Lesson 2 of 7

Map the value chain

Step 3 of 6

Worked example

The AI chip chain

The Semiconductor Industry Association (SIA) describes four stages in making a chip: research, design, front-end manufacturing (fabrication) and back-end assembly, test and packaging. The AI version of the chain, with one example company per link:

Link Example Year Gross / operating margin
Chip-making machines ASML 2025 52.8% / 34.6%
Foundry TSMC 2025 59.9% / 50.8%
Chip designer Nvidia (whole company) FY to 25 Jan 2026 71.1% / 60.4%
Cloud buyer Microsoft FY to 30 Jun 2026 67.9% / 46.8%

Sources: ASML and TSMC 2025 annual reports on Form 20-F (the annual report foreign companies file with the SEC), Nvidia's and Microsoft's 10-Ks. ASML reports in euros and TSMC in Taiwan dollars, but margins are percentages, so they compare.

The arithmetic:

ASML       gross 17,258 ÷ 32,667 ≈ 52.8%   operating 11,301 ÷ 32,667 ≈ 34.6%  (€M)
TSMC       gross margin 59.9%, operating margin 50.8%  (as stated in its 20-F)
Nvidia     gross 153,463 ÷ 215,938 ≈ 71.1%   operating 130,387 ÷ 215,938 ≈ 60.4%  (total company revenue)
Microsoft  gross 225,465 ÷ 331,839 ≈ 67.9%   operating 155,237 ÷ 331,839 ≈ 46.8%

In this chain, the chip designer has the highest margins of the four. One caution: Microsoft's margin covers all of Microsoft (Office, Windows, games), not only AI. A whole-company margin can hide what one link earns.

Concentration, in Nvidia's own words

Suppliers. Item 1 of Nvidia's FY2026 10-K names them:

  • foundries TSMC and Samsung make its wafers (the silicon discs chips are cut from);
  • memory (the chips that hold data while the processor works) comes from SK Hynix, Micron and Samsung;
  • packaging uses a technology called CoWoS;
  • Hon Hai, Wistron and Fabrinet do assembly and testing.

Few names for each critical input means real supplier power.

Customers. The 10-K says: "For fiscal year 2026, sales to one direct customer represented 22% of total revenue and sales to another direct customer represented 14% of total revenue." Both sales were mainly in the Compute & Networking segment. You turned these into dollars in Module 5 lesson 03: together about $77.7B, over a third of revenue. Here, read them as buyer power.

Three more facts from the same filing:

  • three direct customers held 25%, 18% and 13% of Nvidia's money owed by customers (accounts receivable) at 25 Jan 2026;
  • a year earlier, the biggest direct customers were 12%, 11% and 11% of revenue, so concentration rose;
  • "one AI research and deployment company contributed to a meaningful amount of our revenue" by buying cloud services from Nvidia's customers.

Nvidia does not name these customers, and you shouldn't guess. A direct customer buys from Nvidia (often a server builder). The company that finally uses the chips may be someone else.

The buyer side. Microsoft spent $115.9 billion in cash on property and equipment in FY2026 (year ended 30 Jun 2026), up from $64.6 billion a year earlier. That figure excludes equipment acquired through finance leases (leases that work like buying with a loan; lesson 03). It isn't all AI: the 10-K says it covers "new facilities, datacenters, and computer systems". But it shows how much one customer of the chain is putting in.