Signal Board · Taiwan AI supply chain · Board A daily · Board B monthly
AI-Order Signal Board
An early-warning system for the question every AI-capex position depends on: are orders still flowing through the Taiwan supply chain? Board A watches five price tripwires daily — high recall, fires often, its job is to point. Board B reads the island's mandated monthly revenue filings — high precision, fires rarely, its job is to decide whether A was right. Neither board alone is the signal. The signal is their disagreement.
Board A — five tripwires daily · layer basket vs non-AI control leg · Lake-served
| # | Tripwire | State | Reading |
|---|---|---|---|
| T1 | Breadth < 50% above 50d MA | quiet | 64% above (9/14) |
| T2 | Basket/control RS at 20d low | quiet | RS 2.6043 vs 20d min 2.3367 |
| T3 | Upstream−L5 spread negative 3 sessions | FIRED | upstream-L5 20d spread last 3: -6.21%, -7.66%, -13.33% |
| T4 | Dispersion in top quartile (1y) | FIRED | dispersion 17.79% vs top-quartile floor 16.19% |
| T5 | Shock (−7% single / 3× −4%) | quiet | Lotes -5.0% |
0 green · 1 watch · 2 amber · 3+ red. T3 — upstream underperforming downstream — is the geometry of orders thinning at the front of the chain while backlog still ships out the back. Rising dispersion (T4) with healthy breadth usually reads as substitution, not contraction. The board runs untuned through its first two quarters; the false-positive log is the only tuning mechanism.
Board B — layer revenue monthly · TWSE-mandated filings · NT$ · 3-month rolling YoY
| Layer | 3M YoY | Δ vs prior | Months decelerating |
|---|---|---|---|
| L2 · Substrate / PCBUnimicron, Zhen Ding — the earliest monthly signal in the chain. | +37.6% | +5.5% | 0 |
| L3 · Power / ThermalDelta, AVC (Auras pending) — the binding constraint at 200kW racks. | +51.4% | +0.2% | 0 |
| L4 · Interconnect / OpticsAccton, Lotes, BizLink (LandMark pending) — mid-chain. | — | — | 0 |
| L5 · Assembly / ODMWiwynn (purity anchor), Quanta, Wistron, Hon Hai — AI-weighted; confirms, rarely warns. | +59.9% | +1.5% | 0 |
| L6 · Silicon / SubstitutionTSMC, Alchip, Global Unichip — diagnostic, not directional; ASIC pair feeds the substitution filter. | +46.9% | +10.8% | 0 |
M-tests — deceleration breadth 0/4 · sequence quiet · substitution filter quiet · ASIC pair 3M YoY +85.0%
A real warning requires deceleration in ≥1 layer for 2 straight months (M2), in ≥3 of the four signal layers at once (M3), appearing upstream before the ODMs (M4) — and it must survive the substitution filter (M5): ASIC names accelerating while ODMs slow is mix shift, not demand. January and February are always combined (lunar new year moves between them); everything is computed in NT$, the reporting currency.
Methodology
The layer taxonomy runs L2 (substrate — earliest signal) through L5 (assembly — latest, confirms rather than warns), with L6 silicon as a diagnostic and a six-name non-AI control leg (Largan, Compal, Acer, Catcher, Chicony, Merry) separating AI signal from island-wide cyclicality. L5 is weighted by AI-attributable revenue, not total revenue — at full weight, Hon Hai's handset seasonality would drown the board. Expected propagation order is L2/L3 → L4 → L5; anything violating that order is suspect by construction.
Board A fires amber several times a quarter by design — that recall is the job. A screaming Board A against a clean Board B is a legitimate state: revenue lags bookings by a quarter or more under non-cancelable order structures, so A may simply be early. That state is held open and logged, never auto-resolved. Every Board-A red is recorded with Board B's state at the time; after two quarters that log sets the thresholds.
A signal board is a diary entry on our process. Closelook publishes research, not investment advice.