Wars of the Semis · Ethan Ding

Each ribbon is share passing from the absorbed to the absorber — or, in the founding decade, people: the genesis ribbons out of Shockley and Fairchild carry talent and technology, not booked revenue, because in this industry the engineers were the land. Red ribbons = company destroyed — bankruptcy, dismantlement, or defection. Gold ribbons = merger, carve-out or acquisition. Hover any ribbon for the deal or collapse behind it. Note who is absent: Samsung and Qualcomm have no ribbons at all — the only great powers that grew purely by building, never by buying.

Why they got so powerful. Three structural advantages, compounding for four decades. (1) The vow: when Morris Chang — 56 years old, passed over for the top job at Texas Instruments — founded TSMC in 1987 to build Taiwan an industry, he made one promise the incumbents couldn't copy: TSMC would never design a chip, never sell a product, never compete with a customer. Every IDM that offered spare foundry capacity (IBM, Samsung, TI) was asking startups to hand their designs to a competitor. TSMC's emptiness was the moat. (2) Learning from everyone's war: an IDM refines its process on its own products alone; TSMC refined its process on every customer's volume at once — Nvidia's GPUs, Qualcomm's modems, Apple's phones, AMD's CPUs, thousands of designs a year. By the 2010s its process learning curve compounded faster than anyone flying solo could match, and each fabless victory anywhere in the world was automatically a TSMC victory too. (3) No succession wars: one founder at the helm for 31 years — with one legendary exception: in 2009, aged 77, Chang un-retired, removed his own hand-picked successor, and doubled capital spending into the teeth of the financial crisis to catch the smartphone wave. That single counter-cyclical bet — building capacity while everyone else cut — is arguably the greatest capital-allocation decision in the industry's history.

Why they stayed on top while Intel, Japan and Korea warred. Strategy: they never marched on Kyoto. Intel and AMD bled each other in the x86 wars; Japan and Korea annihilated each other in DRAM price crashes; everyone with a product fought everyone else with a product. TSMC fought no one — it armed all sides and banked the margin, spending the difference digesting the world's leading-edge capacity. "Real men have fabs," Jerry Sanders of AMD had scoffed. By 2020, almost no real men still did.

Why the peak is also the exposure. The same concentration that made them indispensable made them a single point of failure: ~90% of the world's leading-edge logic is produced on one island, within range of one strait, dependent on one Dutch company's EUV machines. The "silicon shield" cuts both ways — the fortress that no rival could take is now the fortress every superpower plans around.

Walking away from DRAM looks like a rout until you see what Intel kept. It was a forced swap, not a choice freely made: Intel had invented the DRAM in 1970 and been built on memory — but by 1984 Japan's VLSI-project champions (NEC, Hitachi, Toshiba, Fujitsu) were selling below cost with patient bank capital, and Intel was losing money on every chip. In the famous telling, Grove asked Moore: "If the board threw us out and brought in new management, what would they do?" "Get out of memories." "Then why don't we walk out the door, come back in, and do it ourselves?" Intel surrendered the business it was founded on — its five ancestral provinces — and kept one thing: the microprocessor, and with it the 1981 IBM PC design win.

From Grove's chair, the logic was concentration: (1) memory was a commodity war of capital endurance that American balance sheets could not win in 1985; (2) the microprocessor was differentiated, defensible, and — thanks to IBM's accidental gift of an open PC standard — about to become the tollbooth of the entire computer industry; (3) the x86 instruction set was a moat measured in decades of software, not months of process lead; and (4) Japan, focused on the memory prize, wasn't even contesting it. Land was the only currency of the old war; Intel switched currencies.

The gamble paid on every axis — for thirty years. The exile became the throne: from 1992 Intel was the #1 chipmaker on Earth for 25 consecutive years, at times capturing over 60% of the entire industry's profits on a sixth of its revenue. And then the same playbook was run against them. Intel passed on making the iPhone's chip in 2006 — a commodity too cheap, a margin too thin — and the mobile wave went fabless, feeding TSMC exactly the way the PC had fed Intel. AMD, dying, executed its own Kantō transfer in 2009: it gave up its ancestral fabs (spun off as GlobalFoundries) to survive as a design house — and the clean-sheet move let it come back a decade later, on TSMC's process, to besiege Intel's castle. The lesson Intel taught in 1985 — that what you surrender matters less than what the surrender lets you become — was learned best by its enemies.

In 2001, roughly 25 companies could manufacture a leading-edge logic chip. Each halving of the transistor node doubled the price of staying at the table — a fab that cost $1B in 2000 costs $20B+ today — and one by one the claimants folded their hands. The bars show how far down Moore's law each got before quitting (each unit = one halving from 90 nm): only TSMC, Samsung and Intel still hold seats, and only one of the three is not struggling. Hover a bar for each claimant's fate.

The era's major houses, in order of when they folded. "Folded" takes three forms: destroyed (bankrupt, dismantled, or scattered), absorbed (merged or carved out into a survivor), or still standing in 2025. Note the pattern: almost no one on this table was beaten on engineering — they were beaten by capital cycles: whoever could keep spending through the bottom of a memory crash or a node transition took the land of whoever couldn't.

How the share estimates were made (Fermi method): Anchors are the Dataquest/Gartner and IC Insights vendor rankings and WSTS market totals, by revenue across the merchant semiconductor market plus pure-play foundry. Documented checkpoints keep the bands honest: TI was the #1 merchant chipmaker from the late 1950s through the mid-1980s (Fairchild briefly #2 at its mid-'60s peak); NEC held #1 from 1985 to 1991 and Japan as a whole peaked near half the world market around 1988; Intel was #1 for 25 straight years (1992–2016), at times taking a majority of all industry profits; Samsung first dethroned Intel in 2017 on the memory super-cycle; TSMC passed 60% of all foundry revenue and ~90% of leading-edge output by the mid-2020s; and Nvidia became the largest semiconductor company in history in 2024–25, counting its chip content rather than full AI-system revenue. Between checkpoints, shares are interpolated from ranking tables and fab capacity. Error bars are roughly ±1–2 points.

Two things that look wrong but aren't. (1) Foundry share is double-counted — on purpose. TSMC's wafers also appear inside Nvidia's, AMD's, Qualcomm's and Apple's costs, so a purist would net it out. It is drawn as its own band anyway, because control of the wafer is the story of the modern era: roughly 40 points of "fabless" share on this chart physically passes through the TSMC band on its way to a customer. Deduct it if you like — the shape of history doesn't change. (2) The biggest chipmakers of the early decades aren't on the chart at all. Captive producers — Western Electric (AT&T) in the transistor age and IBM, arguably the largest and most advanced semiconductor manufacturer on Earth through the 1970s–80s — sold nothing on the open market, so they hold no merchant share. Like the Muromachi shogunate, their power was real but sat outside the land system; when the vertical-integration order collapsed, there was no merchant business underneath to fall back on — IBM ultimately paid GlobalFoundries $1.5B to take its fabs away (2015).