Silicon's Final Exam

Correction, August 16: an earlier version of this piece said Teradyne entered the model portfolio at the May rebalance and was its newest position. That was written against a stale copy of the database and was wrong. The live record — including the part where the position was sold too early — is below. On this site the record wins.

Teradyne makes the machines that test chips. Semiconductor companies don't ship anything untested — every die gets probed at the wafer stage and again after packaging, and the equipment doing that probing comes almost entirely from two vendors, Teradyne and Advantest. It's been a duopoly for a long time.

The thesis

Test has historically been a mediocre business to own. The equipment doesn't add any performance to the chip; it's a cost, and a deferrable one, so orders track chipmaker capex with a vengeance. In good years everyone re-tools, in bad years the phone stops ringing. Teradyne has had years where revenue roughly halved.

What's different now is the product being tested. An AI accelerator isn't a single die anymore — it's a logic die plus several stacks of high-bandwidth memory bonded into one package, and the package is worth tens of thousands of dollars. That changes the arithmetic of when you test and how hard. A bad die caught at wafer sort costs you a die. The same bad die caught after it's been bonded into a finished package costs you the package, including all the good silicon around it. So as packages get more complex and expensive, manufacturers test earlier, more often, and for longer per device. The testing bill per chip is going up even before you count the growth in the number of chips.

Fundamentals

The thesis was, and remains, roughly right. Revenue in the latest reported quarter roughly doubled year over year, on top of a string of large earnings beats stretching back through 2025. The market pays about 57 times trailing earnings and 36 times forward for it, on a $65bn market cap — a full price for a famously cyclical business, justified only if AI-driven test demand is a structural shift rather than one hot capacity build.

What the record shows

The position was bought on March 2 at $325.71. One month later the scores had deteriorated enough that the rules said sell, and it went at $312.08 — a 4% loss. The stock then went on a run nobody here participated in: as high as $484, and $419 as of Friday's close. Measured from the sale, that's a 34% move watched from the sidelines.

There's no way to dress that up, so it won't be dressed up. A rules-based system re-scores every position on a schedule and acts on what the scores say that day. The scores fell, and the system did what it was built to do. The cost of never overriding the rules is that sometimes the rules are wrong, and this one was expensive. The benefit is that the same discipline exits genuinely broken positions without argument — and that the record stays honest, because there's no human in the loop to quietly forget the misses.

Could it come back into the book? The long-horizon score is still strong — the thesis above hasn't gone anywhere — but the short-term reading is negative after the run-up, and the composite doesn't clear the entry bar. If that changes at a future rebalance, the position would be re-opened 30% above where it was sold, without embarrassment. Machines are good at that part.