AI infrastructure fell on a 2028 fear. The next few weeks will test it.

Hey {{first_name|Investor}} -

Three of the earliest-committing links in the AI chain, ASML, TSMC, and Aehr, all raised their outlooks this month, with order books that now reach into 2028. The market had already sold the stocks as if the buildout were ending. That divergence, a falling narrative against rising bookings, is the setup worth working through.

When price and fundamentals disagree, the useful question is which clock the market is pricing. Two clocks run through this whole debate. The near one, 2025 to 2027, is orders, capacity, and margins, the factory clock. The far one, 2028 and beyond, is absorption, financing, and permitting, the market's fear clock. The selloff is pricing the far clock. The transcripts answered the near one.

The fastest way to lose money in a selloff is to assume everyone else is dumb. The strongest version of why these stocks fell appears almost nowhere in the three transcripts.

It started with one word: excess. On July 1, Bloomberg reported that Meta is standing up a cloud business to resell its spare AI compute, and the market read the move as the first hyperscaler quietly conceding it had overbuilt. The group was already primed for that reading. Broadcom's June guide had come in a shade under hopes and it declined to raise its full-year AI number, so sentiment was fragile before Meta ever hit the wire. Layer on oil spiking with Middle East risk, rates pressuring every long-duration equity, and a memory price spike bleeding into consumer-hardware demand, and you get a six-week derating of the entire complex.

The one bear thesis worth genuine respect is not surplus. It is the mismatch between how this buildout is financed and how fast it can be absorbed. A rising share of hyperscaler capex is now debt-funded rather than paid from operating cash flow, and the marginal appetite for that debt looks thinner than it did in the spring. Reporting this month described demand for big-tech bond deals cooling, with cover ratios on some issues said to have fallen from around five times earlier in the year toward two. The distinction that matters is between funding getting more expensive and funding drying up, and the evidence points to the former, which pressures returns rather than halting projects.

On the demand side, UBS has sketched capital-spending growth cooling from the mid-70s percent this year toward the mid-20s next year and single digits by 2028, on its estimates, and that trajectory is what matters for anyone modeling second-derivative revenue off these names. And a recent industry estimate put the share of US data-center projects facing local resistance over power and water above two-thirds, with at least one state pausing large builds. Treat each figure as an analyst's or reporter's estimate rather than settled fact, but the vector is real, and it is the case a serious bull answers rather than dismisses.

A firm genuinely drowning in idle compute does not open a storefront to rent that compute out; Meta's decision to resell is more consistent with internal lease economics that already clear its cost than with a glut, and it has kept committing heavily to infrastructure even as it moves to monetize the overflow. That reframes "excess" as arbitrage, not oversupply. But notice the timescale. Every worry above, renewal risk, financing cost, the deceleration in spending growth, permitting, lives in the back half of the decade. The market sold them as facts about this quarter. The three factories answered for this quarter and the next two years and said almost nothing about the far clock. The selloff, then, is not wrong so much as early, pricing a question the calls did not attempt to settle. That gap is where the work is.

The rest of this issue, the upside management left out of the numbers, the dated catalyst calendar, and how I separate conviction from timing, is for paid members.