> ## Content Index
> Fetch the complete content index at: https://www.sotkn.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# What Is Fabless? How Chip Companies Like Nvidia Sell Chips They Never Make
- URL: https://www.sotkn.com/explained/fabless-explained/
- Published: 2026-11-16T10:28:37.000Z
- Updated: 2026-11-16T10:28:37.000Z
- Description: A fabless company designs chips but does not own a factory; a foundry such as TSMC makes them. Nvidia, Broadcom and AMD all work this way. The top 10 fabless firms took in US$141bn in the second quarter of 2026, 2.6 times the top 10 foundries. The catch: they have to book factory space years ahead.
- Author: Sean
- Tags: Mechanism, #desk-explained

EXPLAINED · Mechanism · Last reviewed: 16 Nov 2026 · Next review: after TrendForce’s third-quarter 2026 fabless and foundry tables, or TSMC’s fourth-quarter results

A fabless company is a chip company that designs semiconductors but does not own a fabrication plant, or fab. It sends its designs to a foundry, such as TSMC, which makes the chips for a fee, and then sells the finished chips under its own name.

**In short.** A fabless company owns the chip design and the customer, and buys almost everything physical from someone else: wafers from a foundry, memory from a memory maker, packaging and testing from specialists. It is now the bigger half of the chip business by revenue. The 10 largest fabless companies took in about US$141 billion in the second quarter of 2026, by TrendForce’s figures, 2.6 times the US$53 billion taken by the 10 largest foundries. Nvidia alone took about US$91 billion, 2.3 times TSMC’s US$40 billion. Our read: the stuck step is booked factory space. TSMC makes 72.5% of the world’s foundry revenue, plans capacity years ahead and does not let customers sublet their slots, so a fabless company that wants more chips has to commit early, and carry the risk if demand turns.

Why an investor cares: most of the money in AI chips is booked by companies that own no factory. Their profits depend on how much factory space, memory and packaging they can secure, and at what price, from suppliers that are themselves short of capacity. Fabless is the model behind Nvidia, Broadcom and AMD, and it sits at the centre of our [Semiconductors](https://www.sotkn.com/stack/semiconductors/) layer.

## What does fabless mean?

Fabless means without a fab. A fab, short for fabrication plant, is the factory where chips are printed onto silicon wafers. A fabless company does the design, the software and the selling, and pays a foundry to do the manufacturing.

The model splits the chip business into two jobs. One company decides what the chip does and finds customers for it. Another company owns the clean rooms and the machines and runs the wafers through them. The fabless company pays the foundry for each wafer and keeps whatever it can sell the finished chips for above that cost.

There is a third way to run a chip company: do both. An integrated device manufacturer, or IDM, designs its chips and makes them in its own fabs. Intel has long worked this way, and so do the memory makers. Our [who makes NAND](https://www.sotkn.com/explained/who-makes-nand/) page, published this morning, covers six companies that all own their memory fabs. The table sets out the three models side by side.

|                    | Fabless                                                               | Foundry                                                           | IDM (integrated device manufacturer)                          |
| ------------------ | --------------------------------------------------------------------- | ----------------------------------------------------------------- | ------------------------------------------------------------- |
| **Designs chips?** | Yes                                                                   | No, makes other companies’ designs                                | Yes                                                           |
| **Owns fabs?**     | No                                                                    | Yes                                                               | Yes                                                           |
| **Sells**          | Finished chips, boards and systems under its own name                 | Processed wafers, priced per wafer                                | Finished chips under its own name                             |
| **Examples**       | Nvidia, Broadcom, AMD, Qualcomm, MediaTek, Marvell                    | TSMC, UMC, GlobalFoundries, SMIC                                  | Intel, Micron, SK hynix; Samsung is both an IDM and a foundry |
| **Big cost**       | Engineers and design tools, then wafers, memory and packaging it buys | Fabs and tools: TSMC plans US$60–64bn of capital spending in 2026 | Both                                                          |
| **Main risk**      | Committing to supply before demand is known                           | Building capacity that customers later do not fill                | Both, on its own balance sheet                                |

Examples from TrendForce’s second-quarter 2026 tables and company filings listed under References. Not a rating of companies or investments.

## How does the fabless model work, step by step?

A fabless chip passes through five stages, and the fabless company owns only the first one outright.

- **Design.** Engineers describe the chip in code, check it in simulation and lay it out with specialist design software. Many chips also use licensed building blocks, such as processor cores or interface circuits, bought from other firms.
- **Tape-out and masks.** When the design is final, it is sent off to be turned into a set of [photomasks](https://www.sotkn.com/explained/photomask-explained/), the stencils the factory uses to print each layer. The name comes from the days when designs were shipped on magnetic tape.
- **Wafers at the foundry.** The foundry prints the chips on [silicon wafers](https://www.sotkn.com/explained/silicon-wafer-explained/) over hundreds of steps and several months. The share of chips on each wafer that work, the [yield](https://www.sotkn.com/explained/semiconductor-yield-explained/), decides how many sellable chips each wafer gives.
- **Packaging and memory.** The chips are cut from the wafer and packaged. For AI chips this means joining the processor to stacks of high-bandwidth memory on an [interposer](https://www.sotkn.com/explained/interposer-explained/), a step known at TSMC as [CoWoS](https://www.sotkn.com/explained/cowos-explained/).
- **Test and ship.** Finished parts are tested, then shipped to server makers and cloud companies. Our [who tests AI chips](https://www.sotkn.com/explained/who-tests-ai-chips/) page covers this step.

![How a fabless chip gets made: design, tape-out and masks, wafers at the foundry as the stuck step, packaging and memory, test and ship](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/11/fabless-flow.png)

How a fabless chip gets made. Second Order.

The fabless company manages all five stages and is responsible to its customer for the result. But it owns only the first. Every later step is capacity it has to book from a supplier, and each one has its own lead time.

## Who are the biggest fabless companies?

By revenue, the fabless group is now dominated by AI chips. TrendForce’s figures for the second quarter of 2026 put the combined revenue of the 10 largest fabless companies at US$141.45 billion, up 73% from a year earlier.

Nvidia took about US$91.16 billion of that, or 64%. Broadcom, at US$18.89 billion, and AMD, at US$11.53 billion, followed. By our arithmetic, the top three together took about US$121.6 billion, or 86% of the top 10’s revenue. Nvidia alone took 4.8 times as much as Broadcom. TrendForce’s numbers are calendar-quarter estimates; Nvidia’s own quarter to 26 July 2026 shows revenue of US$96.2 billion.

| Company             | Model                 | 2Q26 revenue, US$bn | What it mostly sells                                                                                               |
| ------------------- | --------------------- | ------------------- | ------------------------------------------------------------------------------------------------------------------ |
| **Nvidia**          | Fabless               | 91.16               | AI GPUs, networking and full server systems                                                                        |
| **TSMC**            | Foundry               | 40.2                | Wafers for fabless and other chip companies; 68% of its third-quarter revenue came from high-performance computing |
| **Broadcom**        | Fabless               | 18.89               | Custom AI chips designed with customers, and networking chips                                                      |
| **AMD**             | Fabless               | 11.53               | Server and PC processors, AI GPUs                                                                                  |
| **Qualcomm**        | Fabless               | 8.5                 | Phone and car chips                                                                                                |
| **MediaTek**        | Fabless               | 4.81                | Phone chips, and custom AI chips for data centres                                                                  |
| **Samsung Foundry** | Foundry arm of an IDM | 3.26                | Wafers for outside customers, including HBM base dies                                                              |

TrendForce estimates by calendar quarter, published 9 and 11 September 2026\. Revenue is total company revenue in the quarter, not only chips. Listed by revenue, not as a judgement of quality or value.

Broadcom shows a second fabless model: designing chips for someone else. TrendForce says the first chip Broadcom helped design for OpenAI and Google’s TPU 8i began shipping in the quarter, and that it has six major custom chip customers. In that case the cloud company or AI lab sets the goals, Broadcom does much of the design and often arranges manufacturing, and a foundry makes the wafers. Our [who makes custom AI chips](https://www.sotkn.com/explained/who-makes-custom-ai-chips/) page covers this group, and [TPU Explained](https://www.sotkn.com/explained/tpu-explained/) covers Google’s chip.

The rest of the group still lives mostly on phones, PCs and cars. Qualcomm took over US$8.5 billion and MediaTek about US$4.81 billion, both held back by phones, TrendForce says, though MediaTek expects more than US$2 billion of data centre revenue in 2026.

## Fabless vs foundry: who earns more?

The designers now out-earn the factories by a wide margin. The 10 largest foundries took in about US$53.49 billion in the second quarter of 2026, by TrendForce’s count. The 10 largest fabless companies took in 2.6 times as much, by our arithmetic.

![Bar chart of second-quarter 2026 revenue in US dollars: Nvidia 91.2 billion, TSMC 40.2 billion, Broadcom 18.9 billion, AMD 11.5 billion, Qualcomm 8.5 billion, Samsung Foundry 3.3 billion](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/11/fabless-revenue-chart.png)

Second-quarter 2026 revenue of the largest fabless companies and foundries, in US$ billion. TrendForce estimates, published September 2026; chart by Second Order.

The foundry side is far more concentrated. TSMC took about US$40.2 billion, a 72.5% share of foundry revenue, more than 12 times Samsung Foundry’s 5.9%. SMIC had 5.4%, UMC 3.9% and GlobalFoundries 3.2%. The top five together had 90.9%, by our arithmetic. Our read: in practice, almost every fabless company making leading-edge chips is a TSMC customer. Our [what TSMC does](https://www.sotkn.com/stack/semiconductors/cowos/what-does-tsmc-do/) page covers the company.

Margins tell the same story from another angle. Nvidia’s gross margin was 75.0% in its quarter to 26 July 2026\. TSMC’s was 67.7% in its April–June quarter, 7.3 percentage points lower, and 66.4% in the quarter to September. Both are high. But the fabless company keeps a larger share of each dollar of its revenue than the foundry does, because it owns the design the customer is buying.

Nvidia’s cost of revenue was US$24.1 billion against revenue of US$96.2 billion, about 25%, by our arithmetic. That figure covers everything Nvidia buys to make what it sells: wafers, memory, packaging, boards, and the other parts of complete server systems, not only TSMC’s share.

## How does a fabless company pay for its chips?

By the wafer, booked in advance. A foundry charges for each wafer it processes, and the price rises with how advanced the process is. TSMC does not publish prices, but its results give a rough average.

In the third quarter of 2026 TSMC shipped 4.476 million wafers, counted as 12-inch equivalents, and took NT$1,494.24 billion in revenue. That works out to about NT$334,000, or about US$10,400, per wafer, by our arithmetic. A year earlier the same sum gave about US$8,100\. The average is crude: it mixes old and new nodes and includes packaging and other revenue. It is still the clearest public sign that each wafer now earns more. The newest nodes carry the most: 2nm and 3nm together made up 41% of TSMC’s wafer revenue in the quarter. Our [2nm chip Explained](https://www.sotkn.com/explained/2nm-chip-explained/) covers the newest node.

The bigger cost is the commitment. To be sure of supply, a fabless company signs agreements to buy wafers, memory and packaging years ahead. Nvidia’s supply and capacity commitments rose from US$119 billion to US$279 billion in a single quarter, as of 26 July 2026, mainly for memory, the company said. By our arithmetic that is 2.3 times the previous quarter’s figure and 2.9 times the revenue of the quarter. About US$267 billion of it, 96%, falls due by the end of Nvidia’s 2029 fiscal year, which ends in January 2029.

Nvidia’s inventory rose from US$25.8 billion to US$31.6 billion in the same quarter, up 22%, as it prepared for its next chip generation. Its 10-Q says it is experiencing certain supply constraints and has entered significant commitments to secure inventory and capacity for the next several years. Our [purchase commitment Explained](https://www.sotkn.com/explained/purchase-commitment-explained/) and [prepayment Explained](https://www.sotkn.com/explained/prepayment-explained/) cover how these promises work.

TSMC sets the terms of the booking. On its 15 October call, TSMC’s chairman said the company plans capacity years ahead with its customers, and that it does not allow customers to sublet or auction off their capacity. So a fabless company that books too much cannot easily pass the surplus to another chip designer, and one that books too little cannot easily buy someone else’s slots.

## Why go fabless at all?

Because the factory is the expensive, slow and risky part. TSMC plans US$60 billion to US$64 billion of capital spending in 2026 and says it is building 13 leading-edge and advanced packaging fabs in Taiwan over the next several years, plus five or six overseas. Few chip designers could fund that, and fewer could keep such fabs full.

Going fabless turns a factory into a bill. The designer can switch its money to engineering, spread its orders across processes, and move to a newer node when the foundry offers one. The foundry, in turn, fills its fabs with orders from hundreds of designers, which keeps them busier than any one company could.

The trade-off is control. A fabless company cannot add capacity by building its own line, and it competes with other customers for the same slots at the same foundry. When demand runs ahead of supply, as it has for AI chips this year, that is the constraint that decides how much it can sell.

## What is the risk of being fabless?

The fabless company carries the demand risk between the booking and the sale. It commits to wafers and memory months or years before it ships the chip. If demand falls, or a government blocks sales, it is left with stock it cannot sell and purchase obligations it still has to meet.

Nvidia shows how large this can get. It took a US$4.5 billion charge in the first quarter of its 2026 fiscal year for excess inventory and purchase obligations tied to its H20 chip, after US export rules stopped sales to China. By our arithmetic that equals about 6% of the gross profit it earned in its latest quarter. It took a further US$0.4 billion charge for H200 in the first half of its 2027 fiscal year.

Concentration is the other risk. With TSMC making most leading-edge chips, many of the world’s fabless companies depend on one supplier, most of whose leading-edge fabs are in Taiwan. TSMC is adding fabs in Arizona, which it says its large US customers asked for, and in Japan.

## Who gets paid?

Our read: the fabless company is paid last and most, but only after everyone it books from has been paid or promised.

- **The foundry, per wafer.** High-performance computing, mostly AI and data centre chips, made up 68% of TSMC’s third-quarter revenue, about US$31.8 billion, by our arithmetic. Customers based in North America, where most large fabless companies have their headquarters, accounted for 79%, about US$36.9 billion.
- **Memory makers.** Nvidia says its new commitments are mainly for memory. High-bandwidth memory comes from IDMs that own their fabs; our [who makes DRAM](https://www.sotkn.com/explained/who-makes-dram/) page covers them.
- **Packaging and test specialists.** Advanced packaging and final test are bought in from the foundry itself or from outside assembly and test firms. Our [who makes advanced packaging](https://www.sotkn.com/explained/who-makes-advanced-packaging/) page covers them.
- **Tool makers, one step back.** Every new fab a foundry builds for its fabless customers is filled with machines. TSMC said on its October call that tool supply is tight and tool prices are rising. Our [who makes wafer fab equipment](https://www.sotkn.com/explained/who-makes-wafer-fab-equipment/) page covers this group.
- **The fabless company, on the margin.** What is left after wafers, memory, packaging and engineering is the designer’s. At a 75% gross margin, that is most of each dollar of revenue before operating costs.

## Where is the stuck step?

Our read: booked leading-edge capacity at the foundry, together with the packaging that follows it. A fabless company can finish a design in months, but it cannot conjure the wafers. TSMC said its 2nm capacity should more than double in 2027, a new 3nm fab in Tainan should start volume production in the first half of 2027 and another in the first half of 2028, and its second Arizona fab in the third quarter of 2027\. Its A14 node is due for risk production in 2027 and volume production in 2028\. Those dates are the ceiling on what its fabless customers can ship. Our [TSMC CoWoS stuck-step analysis](https://www.sotkn.com/analysis/tsmc-cowos-stuck-step-what-price-assumes/) looks at the packaging side.

The clock runs both ways. TSMC’s fourth-quarter 2026 guidance of US$52.3 billion to US$54.3 billion is, at the midpoint, 14% above the third quarter, by our arithmetic. It is a sign of how much its customers have already ordered. Our [Taiwan Markets door](https://www.sotkn.com/markets/taiwan/) tracks TSMC’s monthly revenue, the earliest public read on those orders.

## Fabless and Singapore

Singapore is a foundry location rather than a home for large fabless companies, and that is where it meets this model. On 28 September 2026, VisionPower Semiconductor Manufacturing Company, or VSMC, opened its first 12-inch fab in Tampines. VSMC is a joint venture of the Taiwanese foundry Vanguard International Semiconductor and the Dutch chipmaker NXP.

The fab makes chips on older processes, from 130nm to 40nm, for power management, analog and mixed-signal parts, and interposers, the silicon layer that links an AI processor to its memory. It has entered risk production, with volume production due in the first quarter of 2027 and about 44,000 wafers a month and 1,600 jobs by 2029, it says. By our arithmetic, 44,000 wafers a month is about 2.9% of TSMC’s average monthly shipments in the third quarter of 2026, a useful scale check rather than a like-for-like comparison.

For fabless companies, plants like this matter because the scarce part of an AI chip is not only the leading-edge die. Interposers and power chips come from older processes, and TrendForce says demand for AI-related power chips and other peripheral parts lifted orders at older-process foundries in the second quarter. Our [Wire on GlobalFoundries making interposers for TSMC](https://www.sotkn.com/wire/globalfoundries-tsmc-2b-silicon-interposer-wire/) covers the same shift, and our [Singapore Markets door](https://www.sotkn.com/markets/singapore/) follows the local names.

## What would change the view?

- **A second leading-edge foundry wins big AI orders.** If Samsung Foundry or Intel’s foundry business takes a large share of leading-edge AI wafers, fabless companies gain bargaining power and the stuck step loosens.
- **Fabless companies build fabs again.** On its October call, TSMC’s chairman said it had talked to a customer that is trying to build its own fab. If large designers start owning manufacturing, the line between fabless and IDM blurs.
- **Demand turns.** If AI spending slows, the commitments that secure supply today become the purchase obligations that cost money tomorrow, as the H20 charge showed.
- **Leadership and customer changes.** Nvidia’s chief executive resigned in November 2026\. Taiwan’s central bank governor said on 16 November that he expects only short-term turbulence for Taiwan’s AI suppliers, and that uncertainty will not ease until a new chief executive sets out the company’s plans. Fabless orders are the foundry’s revenue, so a change in a large designer’s plans moves the whole chain.

The clocks: TSMC’s monthly revenue, around the 10th of each month; TSMC’s fourth-quarter results in January 2027; TrendForce’s third-quarter fabless and foundry tables; Nvidia’s next quarterly filing and its commitments table; and VSMC’s volume production in the first quarter of 2027.

## Common misconceptions

- **People assume Nvidia makes its own chips.** Actually it designs them and pays others to make them: a foundry for the wafers, memory makers for the memory, and specialists for packaging and assembly.
- **People assume fabless means low-risk.** Actually the fabless company carries the demand risk. Nvidia’s supply and capacity commitments stood at US$279 billion in July 2026, and it took a US$4.5 billion charge when one chip could not be sold.
- **People assume the factory earns most of the money.** Actually the 10 largest fabless companies took in about 2.6 times as much as the 10 largest foundries in the second quarter of 2026.
- **People assume a fabless company can switch foundries easily.** Actually a design is built for one foundry’s process, and moving it means redesign and new masks. With TSMC at 72.5% of foundry revenue, the choice at the leading edge is narrow.
- **People assume fabless and foundry are rivals.** Actually they are two halves of one supply chain. The foundry’s revenue is the fabless company’s cost.

## FAQ

**What does fabless mean?**  
Fabless means a chip company designs its chips but does not own the factory that makes them. It pays a foundry, such as TSMC, to manufacture the chips, then sells them under its own brand. The word is short for “fabrication plant-less”.

**Is Nvidia a fabless company?**  
Yes. Nvidia designs its GPUs and networking chips and relies on outside foundries, chiefly TSMC, to make them, plus outside suppliers for memory, packaging and assembly. TrendForce puts Nvidia first among fabless companies, with about 64% of the top 10’s revenue in the second quarter of 2026.

**What is the difference between fabless and a foundry?**  
A fabless company designs chips and sells them; a foundry makes chips designed by others and is paid per wafer. Nvidia and Broadcom are fabless. TSMC, UMC and GlobalFoundries are foundries. They are two halves of one supply chain.

**What is the difference between fabless and IDM?**  
An IDM, or integrated device manufacturer, designs chips and makes them in its own fabs. Intel, Samsung and the memory makers work this way. A fabless company skips the factory and buys manufacturing from a foundry instead.

**Why do chip companies go fabless?**  
Because a leading-edge fab costs tens of billions of dollars and has to run nearly full to pay off. TSMC alone plans to spend US$60 to US$64 billion on capital projects in 2026\. A fabless company puts its money into design and leaves the factory risk to the foundry.

**What are the biggest fabless companies?**  
By TrendForce’s second-quarter 2026 figures: Nvidia, Broadcom, AMD, Qualcomm and MediaTek, followed by Marvell, Realtek, MPS, Novatek and OmniVision. The 10 together took in about US$141 billion in the quarter.

**What is the risk for a fabless company?**  
It has to commit to wafers, memory and packaging long before it knows final demand. If demand falls or an export rule changes, it is left with stock and purchase obligations. Nvidia took a US$4.5 billion charge in 2025 when its H20 chip could not be sold to China.

**Is this advice to buy or sell any company?**  
No. This page explains how the fabless model works and who it pays. It does not rate companies as investments or give buy or sell calls. Not investment advice. Do your own research.

Related: [who makes AI chips](https://www.sotkn.com/explained/who-makes-ai-chips/); [foundry and advanced packaging](https://www.sotkn.com/explained/foundry-advanced-packaging-explained/); [custom AI chips](https://www.sotkn.com/explained/who-makes-custom-ai-chips/); [semiconductor yield](https://www.sotkn.com/explained/semiconductor-yield-explained/); [who makes NAND](https://www.sotkn.com/explained/who-makes-nand/); [purchase commitments](https://www.sotkn.com/explained/purchase-commitment-explained/).

## Stack doors

Layer: [Semiconductors](https://www.sotkn.com/stack/semiconductors/). Players: [TSMC](https://www.sotkn.com/stack/semiconductors/cowos/what-does-tsmc-do/), [Samsung](https://www.sotkn.com/stack/semiconductors/hbm/what-does-samsung-do/) and [ASML](https://www.sotkn.com/stack/semiconductors/what-does-asml-do/). Market read: [TSMC’s third quarter and CoWoS](https://www.sotkn.com/markets/taiwan/taiwan-tsmc-q3-cowos-october-2026/).

[Get the Wire](https://www.sotkn.com/#/portal/signup) — short daily notes on where AI computing gets built and who pays for it.

Not investment advice. Do your own research.

**How we check this**

Fabless and foundry revenue and shares come from TrendForce’s second-quarter 2026 tables, published on 9 and 11 September 2026; these are a research firm’s calendar-quarter estimates, not company-reported figures. TSMC revenue, margins, wafer shipments and the split by node, platform and customer region come from its third-quarter 2026 management report of 15 October 2026, and its capacity plans and capacity rules from the earnings call the same day. Nvidia’s revenue, cost of revenue, gross margin, commitments and inventory come from its CFO commentary for the quarter to 26 July 2026, and its supply risks and inventory charges from its Form 10-Q filed on 26 August 2026\. The Singapore fab figures come from the VSMC release published by Singapore’s Economic Development Board on 28 September 2026\. The note on Nvidia’s leadership change comes from Focus Taiwan’s report of 16 November 2026\. Revenue per wafer is our rough arithmetic: TSMC’s total revenue divided by its 12-inch-equivalent wafer shipments, so it includes packaging and other revenue and is not a price for any one wafer. Ratios, shares, sums and conversions are our arithmetic, checked in code and labelled as ours on the page. Who gets paid and the stuck step are our read. The featured photograph is credited in its caption. This page explains how the fabless model works and who it pays. It does not rate companies as investments or give buy or sell calls. Not investment advice. Do your own research.

**Last reviewed:** 16 Nov 2026 · **Next review:** after TrendForce’s third-quarter 2026 fabless and foundry tables, or TSMC’s fourth-quarter results

Spot an error? [Tell us](mailto:sean@thekopinotes.com) and we will correct it and note the change here.

[Our editorial standards →](https://www.sotkn.com/editorial-standards/)

## References

1. [TrendForce — Top 10 Global Fabless IC Design Houses Post 73% YoY Revenue Growth in 2Q26; AMD Breaks into Top Three (11 Sep 2026)](https://www.trendforce.com/presscenter/news/20260911-13231.html?ref=sotkn.com) — Second-quarter 2026 revenue of the 10 largest fabless companies: US$141.45bn combined, Nvidia US$91.16bn (64%), Broadcom US$18.89bn, AMD US$11.53bn, Qualcomm US$8.5bn, MediaTek US$4.81bn, Marvell US$2.63bn
2. [TrendForce — Global Foundry Revenue Approaches US$53.49 Billion in 2Q26 as SMIC Narrows Market Share Gap with Samsung (9 Sep 2026)](https://www.trendforce.com/presscenter/news/20260909-13225.html?ref=sotkn.com) — Top 10 foundries: TSMC US$40.2bn and 72.5% share, Samsung Foundry 5.9%, SMIC 5.4%, UMC 3.9%, GlobalFoundries 3.2%; Samsung’s orders include HBM base dies
3. [TSMC — 3Q26 Management Report and quarterly results (15 Oct 2026)](https://investor.tsmc.com/english/quarterly-results/2026/q3?ref=sotkn.com) — Revenue NT$1,494.24bn (US$46.72bn), gross margin 66.4% (2Q26: 67.7%), wafer shipments 4,476k 12-inch equivalent, revenue by node and platform, 79% from North American customers, 4Q26 guidance US$52.3–54.3bn
4. [TSMC — Third-quarter 2026 earnings call transcript (15 Oct 2026)](https://investor.tsmc.com/english/encrypt/files/encrypt%5Ffile/reports/2026-10/eba4fac72873bf546b669bd2b5d973df5c93215d/TSMC%203Q26%20Transcript.pdf?ref=sotkn.com) — 2026 capital budget US$60–64bn; N2 capacity to more than double in 2027; new 3nm fabs in Tainan and Arizona; A14 risk production 2027; customers may not sublet or auction capacity
5. [NVIDIA — CFO Commentary on Second Quarter Fiscal 2027 Results, SEC exhibit 99.2 (quarter ended 26 Jul 2026)](https://www.sec.gov/Archives/edgar/data/1045810/000104581026000073/q2fy27cfocommentary.htm?ref=sotkn.com) — Revenue US$96.2bn, gross margin 75.0%, cost of revenue US$24.1bn; supply and capacity commitments US$279bn, up from US$119bn, mainly memory; inventory US$31.6bn
6. [NVIDIA — Form 10-Q for the quarter ended 26 Jul 2026 (filed 26 Aug 2026)](https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm?ref=sotkn.com) — Supply constraints, multi-year manufacturing and supply agreements, US$4.5bn H20 charge for excess inventory and purchase obligations, H200 charge
7. [Singapore EDB — VSMC Celebrates the Grand Opening of Its First 300mm Fab in Singapore (28 Sep 2026)](https://www.edb.gov.sg/en/news-and-insights/vsmc-celebrates-the-grand-opening-of-its-first-300mm-fab-in-singapore?ref=sotkn.com) — VIS–NXP joint venture fab in Tampines: 130nm to 40nm, interposer applications, risk production, volume production 1Q27, about 44,000 12-inch wafers a month and 1,600 jobs by 2029
8. [Focus Taiwan (CNA) — Jensen Huang’s exit to have ‘short-term’ impact on Taiwan suppliers: Yang (16 Nov 2026)](https://focustaiwan.tw/business/202611160007?ref=sotkn.com) — Taiwan’s central bank governor on Nvidia’s chief executive resigning and the effect on Taiwan’s AI suppliers

S

**[Sean](https://www.sotkn.com/author/sean/)**

Writes Second Order by TKN — a plain-English map of AI infrastructure and semiconductors for non-specialist investors. Focus: who gets paid, where money sticks, and which step is stuck.

Covers AI infrastructure, semiconductors, and second-order stack reads.

[Author page →](https://www.sotkn.com/author/sean/) · [Editorial standards →](https://www.sotkn.com/editorial-standards/)