The market says the AI buildout has peaked. Three chip suppliers just disagreed.

On July 1, Bloomberg reported that Meta is building a cloud business to sell its excess AI computing power. The market heard one word, "excess," and decided the buildout had produced a glut. If the biggest spenders have compute to spare, the logic ran, then spending has peaked and the trade is over. AI-infrastructure stocks tumbled, and weeks of jittery macro, soft chip guidance, higher oil, rate nerves, did the rest. Much of the complex is well off its highs now. Some of the smaller, higher-beta names have lost 40% or more. Even the steady anchors got pulled in, though by less: TSMC came into this week only about 10% under its own high.

I've been bullish on where these businesses head over the medium term. Bullishness is a feeling, though, so I went looking for evidence in the place it is hardest to fake: the order books of the companies that commit money years before anyone else. A chipmaker can wait for demand to arrive. The three companies below have to build capacity two or three years ahead of it, so their bookings are a bet on the future written in real money.

Three of them reported inside 72 hours, between July 15 and 16, from three different heights of one supply chain. All three raised. That is the test, and it just came back.

Follow a leading-edge AI chip from raw silicon to a finished part, and three chokepoints show up. Each is a company on this list.

The machines that make the chips are ASML ($ASML). ASML builds the systems that print the smallest features onto advanced chips, and it is effectively the only company on earth that makes them. Each system costs about as much as a small building and takes years to deliver.

The wafers themselves are TSMC ($TSM). TSMC's fabs build most leading-edge AI chips, whatever logo goes on the box. A fab is the factory that manufactures the chips. Its quarterly call is a demand check on the whole compute stack at once.

The quality gate is Aehr Test Systems ($AEHR). Most readers won't know this one, and it is the sharpest tell of the three. Aehr makes burn-in equipment. Burn-in means stress-testing chips at high heat to catch the ones that would fail early, before they ship. Nobody adds burn-in capacity for chips they don't plan to ship, so demand here echoes real production plans rather than forecasts. It is a small company, so it carries the extra risk any small cap does. As a signal, it is clean.

The most important thing ASML said had nothing to do with the quarter it just printed. It was about 2028.

On the call, ASML said it is nearly fully booked for its main EUV systems in 2027, and has already taken a meaningful number of orders for 2028. Its finance chief called booking that far out something the company "haven't enjoyed in many years." A company at the very front of the chain is filling its 2028 order book right now, while the market acts as if the story is ending.

On that demand, ASML said it is studying back-to-back capacity increases of roughly 30% for both 2027 and 2028. Next year's machines run faster than this year's, so the same 30% more units works out to closer to 45% more actual chip output, on the company's own math. The raise is bigger than it looks. ASML also expects its memory-related system sales to grow more than 75% this year, on its guidance, as memory makers race to add lines.

It confirmed one milestone the industry has waited years for. Intel is now running ASML's newest, most advanced machines in real production. That is the first proof the next generation of chip printing works on a factory floor, not just in a lab.

I read this call as the cleanest data point in the batch. You don't quietly book 2028 capacity to flatter one quarter.

On any TSMC call, the sentence I care about most is never the quarter itself. It is the full-year guide. A raise says the buildout is still speeding up; a hold says it is merely healthy. They raised it. TSMC lifted its full-year 2026 growth outlook to a little above 40% in dollar terms, up from above 30% a quarter earlier.

The capital plans said it louder. TSMC raised its 2026 capital budget again, to roughly $60 to $64 billion, the third increase this year, up from the $52-to-$56-billion range it started with. It announced another $100 billion for Arizona, which the chairman said means about four more fabs. And it still calls its advanced-packaging capacity so tight that, in his words, it is "limiting my customers' growth." Advanced packaging is the step that binds a processor to its memory in one module. When a company says its own capacity caps its customers, that reads as unmet demand, not a glut.

The mix shows where the growth comes from. High-performance computing, the segment that holds AI and data-center chips, now runs at roughly two-thirds of TSMC's revenue, and gross margin sat near 68% on the company's reporting, high even by its own history. A newer thread ran through the call too. Management leaned into agentic AI reviving demand for CPUs, on top of the GPU accelerators, a second demand curve laid over the first. TSMC builds the leading-edge version of both.

The chairman also said TSMC is actively checking its customers' data-center build progress to make sure the chips it ships get used rather than sitting in inventory. That speaks to the quality of the demand, not just its size: a company watching for a glut on your behalf, not ignoring the risk. Its next node, called A14, is on track for volume production in 2028, so the roadmap that carries this past the current cycle is already moving.

I'll give the other side its due in Part 2, because TSMC also said careful things about consumer weakness and margins. On the core question, the factory answered by opening its wallet wider.

The small one has the most dramatic numbers. For its new fiscal year, Aehr guided revenue to roughly $130 to $150 million, about two-and-a-half to three times what it just did. Recent-quarter bookings jumped more than fivefold from a year earlier, and backlog hit a record that now runs well ahead of shipments. The concrete orders are landing in real time. Aehr called out a roughly $41 million order from a large data-center customer, and about $8 million in fresh silicon-carbide orders inside a single month. Its CEO thinks burn-in is the fastest-growing segment in all of chip testing.

The mix is the interesting part. Two years ago the company was almost entirely tied to silicon-carbide chips for electric vehicles. Today most of its business comes from AI processors and the optical parts that shuttle data around AI data centers, and it just added a new silicon-photonics customer it described as a global networking leader. A company whose whole job is testing chips before they ship has rebuilt itself around AI, and its customers are ordering that test capacity ahead of production they have already committed to.

The caveats are real. Much of the guide leans on a few large customers, management said the most exciting wins aren't even in the number yet, and results in a company this size will be lumpy. It also competes with other test and burn-in houses, and its customers could adopt different test approaches, so the signal is real without being guaranteed. As downstream confirmation that real chips are getting ready to move, though, the quality gate tells the same story as the two heights above it.

Line the three up and one sentence keeps repeating in different words. These companies are limited by how fast they can supply, not by how much anyone wants to buy. ASML is booking 2028. TSMC calls its own packaging the limit on customer growth and raises capex a third time. Aehr's backlog runs ahead of shipments. Visibility across all three is getting longer, not shorter, which is the opposite of what you hear at the top of a cycle.

Here are the receipts in one view.

Those figures come from each company's July call and guidance, so confirm them against the releases before leaning on any one number.

The market sold the word "excess." The order books say the opposite. That gap is the point.

The market isn't stupid. The complex didn't sell off because traders can't read an order book. It sold off because the tape is worried about a different clock. There are two of them. The near clock is 2026 and 2027: orders, capacity, and margins, and the three factories just answered it. The far clock is 2028 and beyond: whether all this capacity gets absorbed, and whether the debt funding it stays cheap. The transcripts are loud on the near clock and quiet on the far one. And the same memory boom lifting ASML and Aehr is quietly squeezing the consumer end of TSMC's business.

Part 2 takes that apart. What the drawdown is really pricing, what management left out of the numbers, a dated calendar of what to watch across the whole buildout, and how I think about acting on medium-term conviction without catching a falling knife.

Three of the earliest-committing links in the chain just told you, in the language that is hardest to fake, that they are still building faster. That is not a reason to buy tomorrow. It is a reason to trust the "it's over" story far less than the share prices want you to.

Stay disciplined - Koh

Disclosure: I have a position in $TSM. The other companies named in this issue are not positions for me.

Disclaimer: Nothing in this newsletter constitutes investment advice or a recommendation to buy or sell any security. Numbers and observations are as of publication. I may hold positions in companies discussed above. Always do your own research and consult a licensed financial advisor before making investment decisions.