holocron
中/en
← Business
Business

The Semiconductor Power Chain

In one sentence: TSMC turned Taiwan's lack of design talent, brand, and IP into an almost unclonable monopoly over chip manufacturing, using the pure-play foundry model plus learning-curve pricing. NVIDIA turned three near-death bets into an equally durable monopoly over chip design and the AI developer ecosystem, using CUDA plus a shipping cadence that outran the rest of the industry. The two monopolies are each other's only customer and only supplier — almost every NVIDIA chip is made by TSMC, and NVIDIA is the single largest buyer of TSMC's most advanced capacity — and together they form the underlying power structure of AI-era compute. The honest other half of that story: the chain's output, pricing, and direction all ultimately run through one island whose sovereignty is contested, and through a personal, nearly thirty-year relationship between two founders sealed over a handful of phone calls and one dinner at Morris Chang's house.

TSMC: Hegemon of Manufacturing (sources: TSMC：纯代工模式发明者, Morris Chang 访谈：张忠谋亲述 TSMC)

In 1987, at 56, having been pushed out of Texas Instruments and then General Instrument, Morris Chang（张忠谋） was given three days by KT Li to produce a business plan for what became TSMC. His founding logic was an honest inventory of constraints, not ambition: Taiwan had "no strength in research and development... no strength in circuit design... little strength in sales and marketing, and... almost no strength in intellectual property." The only possible strength, and even that was just potential, was wafer manufacturing — so "what kind of company would you create to fit that strength and avoid all the other weaknesses? The answer was a pure-play foundry." That decision — designing the company for its constraints rather than its ambition — grew into two separate moats. First, never compete with your customers: a pure-play foundry hands all the brand and design credit to its clients (Apple announces "3-nanometer process" on stage and keeps every ounce of the glory), which is exactly why Apple, NVIDIA, and Qualcomm trust it with their most sensitive designs, unlike Samsung, which both designs and manufactures and collides with its own customers at every turn. Second, the 学习曲线（Learning Curve） pricing Morris co-developed with BCG at TI and carried straight into TSMC's founding playbook: start low, cut prices automatically every quarter whether the market demands it or not, trade lower margin for fuller fab utilization and faster yield learning, and let scale and cost compound into a flywheel — profit into capex, capex into process leadership, process leadership into customer success, customer success into more profit.

The fuel for that flywheel is the exponential capital wall created by Moore's Law stacked on Rock's Law (fab cost doubling roughly every four years): the number of players who can keep up shrinks exponentially too. By the time of the 2021 recording, the field of leading-edge players had already collapsed from 22 in the early 2000s to just TSMC and Samsung, with TSMC holding over 90% of 5nm capacity, 50%+ of foundry market share, and a staggering 95%+ of the industry's profit. Running TSMC through the 7 Powers 护城河框架 framework, the show concludes process power here is "the clearest case I can imagine" — forty years of accumulated IP, people, know-how, and relationships with ASML that "no amount of money can replicate"; even SMIC, with money to spend on equipment, reportedly "doesn't know how to use it, and not because they're stupid." That monopoly has started cashing in its pricing power too: in 2021 TSMC first stopped cutting prices, then raised them roughly 20% — a direct reversal of Morris's own founding doctrine of automatic quarterly price cuts. According to the January 2025 interview, the hosts' own post-game synthesis pushes the endgame further out still: fab costs are projected to climb from roughly $20B toward $40B, $80B, even $100B, and "the number of players who can write a $100B check for a building full of machines will only keep shrinking... the endgame of this industry is a single dominant player."

NVIDIA: Hegemon of Design + the CUDA Ecosystem (sources: NVIDIA 之一：GPU 公司（1993-2006）, NVIDIA 之二：机器学习公司（2006-2022）, NVIDIA 之三：AI 时代的黎明（2022-2023）)

Founded in 1993 over a meal at Denny's, Jensen Huang（黄仁勋） and his two co-founders clawed their way out of a graphics-card bloodbath with 90 undifferentiated competitors, then nearly died from cash exhaustion three separate times: down to nine months of runway and a 70% layoff in 1996; an 80% stock crash in 2008 when CUDA still had zero revenue and the financial crisis hit at the same time; another 50% crash in 2011. The real turning point was the 2006 bet nobody could see a market for — CUDA. The visible addressable market at the time, mostly scientific computing, added up to maybe two or three billion dollars. Jensen's logic wasn't a demand forecast; it was a supply-side necessity: "If you don't build it, they can't come." CUDA shipped free but closed-source, runnable only on NVIDIA silicon — textbook Counter-Positioning（反向定位） in an era when the industry consensus was that open, modular Wintel-style platforms always won. Six years of zero revenue later, AlexNet trained on two consumer GeForce cards in 2012 and knocked the ImageNet error rate from 25% down to 15% — deep learning "landed squarely in NVIDIA's lap," a miracle nobody could have planned for and that even Jensen can't really claim to have foreseen, and it cashed out that supply-side bet. Over the following decade NVIDIA turned CUDA from a language into an entire developer civilization: compilers, SDKs, libraries, developer evangelism, all built in-house, an estimated ten thousand person-years of software engineering cumulatively, with four million registered developers by August 2023.

What that investment bought wasn't a single power but, in the hosts' words, "the whole stew of powers that built Apple and Microsoft" landing on one company at once: scale economies (a thousand-plus-person CUDA team amortized across millions of developers), switching costs (NZS Capital's verdict that data-center revenue and capex are "the stickiest revenue known to humanity"), cornered resources (TSMC's CoWoS advanced packaging capacity, and Mellanox, the last surviving InfiniBand supplier on earth), and branding ("NVIDIA is the modern IBM of the AI era — nobody gets fired for buying IBM," and, as the other host puts it, nobody's getting fired for buying NVIDIA anytime soon either). By the September 2023 recording of the trilogy's closing episode, training a GPT-class model had "one option" — data-center revenue had hit $10.3B in a single quarter, up 141% quarter over quarter, with gross margin climbing past 70%. The show's own nine-step thought experiment on "what it would take to compete with NVIDIA head-on" — design an equally good chip, build chip-to-chip interconnect at NVLink's level, build assembly relationships, build rack-to-rack networking to match Mellanox, convince customers to switch, secure TSMC's advanced packaging capacity, build software as good as CUDA, convince developers to abandon CUDA, and do all of it faster than NVIDIA keeps moving — lands on one conclusion: head-on competition is "close to impossible."

The Interdependence (sources: TSMC：纯代工模式发明者, Morris Chang 访谈：张忠谋亲述 TSMC, NVIDIA 之一：GPU 公司（1993-2006）, NVIDIA 之三：AI 时代的黎明（2022-2023）)

The single most important link in this chain isn't either company's moat on its own — it's that the two are irreplaceable customer and irreplaceable supplier to each other. Three independent sources tell the same origin story from three different angles: TSMC's San Jose office ignored this near-bankrupt small customer for years, so Jensen mailed a physical letter to TSMC's Hsinchu headquarters. Morris read it and called NVIDIA's office directly, walking straight into a room where the whole company was hand-testing freshly fabbed RIVA 128 chips — Jensen shouted over the chaos, "Everybody shut up. Morris Chang is on the phone." Morris's own account fills in the motive: he'd always told his sales staff "we should never be negligent in talking to future customers, even if the customer seems to be a very small one." Within two to three years, NVIDIA was a top-five TSMC customer.

The relationship survived its worst moment through personal trust rather than contract language. In 2009, a 40nm yield-and-delivery crisis left NVIDIA carrying the industry's biggest losses; the previous CEO's response had been to pay zero. Morris, newly returned as CEO, spent nearly half of his first four or five weeks back learning the facts cold, then took Jensen to his study after a family dinner and named a number on the spot: over $100M, good for 48 hours, no negotiation. Jensen accepted within two days, and the two companies have since done business worth, as one host put it, "many, many, many billions." That wasn't an isolated negotiation — it's a sample of two industry power-chain principals substituting personal trust for haggling at the moment it mattered most.

Structurally, that interdependence is baked into the industry's own division of labor. The shows sketch a five-layer value chain — IP/architecture, EDA, fabless design (NVIDIA's layer), equipment, and foundry (TSMC's layer) — and argue that Moore's Law simply moves too fast for any one company to stay best-in-class at every layer at once, forcing the industry from vertical integration into horizontal specialization. That's how NVIDIA ended up with a software company's capital structure: roughly $1B a year in capex against TSMC's $30B, and Apple, Microsoft, and Google's own tens of billions — prompting one host's needling aside, "Thank you, Morris," a nod to the man who invented the fabless split in the first place. In the other direction, CoWoS advanced packaging for AI training chips is the fastest-growing slice of TSMC's capacity, and NVIDIA has locked up an outsized share of it. Each company is a load-bearing input into the other's growth curve: TSMC's process leadership is the physical precondition NVIDIA's chips cannot do without; NVIDIA's demand fills TSMC's advanced packaging lines and helps amortize the next fab generation's enormous capex.

Geographic and Geopolitical Concentration (sources: TSMC：纯代工模式发明者, NVIDIA 之三：AI 时代的黎明（2022-2023）)

The physical base of this entire chain sits on one island. In the 2021-recorded TSMC：纯代工模式发明者, the chip shortage had already forced Ford to halt F-150 production — a live example of a single point of failure propagating straight into the real economy. The hosts don't dodge the "China takes Taiwan" bear case: lose TSMC and "imagine we can never get leading-edge semiconductors again — that's everything." Their thought experiment on whether "process power can be airlifted" — evacuate the people, ship out ASML's tools, does process power come with them? — gets a candid "I don't know" and no more. Arizona gets explicitly framed as built for "customer and government reasons," not economic optimality: Morris has said publicly there's no commercial logic to putting leading-edge capacity anywhere outside Taiwan, because the ecosystem only exists there. Geographic diversification, as of this material, is not a real hedge yet.

The single points of failure nest inside each other rather than sitting in one layer. Inside the island's concentration is a second bottleneck: CoWoS advanced packaging is only 10–15% of TSMC's total capacity, new capacity requires years of new fab construction, and NVIDIA has locked up a disproportionate share of it — a secondary chokepoint inside the primary one. One layer further down, the high-bandwidth rack-to-rack interconnect that large-model training depends on runs through a single surviving supplier, Mellanox. In September 2022, US export controls on advanced computing to China produced the downgraded A800/H800 SKUs — "even the cut-down version is still the best hardware available in China" — and by the September 2023 recording, China made up roughly 25% of NVIDIA's prior-year revenue, which is at once evidence that this chain can be wielded as geopolitical leverage and evidence of NVIDIA's own exposure to policy that can tighten without warning. None of the challengers that surface in the source material amount to an actual crack yet: Google TPU is sold only through GCP, not as an industry-standard product, and the show frames it as an "Android strategy for the data center" counter-position rather than a head-on threat; wafer-scale and novel-architecture startups like Cerebras and Graphcore are explicitly judged as "not there yet, but worth watching"; AMD is the one competitor both hosts credit as real, and it still lacks both the CUDA developer ecosystem and reserved TSMC advanced-packaging capacity. TSMC's own self-assessment on tape might be the most honest footnote to the whole chain: "It's everything but bulletproof."

Shared Mechanisms

Learning curve, scale, and ecosystem lock-in don't stay contained inside TSMC or NVIDIA individually — they nest along the length of this chain and amplify each other:

  • Learning-curve pricing was first worked out inside TI by Morris Chang and BCG, then carried whole into TSMC's founding logic: low prices buy volume, volume buys learning, learning buys yield, yield buys cost advantage. That curve is the ignition switch for TSMC's own profit-to-capex-to-process-leadership flywheel. It explains who survives and how fast — turning survival into monopoly is the job of the next layer.
  • Scale economies show up in opposite forms on the two ends of the chain, and feed each other. On TSMC's side it's the exponential capital wall created by Moore's Law stacked on Rock's Law — fab costs rising exponentially, pricing competitors out, with three-year capex running into nine figures of billions. On NVIDIA's side it's software-platform scale economics — CUDA's fixed investment amortized across millions of developers, at the cost of a decade and ten thousand person-years sunk. The chemistry between them: because NVIDIA is fabless, it never had to carry manufacturing's capex burden in the first place, which is exactly what freed up the budget to fight a second scale war — CUDA — while NVIDIA's own demand (especially for CoWoS packaging) fills and amortizes TSMC's capital wall in return. One company's scale economics transfuses the other's.
  • Ecosystem lock-in holds independently on each end, then chains end to end. On TSMC's side: "you can't switch off TSMC" — moving to GlobalFoundries isn't relocating an order, it's years of decoupling engineering. On NVIDIA's side: CUDA converts every developer's sunk investment automatically into loyalty to NVIDIA hardware. These two locks aren't parallel — they're in series. An AI developer locked into CUDA is locked into hardware that only runs on NVIDIA silicon, and NVIDIA silicon can only be manufactured on TSMC's leading-edge process with CoWoS packaging. Choosing CUDA is, several steps removed, also a vote for how TSMC's capacity gets allocated.

The compound result is a chain-level power structure that's harder to breach than any single company's moat: an attacker would need to clear the learning curve's years, the capital wall of scale economics, and the developer- and customer-side inertia of two separate ecosystem locks, all at the same time. The nine-step thought experiment in NVIDIA 之三：AI 时代的黎明（2022-2023） is, at bottom, a step-by-step test of exactly how hard that simultaneous breach would be.

Tensions and Fragility

This chain's strength and its fragility come from the same fact: extreme concentration.

  • Taiwan Strait concentration is the central, most openly acknowledged risk in the material. The hosts offer no reassurance — "can process power be airlifted" stops at "I don't know" — and the cornered resource that Taiwan's government built by funding half of TSMC's founding capital and personally arm-twisting local industrialists for the rest has, in one host's own reframing, flipped into an "anti-power": the very location that was once the asset is now the liability. This genuinely intersects with concepts discussed in the PH repository — Fortress America, the technate — where Taiwan's chip geography is a central entity in exactly this "where is technology concentrated, and who controls it" narrative; this page reports only the business-side analysis the hosts themselves give in the Acquired material, without importing PH-repo conclusions.
  • The single points of failure nest inside each other rather than sitting at one layer. Inside the island's concentration sits CoWoS packaging, only 10–15% of TSMC's total capacity and heavily locked up by NVIDIA; inside that sits the rack-to-rack interconnect for large-model training, which runs through one surviving company, Mellanox. A failure at any layer can propagate straight up the chain.
  • Export controls cut both ways. They're a geopolitical lever the US can use to precisely throttle this chain's output, and simultaneously a source of exposure for NVIDIA to Chinese market policy that can tighten without warning. "Even the cut-down version is still the best hardware available in China" quantifies both NVIDIA's lead and the practical limits of the controls at once — both readings hold, and they constrain each other.
  • The challengers on record don't amount to a real crack yet, but deserve honest documentation. Google TPU, sold only through GCP rather than retailed, is framed as a counter-position rather than a frontal challenge; wafer-scale and novel-architecture startups like Cerebras and Graphcore are explicitly "not there yet, but worth watching"; AMD is the one competitor both hosts credit as genuine, yet it still lacks the CUDA ecosystem and reserved TSMC advanced-packaging capacity. These are candidate cracks that genuinely appear in the source material — not confirmed ones.
  • Power concentrated in a handful of individuals is a distinct layer of fragility, separate from the structural kind. According to the January 2025 interview, Morris Chang, at 93, is still personally reconstructing decisions he made more than four decades ago; according to the September 2023 episode, Jensen Huang is explicitly described as having "no succession bench — the company is an extension of Jensen's mind, will, drive, and belief in the future." Both power chains run through founders who are either very old or effectively irreplaceable, and that personal risk is a different kind of thing from the structural risks above it, but no less real.
  • Even TSMC's own on-tape self-assessment leaves room for doubt: "It's everything but bulletproof." This chain looks unbreakable today, and the source material itself stops short of claiming otherwise.

This page draws on TSMC：纯代工模式发明者, Morris Chang 访谈：张忠谋亲述 TSMC, NVIDIA 之一：GPU 公司（1993-2006）, NVIDIA 之二：机器学习公司（2006-2022）, and NVIDIA 之三：AI 时代的黎明（2022-2023）. On the people side, see Morris Chang（张忠谋） (architect of manufacturing-side power) and Jensen Huang（黄仁勋） (architect of design- and ecosystem-side power). On the concepts side, see 学习曲线（Learning Curve） (the origin point of TSMC's pricing and flywheel), 7 Powers 护城河框架 (the moat framework applied throughout the chain), and Counter-Positioning（反向定位） (the shared strategic structure behind both the CUDA bet and TSMC's founding). Chinese version at 半导体权力链.