AI Trade — The Suppliers' Suppliers Move First

Micron and TSMC are the two numbers the AI trade watches most: Micron for memory, TSMC for every advanced chip Nvidia, AMD, Apple and Google sell. But both are customers too. They buy machines to build factories, and they hand work to test houses, packaging plants, substrate makers and design partners. Those suppliers report earlier or more often — in Taiwan every listed company has to publish its revenue by the 10th of the next month — so their numbers show up before Micron's and TSMC's own. Read the chain from the back, and you see the AI trade a few weeks ahead.

Two kinds of supplier, two different clocks

The companies TSMC and Micron pay fall into two groups, and they tell you different things.

• Paid per chip shipped. Every AI chip has to be tested (Advantest in Tokyo, Teradyne in the US), mounted on a substrate (Unimicron in Taiwan, Ibiden in Japan) and packaged (ASE). TSMC's design partner Global Unichip and the custom-chip designer Alchip are paid per project. Further down sit the server builders, cooling and power suppliers. These companies see volume as it happens — their monthly revenue is the closest thing to a live count of AI chips leaving the factory.
• Paid for new factories. ASML, Applied Materials, Lam Research, KLA and Tokyo Electron sell the machines that make chips. They are paid when TSMC and Micron decide to add capacity, often a year or two before that capacity produces anything. Their orders show what the two customers expect demand to be in 2027 and 2028.

Since the end of June the two groups have split. The per-chip suppliers with the most AI exposure are at or near record highs: Global Unichip is up 77.5%, Auras 42.0%, Advantest 25.9%, Unimicron 23.8%. The machine makers fell: KLA −31.4%, Applied Materials −25.3%, Lam −19.8%, Tokyo Electron −19.0%. The market paid for chips being shipped today and doubted the factories being planned for tomorrow.