Structural shift

Resolved PARTIAL on 2026-10-01. The call made a claim about the world and a claim about stocks. The first held. The second missed by a wide margin, and in the wrong direction.

The claim about the world was that the limit on AI compute would move off the GPU and onto memory, networking, power and cooling, and that someone at the top of the chain would say so in the third quarter. NVIDIA did in August, blaming extreme memory pricing for margin pressure and describing its fiscal 2028 outlook as limited by supply. Micron did on September 30, saying the industry stays short through 2027 and 2028 and that customers are asking for more than it can commit. That condition passed without argument.

The claim about stocks was that ten constraint names (Micron, Lumentum, Coherent, Marvell, Eaton, Vertiv, Constellation, Vistra, Modine and Chemours, equal weight) would beat an equal-weight basket of NVIDIA, AMD and Broadcom by at least five percentage points between June 12 and September 30. The constraint basket returned minus 10.3% while the GPU basket returned plus 7.8%, which is eighteen points the wrong way. Nothing about that is a near miss on a threshold. The trade was wrong.

Being right about the bottleneck and wrong about the stocks is the whole lesson, so it's worth saying plainly what the record does and doesn't explain. It doesn't hold a per-name breakdown, so which of the ten did the damage isn't something this note can claim. It does hold two things. Memory and its neighbors sold off in the last week of the window after OpenAI paused frontier training, with Micron down more than 4% in a day. And the thesis's own argument about power and cooling says that hyperscaler exposure is already priced across the obvious names, several of them up two to five times, and that what's left is in the less obvious ones. Four of the ten basket names are on that list of the obvious. The basket was built from the companies the thesis elsewhere calls priced, and then asked to outperform.

There's an awkward detail in the dates. The call was made on May 10 and described a six-month outperformance. The test was written on June 12, and the measurement window starts there, five weeks later and after those five weeks of price moves. A window opened on the original date might have scored differently. It might not. The record doesn't say, and the result stands on the window that was written down.

The four thesis frames all carried this call at high conviction into the horizon: the September 22 through October 1 assessments in the optical, power, materials and custom-chip contexts read high every day. They were tracking the condition that passed. None of them looked at where the two baskets stood, though prices were observable the whole time. A conviction grade that can't see half of what's being scored isn't measuring the call.

What it changed in the thesis: nothing was withdrawn, because the constraint claim was the one the arguments rest on and it held. If anything the argument that the remaining edge sits outside the obvious names came out stronger, at this call's expense.

The authoring lesson: a constraint locates the scarcity. Whether the market has already paid for it is a separate question, and probably the harder one, and this call never asked it. A call that wants to claim outperformance has to say why the market hasn't paid for it yet, and pick its basket from names where that's true.200200