A nine-name cohort, built in a weekend.

The cohort runs the full chip-to-grid stack — from the substation to the transistor — as a single analytical frame. Two of the nine names are not semiconductor companies. They sit here because the dollar-weighted center of mass of the AI-datacenter build sits at the electrical layer (G1–G4), not the silicon layer (L8b/L8c). Per-MW addressable content at the box-builder tier runs $700K–$2.8M; the power-semi BOM at the same MW is roughly twenty to sixty times less. portfolio-summary §1 · refinement-log Findings 8 + 11

The book is seven longs, two shorts — net long ~15.25% NAV, gross long ~17.15%. Core longs anchor at the silicon tier (TSM 5/5, NVDA and AVGO 4/5) and scale down through TXN at the system-BOM tier (3/5) and ETN/VRT at the grid-system tier (3/5) to NVTS as a probe (2/5). Shorts are WOLF (4/5, structural SiC share break) and INTC (3/5, foundry-leadership binary). portfolio-summary §2

Two non-negotiable construction rules emerged from the refinement loop. ETN+VRT combined ≤ 1.5–2.0× TXN NAV — both names are AI-capex amplifiers; uncapped, they convert the cohort into a single-factor bet wearing two tickers. And chip-layer and grid-layer names are held independently, not as a chain — box-builders are vendor-agnostic on power-semi supplier by commercial design. A strong VRT order book does not mechanically translate into NVTS or TXN P&L. refinement-log Findings 12 + 16

The AIR pattern is documented in its own field guide — see the working diagram. What follows is the roster for this specific run — twelve distinct roles, dispatched in the order they appeared.

Twenty-one atoms, six sub-agent types, one cohort. Every dispatch was logged. Every re-dispatch was logged. The audit trail is 439 lines long and is itself a deliverable.

Three moments where the loop earned its keep.

Stepping into the deck — different voice, same author.

A four-to-six-person team running fifty to two hundred names spends most of its analyst capacity on coverage maintenance — routine memos, earnings updates, model rolls. The cohort run above suggests the binding constraint can move. Coverage breadth handled by the fan-out, the senior desk concentrated on frame-setting and refinement. The PM's job is to brief well, read returning work against the rubric, and re-plan. The analysts' job becomes specialist depth on demand, dispatched per cohort rather than maintained per ticker. The desk gets to spend its scarce senior attention on the questions that actually move conviction.

1. No price discovery. The cohort is built on synthesis, filings, transcripts, and vault notes. It does not read the tape. Sizing rules — "initiate ETN at one-third now, full size below $360" — are framework discipline, not real-time signals.
2. No primary management contact. No 1:1 with a CFO, no broker channel checks beyond what has been published, no off-the-record sell-side colour.
3. No expert-network calls. No GLG, no Tegus, no industry consultants paid for two-hour calls.
4. No real-time market signals. No bookings panels, no hyperscaler capex tracker, no satellite parking-lot data, no foundry utilization scrapes.
5. No execution. This is research output. Orders, fills, hedging overlays, and risk-system integration sit elsewhere.
6. Source-corpus dependence. Output quality is bounded by the vault — four notes, ~41,000 words — plus public filings. ETN's deep-dive explicitly flagged thinness and leaned on filings; the system worked, but the gap was named, not hidden.

A senior desk closes these gaps with a phone, a corporate-access calendar, and a Bloomberg. This work makes those tools more leveraged — it does not replace them.

The cohort is not the deliverable. The loop is.