Acquired · Thread
Semiconductor supply
Each of these episodes is part of the same story about who can make leading-edge silicon and what the rest of the world owes them for it.
Each of the threads on this site is an arc — a set of episodes that only make sense as installments of the same argument. This page is the in-order index of one of them.
Background — the chip industry in one page
Semiconductors are the most concentrated industry in modern capitalism. Nearly all advanced logic chips — the ones in your iPhone, NVIDIA GPUs, AMD server CPUs — come from three fabs in Taiwan, all owned by TSMC. Memory chips are made primarily by Samsung, SK Hynix, and Micron. The tools that make leading-edge fabs possible at all come from one company in the Netherlands, ASML. Each new process node costs $15-25B to build a single fab; the cumulative spend since 2015 across TSMC, Intel, and Samsung's semiconductor division exceeds $575B.
The seven episodes in this thread tell the story from both sides. TSMC and Morris Chang show the supply side: how a pure-play foundry — a company that refuses to compete with its own customers — became the bottleneck of the global economy. NVIDIA and Jensen Huang show the demand side: how a fabless chip designer placed a decade-early bet on CUDA, then captured the entire AI compute wave when it finally arrived. Qualcomm bridges the two — fabless design plus IP licensing as a hybrid power.
Underneath both sides is the same fact: only a few companies on earth can credibly fund the next process node. The chart below shows why.
Live data
Cumulative capex 2015–2024
Cumulative capex (US$ billions) per company, 2015–2024. The biggest infrastructure spend of any tech-industry segment in modern history.
Samsung (DS / Semiconductor)
IDM (designs + manufactures) · South Korea
TSMC
pure-play foundry · Taiwan
Intel
IDM (designs + manufactures) · United States
Micron
memory specialist · United States
ASML
equipment maker (lithography) · Netherlands
NVIDIA
fabless designer · United States
Year-by-year capex (small multiples)
Samsung (DS / Semiconductor)
$13.0 → $35.0BTSMC
$8.0 → $30.0BIntel
$7.0 → $25.0BMicron
$4.0 → $8.0BASML
$0.3 → $2.2BNVIDIA
$0.2 → $3.2BAs of 2026-05 · Figures rounded to nearest $0.5B · Samsung row is the DS division (not parent group) · NVIDIA is fabless (fab capex lives inside TSMC's number)
Three things to know
Why the post-2020 spike happened
Three shocks fired at once
First the pandemic-era chip shortage exposed how thin the leading-edge buffer was — every car, phone, and console maker hit allocation walls simultaneously in 2020-21. Then the generative-AI demand wave landed in 2022-23 and the same hyperscalers that had been buying GPUs in the thousands started buying them in the hundreds of thousands. Finally the 2022 US CHIPS Act made $52B in subsidies available, forcing every foundry to commit to new fab construction or lose the political coverage. The combined effect: TSMC's capex went from ~$17B (2020) to $30B+ sustained from 2021 onward; Samsung DS roughly doubled in two years; Intel's IDM 2.0 strategy added $20B+ annually.
Why ASML matters more than its capex suggests
Bottleneck pricing inverts the math
ASML's own capex (~$2B/yr) is dwarfed by what its customers pay for its machines — TSMC, Samsung, and Intel collectively spend $10-15B/yr on ASML EUV lithography tools (each leading-edge machine runs $200-400M and the next-gen High-NA EUV starts at ~$380M). ASML is the only company on earth that makes them. That's bottleneck pricing: the supplier captures most of the value in the build-out, even though it sits one tier upstream from the visible foundries. Without ASML throughput, the foundries' capex plans are paper.
Geopolitical layer — three sites, one chokepoint
Concentration is the strategy AND the risk
TSMC's leading-edge fabs are in Hsinchu and Tainan, Taiwan. Samsung's are in Hwaseong and Pyeongtaek, South Korea. ASML's EUV manufacturing is in Veldhoven, Netherlands. China's SMIC has been blocked from buying ASML EUV since 2019 export controls and cannot credibly compete at ≤7nm without them. The US CHIPS Act funded TSMC Arizona and Samsung Texas in part to thin this concentration — but the actual production volume remains anchored in East Asia, with US-equipment dependence (ASML, Applied Materials, KLA, Lam Research) as the lever the US uses to police China's access.
Live data endpoint: /api/semiconductor-capex — returns JSON {meta, companies: [...]}, cached 1h/24h
Concepts the thread keeps returning to
Capital intensity as moat
Industries where the cost to build the next unit of capacity is so large that only a few players can afford to play — and the rest of the system bends around them.
Geopolitical moat
Strategic advantage that comes from being indispensable — and difficult to displace — for reasons that are about geography, defence, and policy as much as economics.
The CUDA bet
NVIDIA's decade-early commitment to programmable GPUs as general-purpose compute, before there was a use case — which became the entire AI moat.
The arc · 7 episodes over 4 years
Each of these episodes is part of the same story about who can make leading-edge silicon and what the rest of the world owes them for it.
Episodes in this thread