Semiconductors
How the most complex objects humans build actually work, who controls them, and why a handful of chokepoints hold up the entire economy.
A modern chip is the most complex object humanity mass-produces, and almost all of the best ones are made in one place. Understanding chips means understanding a supply chain built on a few irreplaceable chokepoints.
What this argues
- Chips are the new oilThe input everything else depends on, and just as geopolitically fraught.
- The physics is near its limitBut engineering keeps finding ways around Moore's Law's slowdown.
- Value sits at the chokepointsASML, TSMC and the design-tool makers hold irreplaceable positions.
- Concentration is the real riskOne earthquake or blockade in Taiwan would halt the digital economy.
- Everyone is now reshoringHundreds of billions in subsidies to rebuild what was offshored.
What has changed
Updated November 2026
The chip market is on track to be more than double its 2025 size
- WSTS (Aug 2026): the market reached $702B in the first half of 2026, up 102% year over year. The full-year 2026 forecast was raised to about $1,655B (about +108%), and 2027 to about $2.1T (about +29%).
- Memory is the driver: up 305% in the first half, with a 2026 forecast of +302%. Logic was up 45% in the first half.
- WSTS (Autumn 2025 forecast) had put 2025 at $772B (+22%) and 2026 at $975B, so the forecast has been lifted sharply in under a year.
TSMC posted record results, with 2 nm already a visible share of revenue
TSMC (Q3 2026 results): revenue NT$1,494.24B (US$46.72B), net income NT$763.06B, gross margin 66.4%, operating margin 58.7%. By wafer revenue, 3 nm was 33%, 5 nm 28%, 7 nm 10% and 2 nm 8%, so 7 nm and below made up 79%. Q4 guidance is US$52.3-54.3B, with management citing the steep 2 nm ramp.
ASML is guiding to a record year
ASML (Q3 2026 release, 14 Oct 2026): net sales EUR 11.7B, gross margin 57.5%, net income EUR 5.0B, 113 new lithography systems sold. Guidance for full-year 2026 is EUR 44-45B in sales at about 56% gross margin, and the CEO said some customers want to place orders for 2029 and 2030.
US export controls moved in both directions
Reports citing BIS: a January 2026 rule moved some advanced-computing exports to China to case-by-case license review, then guidance on 31 May 2026 required a license for advanced AI chips sold to any firm whose ultimate parent is in China, wherever the firm operates. Applied Materials was fined US$252M on 12 Feb 2026 over ion implantation equipment exports to China. Coverage came partly from secondary outlets; exact rule text should be checked in the Federal Register.
Enforcement is a separate strand: Culper Research, a short seller that posted on 13 May 2026 that it is short Nvidia, says over 20% of Nvidia's FY2026 compute revenue stayed driven by China through Southeast Asian intermediaries (Culper Research, 13 May 2026, report not opened; Benzinga, May 2026). Nvidia has said diversion claims about one customer were investigated and debunked, I found no reply to Culper by name, and I found no ruling by a court or regulator (The Hill, 29 Oct 2026). The Nvidia page (/investing/nvidia) sets out both sides.
Intel's foundry effort is advancing, but unproven at scale
TrendForce and Intel (May-Aug 2026): Intel's 18A process is in production, 18A-P entered risk production for foundry customers, and Socionext added 18A-P to its foundry mix alongside TSMC. Reports of a possible Apple deal were unconfirmed, and TSMC's share of the leading edge has not yet moved.
Not refreshed: foundry revenue share (2025 estimates), advanced-node share by country, US manufacturing share, Nvidia market capitalisation and latest results, CHIPS Act disbursement status, High-NA pricing and the transistor-count figure. Old values are left as written.
What a Chip Is
From sand to switches
A chip is just billions of microscopic switches (transistors) etched onto a sliver of silicon
Each switch is either on or off: a 1 or a 0. Everything digital is built from that.
The numbers stop making intuitive sense
Four families do very different jobs
| Family | Job |
|---|---|
| Logic | the 'brains': CPUs and GPUs that compute. The most advanced and valuable. |
| Memory | the 'notepad': DRAM and flash that store data. A brutal commodity cycle. |
| Analog | bridges the real world: power, sensors, radio. Cheap, essential, sticky. |
| Specialized | AI accelerators, image sensors, microcontrollers: the fastest-growing niches. |
Why AI rewired the entire chip market
| CPU, the generalist | GPU, the specialist | |
|---|---|---|
| Cores | a few, very fast | thousands, in parallel |
| Job | one task at a time, quickly | the same math, all at once |
| Analogy | a brilliant lone worker | an army of workers |
| For AI | too slow | the unlock |
How Chips Are Made
Printing at the atomic scale
Making a chip means printing patterns finer than a wavelength of light, hundreds of times, without a single fatal error
A leading-edge chip takes ~3 months and 1,000+ steps to build.
Chipmaking, radically simplified
- Design: engineers lay out billions of transistors as a blueprint.
- Mask: the design becomes a stencil for each layer.
- Lithography: light prints the pattern onto a silicon wafer.
- Etch and repeat: layers are built up 100+ times, then diced into chips.
A leading-edge fab is a wonder, and a fortune
Why 'yield' is the real trade secret
- Yield = good chips per wafer. A brand-new process can start with most chips defective.
- A single speck ruins a chip. Which is why fabs are the cleanest places on Earth.
- Yield climbs with experience. Learning curves slowly turn losses into profit.
- Know-how, not just machines. TSMC's edge is as much operational skill as equipment.
DUV vs EUV lithography
| DUV, the workhorse | EUV, the frontier | |
|---|---|---|
| Light | 193 nm (deep ultraviolet) | 13.5 nm (extreme ultraviolet) |
| Nodes | mature and mid-range chips | every advanced chip, 7 nm and below |
| Maker | ASML, Nikon, Canon | ASML, and only ASML |
| Price | ~$5-80M per system | ~$180M, rising to $380M+ |
The Value Chain
Who does what
No single company makes a chip alone
| Layer | Who |
|---|---|
| 05 Design tools (EDA) | Synopsys, Cadence: the software to design chips. |
| 04 IP and architecture | Arm: the blueprints many chips are built on. |
| 03 Fabless design | NVIDIA, Apple, AMD, Qualcomm: design, don't manufacture. |
| 02 Foundries | TSMC, Samsung: the factories that make the chips. |
| 01 Equipment and materials | ASML, Applied Materials: the tools and inputs. |
The market splits across chip types
View the data
| Market size | |
|---|---|
| Logic | 220 |
| Memory | 170 |
| Analog | 90 |
| Microprocessors | 80 |
| Discrete & sensors | 67 |
Source: WSTS; segment estimates (2024)
A few firms hold irreplaceable positions
| Firm | Position |
|---|---|
| TSMC | makes ~66% of outsourced chips, and >90% of the most advanced ones. |
| ASML | the only maker of EUV machines. No ASML, no advanced chips. |
| NVIDIA | designs the AI accelerators the whole boom runs on. |
| Synopsys / Cadence | their software is used to design essentially every chip. |
| Arm | its architecture sits in nearly every phone on Earth. |
| Applied Materials | the deposition and etch tools every fab needs. |
The biggest buyers are becoming chip designers
| Company | Move |
|---|---|
| Apple | its own phone and Mac silicon, a decade ahead of rivals. |
| TPUs built specifically to train and run AI. | |
| Amazon | Graviton and Trainium chips for its own cloud. |
| Tesla | custom silicon for self-driving. |
| Microsoft & Meta | designing in-house AI accelerators. |
| The logic | own your destiny, cut cost, tune to your workload. |
Who holds leverage in the chain
A two-axis map of market share (horizontal) against ease of replacement, inverted (vertical). ASML, TSMC and EDA tools sit in the top right, the 'Chokepoints' corner (high share, almost impossible to replace). NVIDIA sits just below them, still on the high-share side. Memory makers and assembly sit lower left, toward 'Commodity'. The corners are labeled Niche (top left), Chokepoints (top right), Commodity (bottom left) and Contested (bottom right).
The Chokepoints
A handful of bottlenecks
The entire digital world rests on a few chokepoints: points where one firm, or one country, controls a step no one can skip
This is what makes chips the most strategically contested industry on Earth.
Concentration increases toward the top
| Layer | Concentration |
|---|---|
| EUV lithography | ASML: 100%, no alternative exists |
| Leading-edge foundry | TSMC: >90% of sub-5 nm chips |
| Design software | Synopsys & Cadence: near-duopoly |
| Chip design & memory | more competitive, but still concentrated |
One company makes most of the world's chips
View the data
| Share | |
|---|---|
| TSMC | 66 |
| Samsung | 8 |
| SMIC | 6 |
| Intel | 5 |
| GlobalFoundries | 5 |
| Others | 10 |
Source: Foundry revenue share, 2025 estimates
Moore's Law
Is the magic curve ending?
For 50 years, transistor counts doubled roughly every two years
A log-scale chart of transistors per chip, rising from roughly 1,000 in 1971 (the first commercial chip had 2,300) through 1985, 2000 and 2012 to 208 billion in 2024.
Moore's Law isn't a law of physics: it's an economic promise that chips get twice as good every two years
That promise is getting harder and more expensive to keep.
Chips still improve, but the cost per transistor stopped falling
A log-scale chart of cost per transistor (falling) from 1970 to 2024 shows a steep decline through the 1970s to 2010, then a flat line: cost stops falling at the leading edge.
Each slowdown met with a new invention
- FinFET: 3-D transistors extend scaling
- EUV: new light source keeps shrinking
- Chiplets: stitch small dies into big ones
- Gate-all-around: the 2 nm transistor (8% of TSMC wafer revenue by Q3 2026)
- Backside power: squeeze out more efficiency
Is Moore's Law dead?
| 'It's over' | 'It continues' | |
|---|---|---|
| 1 | Cost: new nodes cost more per transistor | Packaging: chiplets & 3-D stacking |
| 2 | Physics: atoms don't get smaller | Specialization: custom silicon per workload |
| 3 | Slowing: gains now years, not months | Materials: new transistor designs |
| 4 | Fewer players: only 3 firms can afford it | Systems: optimize the whole stack |
Progress moves from shrinking to stacking
Geopolitics
Chips are the new oil
Whoever controls advanced chips controls AI, weapons and the economy, so chips have become an instrument of national power
The result: export controls, subsidy wars and a global scramble to reshore.
Advanced-chip manufacturing is dangerously concentrated
View the data
| Share | |
|---|---|
| Taiwan | 90 |
| South Korea | 8 |
| China | 1 |
| United States | 1 |
| Japan | 0 |
| Europe | 0 |
Source: Industry estimates; advanced-node share (2024)
America went from making most chips to making few
View the data
| US share | |
|---|---|
| 1990 | 37 |
| 2000 | 28 |
| 2010 | 19 |
| 2015 | 14 |
| 2020 | 12 |
| 2024 | 10 |
Source: SIA; industry estimates · values read from the chart, approximate
Governments are spending to rebuild capacity
| Measure | What it does |
|---|---|
| US CHIPS Act | $52B in subsidies to bring fabs onshore. |
| EU Chips Act | EUR 43B to double Europe's share by 2030. |
| Export controls | the US restricts advanced chips and tools to China. |
| Reshoring | TSMC, Samsung and Intel building new US and EU fabs. |
| China's push | hundreds of billions to build a domestic industry. |
| Allied 'friend-shoring' | supply chains rerouted around rivals. |
The risk that overhangs everything
| The risk | The mitigation | |
|---|---|---|
| 1 | Concentration: ~90% of advanced chips, one island | 'Silicon shield': chips deter conflict, both sides need them |
| 2 | Tension: China claims Taiwan | Reshoring: new fabs in US, Japan, Germany |
| 3 | No substitute: years to rebuild elsewhere | Stockpiling: buyers holding buffers |
| 4 | Global stakes: a blockade would freeze tech | Slow diversification: but the edge stays in Taiwan |
What to Watch
Signposts, not predictions
Four dials worth watching
Five signposts for the next 24 months
- Does 2 nm ramp on schedule? The health of the leading edge in one number.
- Do reshored fabs actually work? Subsidies only matter if the chips ship.
- How far do export controls go? Each escalation redraws the supply chain.
- Does AI demand hold? The current engine of the whole industry.
- Does anyone challenge ASML or TSMC? The chokepoints are the whole game.
Reference
Glossary and sources
Plain-language glossary
| Term | Meaning |
|---|---|
| Transistor | a microscopic on/off switch, the atom of computing. |
| Fab | a chip factory (fabrication plant), costing $20B+. |
| Fabless | a firm that designs chips but owns no fab (e.g. NVIDIA). |
| EUV | extreme-ultraviolet lithography, ASML's frontier machine. |
| Node (e.g. 2 nm) | a generation of chipmaking; smaller is roughly better. |
| Foundry | a fab that makes chips for others (e.g. TSMC). |
| Lithography | using light to print circuit patterns on silicon. |
| Moore's Law | the ~2-year doubling of transistors per chip. |
The most complex objects we build are made by a handful of firms, on one island, with one company's machines. Efficiency built that chain; security is now trying to unbuild it.
The five conclusions to take away
- Chips are foundational: the input beneath every other technology.
- Complexity is the moat: only a few firms can operate at the edge.
- Chokepoints hold the power: ASML, TSMC and EDA are irreplaceable.
- Concentration is the risk: Taiwan is the economy's single point of failure.
- Moore's Law bends, not breaks: progress shifts from shrinking to system design.
Sources and further reading
- WSTS: Q2 2026 release (Aug 2026) and Autumn 2025 forecast; TSMC and ASML Q3 2026 results.
- WSTS: World Semiconductor Trade Statistics market forecasts (2024-25).
- Grand View Research: Semiconductor Market Size & Outlook to 2030.
- TSMC: quarterly results and technology disclosures (2024-25).
- ASML: annual report, EUV/High-NA system pricing and roadmap.
- Semiconductor Industry Association (SIA): manufacturing-share data.
- Company filings: NVIDIA, Synopsys, Applied Materials, Arm.
- Industry analyses: foundry and advanced-node market share estimates (2025).
- Figures marked 'estimate' or 'illustrative' are compilations.
Sources cited
WSTS · Company filings · Industry estimates · SIA · Grand View Research · TSMC · ASML · Manufacturer data · WSTS, Aug 2026 (Q2 2026 release) · WSTS, Autumn 2025 forecast · TSMC, Q3 2026 results, Oct 2026 · ASML, Q3 2026 results, 14 Oct 2026 · BIS export guidance, May 2026 (via press reports) · TrendForce, 2026 (Intel foundry) · Culper Research, 'NVIDIA (NVDA): The China Problem', 13 May 2026: report not opened (HTTP 403); Culper's own X thread of 13 May 2026, as unrolled on Thread Reader App, is used instead · Benzinga, 'Tip Of The Iceberg': Culper Research Targets Nvidia's Alleged China Problem, May 2026 (Yahoo Finance copy dated 18 May 2026) · The Hill, Nvidia spokesperson on Megaspeed diversion allegations (article on a New York Times report), 29 Oct 2026
Keep reading
Research, not investment advice. Figures as of , from the sources cited.
Legal information and disclosures
These are my own views as of the update shown on this page. They can change without notice, and I have no duty to update them. Past performance does not indicate future results, and any investment can lose money.
This is for education only. It is not investment, legal, tax, medical or accounting advice, and it is not an offer to sell or a solicitation to buy any security in any jurisdiction. It does not consider your goals or circumstances.
Figures from company filings, public agencies and other third parties are believed to be reliable, but I have not independently verified them and cannot guarantee them. Forecasts and anything marked as an estimate rest on assumptions that may prove wrong.
Please do not copy or republish this material without my written consent.