We tested the obvious AI story. It didn't survive.

Everyone was talking about AI chips and hyperscalers. We tested whether they were actually where the biggest structural changes were happening.

They weren’t.

Across 81 trading sessions and a fixed panel of 358 companies, an AI, data-center and power cohort we named in advance widened no faster than the rest of the market. Both rose 0.016. The category everyone was writing about did not move as a category.

The largest individual changes were one layer down — in optics, interconnect, memory, storage and test equipment.

That was not the hypothesis. It was the result.

VERDICT_START|The claim|AI infrastructure was structurally moving before the market noticed.|Our test|Fixed panel of 358 companies, 81 sessions, 31 cohort names chosen in advance.|Verdict|FAILED — the cohort widened 0.016; everything else widened 0.016.|What survived|Individual suppliers a layer down: optics, interconnect, memory, storage and test equipment.|VERDICT_END

The category did not move

Thirty-one AI, data-center, power and hyperscaler companies, chosen before we looked at any numbers, against the 327 other companies in the panel.

The lines rise together and the gap is the same at the end as at the start. The cohort held four of the top fifty risers against a base rate of nine percent — chance. Measured at the level the story was told, there was nothing there.

The names underneath

Sort the same 358 companies individually and the picture changes completely.

CALLOUT_START|Lumentum’s turbulence rose 0.312. The market’s median rose 0.016. One company moved roughly twenty times as much as the typical one — and it wasn’t one of the companies dominating the summer narrative.|CALLOUT_END

Lumentum’s turbulence went from 0.442 to 0.754. The control’s median rose 0.016. That is roughly twenty times the typical increase, in a single company.

Read the list by what these businesses actually do and eight of the twelve are the same three things: optics and interconnect (Lumentum, Ciena, Corning), memory and storage (Western Digital, Sandisk, Micron), and test equipment (Teradyne). Not the chip designers, and not the hyperscalers.

Four do not fit that pattern — Oracle, AMD, Datadog and Gartner. They are listed because a ranked list with the inconvenient entries removed is not a finding, it is a selection.

When it happened

The session by which each name had completed half of its total increase:

The movement is concentrated in late May and early June. We would like to tell you how that compares with when these companies started drawing attention, and we cannot: Filter Lab does not yet store a history of news volume, so we can measure when the structure moved but not yet when the coverage did. Building that archive is the next thing on our list, and until it exists we are not making claims about what the press did or did not say.

What this does not establish

The cohort was defined in advance and failed. The twelve names were found afterwards, by sorting a finished window and reading the top. That is a description of what the measure captured, not evidence that it predicts — anyone can identify a leader once the race is over. Only the first of those two results is a test.

Turbulence is also range, not direction. Nothing here says these companies rose or fell, or will. It says the spread of their possible outcomes widened while the market’s widened far less.

The selection rule is now fixed and the forward record starts from 2026-08-04.

Why it is worth your time anyway

The useful part is not the tickers. It is the shape of the error.

The public category — “the AI trade” — was too coarse to describe what was actually happening underneath it. Measured as a category, nothing. Measured company by company, a great deal, concentrated in businesses outside the dominant narrative.

That gap between the category and the companies is what Filter Lab is built to find.

How this was measured

83 daily snapshots of the Filter Lab US equity universe between 2026-05-13 and 2026-08-04. Two were partial builds and were excluded, leaving 81 sessions. To stop a growing panel from manufacturing a trend, the sample is fixed at the 358 companies present in every one of those sessions; 31 of the 37 predefined cohort tickers fall inside it, and the remaining 327 companies are the control. Turbulence is the 90-day variance score, a descriptive measure of the range of a company’s movement. Medians throughout, not means, so a single extreme name cannot carry a group.