Buongiorno.
In June’s back-end equipment deep dive — The Fields Have Grown. One Seed Spot Remained. — the one seed spot was Advantest, and I pinned the verification event to its earnings release on July 29. That date has now come and gone. Let me say it upfront: the seed has sprouted. Today’s note is the scorecard on that call — and a promise to keep tracking this field.
The same week brought one more scene. On July 24, Nomura published a lengthy sector report covering the entire semiconductor test value chain, initiating six Taiwanese test names — all at Buy. Somewhere in the past few days, “test is the next bottleneck in AI chips” stopped being a contrarian frame and became the market’s common language.
Step back and my read is this. The binding constraint on AI infrastructure is, in the end, the sheer quantity of chips the industry can supply — the Q. And expanding Q is no longer a corporate capex question; it has become a national-level project. Fail to scale compute fast enough and you either fall behind or pay dearly for what you get. Test is the gate every chip must pass before it ships. When the gate narrows, the whole field’s harvest runs late.
Bigger chips, longer exams
Why test, why now? Because chips are getting bigger and hotter. By Nomura’s estimates, final test time for NVIDIA silicon runs about 4x Hopper’s level for Blackwell and roughly 7x for Rubin. System-level test and burn-in stretch with every generation as well, lifting test’s share of AI chip cost from an estimated 1.9% at Hopper to 3.3% at Rubin.
The mechanics are simple. More transistors mean more items to check; higher power means testing under active thermal management. To ship the same number of chips, you need more testers. Two companies effectively keep this gate: Advantest and Teradyne. In June I kept only the former as a seed — the latter was trading above my fair midpoint at the time, so I left it on the watch list. Teradyne has since resurfaced in my Physical AI deep dive as the parent of Universal Robots, which gives me one more reason to line the two up side by side.
Advantest: the answer came the morning after
Results landed after the close on July 29. Revenue of ¥367.5bn, up 39.3% year over year; operating profit of ¥190.0bn, up 53.3%; an operating margin of 51.7% — a quarterly record across the board. The company raised full-year guidance again, to ¥1,714bn in revenue and ¥846bn in operating profit, explaining that AI inference test demand had already outrun its April assumptions.
Going into the print, the stock was depressed. Contagion from a sharp memory-sector selloff had knocked it down more than 10% the day before, and it closed the release day slightly lower at ¥25,200. On the call, management flagged that Q1 margins had been flattered by one-off items, including an inventory valuation reversal, and warned of memory component cost pressure in the second half. The market’s answer came the next morning anyway: two consecutive double-digit up days, to ¥32,500 as of July 31.
