn did not go away. It moved upstream, out of the market and into the financing decision.
It used to be that the market let entrants form and then killed the losers in the bust, in public, through bankruptcy and fire-sale acquisition. Now the capital markets do the culling before a single wafer is made, by declining to fund the attempt at all. Every potential entrant meets the same verdict in the boardroom: a $20 billion bet against a decade lead is not a fundable proposition, so it never gets built. The selection pressure is identical. Only its timing changed, from after the fact to before it. The industry still weeds out the weak. It just does it at the pitch stage now, silently, which is why you never see the companies that would once have formed, boomed, and busted. They are culled as ideas, not as bankruptcies.
What replaces the entrant
If shortages no longer summon new companies, then new capacity has to come from somewhere else, and there are now exactly two sources.
The first is incumbent expansion. The three survivors add capacity, but at their own measured pace, and deliberately without the reckless overshoot that used to define the cycle, because the consolidated survivors learned that discipline is what finally made memory profitable. Micron has announced a roughly $200 billion long-term US investment plan, which sounds like a flood until you read the timeline: its first new fab does not begin meaningful DRAM production until 2027, and other sites broke ground only in 2026. Total DRAM wafer starts are growing on the order of 6 to 8 percent a year. That is not a market rushing to clear a shortage. It is a small number of disciplined incumbents adding supply on a schedule that suits them.
The second is government. When private capital refuses the bet, a state can absorb the risk that the market won't. Japan's Rapidus exists only because the government and a consortium of national champions pooled more than $16 billion to stand up a 2-nanometer foundry, licensing the process from IBM because it cannot be grown from scratch. China's CXMT scaled behind roughly $1.9 billion in state backing. These are not market entrants in the old sense. They are policy instruments, built because a nation decided the strategic value justified a bet no rational investor would make.
Both replacements share the same properties. They are slower than the old free-for-all, more concentrated in a few hands, and far more political. Neither responds to a price signal the way a crowd of 1990s entrants did.
Why the shortage lasts longer now
This is the consequence that matters, and it ties back to every constraint story in this buildout. The self-healing valve is welded shut. The classic release mechanism for a semiconductor shortage, capital racing in to build new supply, no longer functions at the leading edge, because the entrants who used to carry that capital in can no longer form. So when a shortage hits now, it does not clear on the old timeline. It persists, because the thing that used to clear it has been priced out of existing.
That's why TSMC can say, quarter after quarter, that it cannot build fast enough, and the market does not answer with a wave of new fabs. It answers with the same handful of incumbents expanding on their own schedules and the occasional state-funded moonshot. The shortage lasts as long as those two slow sources take to respond, which is years, not quarters, and there is no crowd of desperate newcomers to shorten it.
The honest other side
There is a real upside to all of this, and leaving it out would be dishonest. The death of the entrant cycle is exactly what finally made the memory industry profitable and stable. The endless boom-bust that destroyed shareholder value for thirty years was a direct product of too many players building at once. The consolidated survivors, freed from that dynamic, now behave with a discipline that produces durable profits for the first time in the industry's history. So the same structural change that removed the market's ability to self-heal a shortage also removed its tendency to self-destruct in a glut. You do not get one without the other.
But the trade is worth naming plainly. The industry bought stability and profitability at the price of responsiveness. It used to overreact to shortages and gluts alike, wildly and wastefully, and in exchange it cleared imbalances fast. Now it does neither. It is calm, concentrated, and slow, and when a genuine shortage arrives, as this one has, there is no mechanism left to resolve it quickly.
You do not remove a bottleneck. You relocate it. The industry relocated its Darwinian cull from the bust to the boardroom, and in doing so it lost the messy, wasteful, self-correcting engine that used to turn a shortage into new companies and new supply. What's left builds slower, answers to fewer people, and increasingly waits on governments to place the bets the market no longer will. The shortage used to build companies. Now it builds subsidies and incumbent capex, and that is precisely why it lasts.