# Semiconductor Demand Early-Warning Plan

## Goal

Build a fast, defensible research product that detects whether AI infrastructure demand is accelerating, plateauing, or slowing before the effect is fully visible in semiconductor revenue.

The product should not start as a global data-center census. It should start as a monthly or quarterly signal stack that tracks the conversion of capital, power, construction, and supply-chain commitments into AI-ready compute deployment.

## Recommended Research Question

Are hyperscaler and AI-cloud infrastructure commitments converting into enough power-ready, rack-ready AI capacity to sustain semiconductor demand growth over the next 2 to 6 quarters, and where are the bottlenecks that could accelerate or slow that conversion?

## MVP Name

Power-to-Silicon Demand Pulse.

## MVP Scope

Use a narrow first cut that can produce a useful answer quickly:

- US core markets: PJM/Virginia, ERCOT/Texas, AEP territory, Georgia, Arizona, Nevada.
- Global swing checks: Malaysia/Johor, India, Japan, Korea, Ireland/Europe, Gulf sovereign AI projects.
- Semiconductor exposure: accelerators, HBM, advanced packaging, custom ASICs, high-speed networking, power management, memory, and AI-server ODMs.

## Core Metrics

1. AI Infrastructure Conversion Index
   - Composite score from hyperscaler capex, signed power, construction conversion, semiconductor supply confirmation, and demand absorption.

2. Probability-Weighted AI-Ready IT MW By Quarter
   - Estimated IT load that is likely to become ready for AI racks, weighted by evidence quality and stage.

3. Power-to-Silicon Bridge
   - Translation from AI-ready MW into directional accelerator, HBM, networking, server, and power-equipment demand.

4. Bottleneck Score
   - Measures whether power, packaging, HBM, networking, financing, construction, or workload economics is the limiting factor.

5. Divergence Monitor
   - Flags when semiconductor guidance, hyperscaler capex, power commitments, and data-center construction stop agreeing with each other.

## Signal Stack

### 0-2 Quarters: Semiconductor Revenue Nowcast

- Hyperscaler capex guidance and revisions.
- Purchase obligations and finance lease commitments.
- Nvidia, AMD, Broadcom, Marvell, TSMC, SK Hynix, Micron, Samsung, Arista, and ODM commentary.
- HBM allocation, advanced packaging capacity, lead times, and pricing.
- GPU cloud utilization and rental pricing.
- AI-cloud financing and customer concentration signals.

### 2-6 Quarters: Deployment Conversion

- Signed leases and precommitments.
- Under-construction AI-capable MW.
- Substation, transformer, switchgear, generator, and liquid-cooling evidence.
- Energization and commissioning milestones.
- Cloud region and AI cluster launch cadence.
- Delayed/cancelled data-center and power projects.

### 6-12 Quarters: Power Pipeline

- Signed utility service agreements, electric service agreements, letters of agreement, and take-or-pay commitments.
- ISO/RTO and utility load forecasts.
- Interconnection queues filtered by deposit, site control, power study progress, and grid-upgrade feasibility.
- Utility capex filings, transmission plans, and resource adequacy constraints.
- Behind-the-meter generation and power-purchase agreements.

### 2-5 Years: Structural Ceiling

- Generation and transmission buildout.
- Water, permitting, local opposition, tariffs, and reliability rules.
- Sovereign AI policy commitments.
- AI workload economics, model-training demand, inference monetization, and cloud margin pressure.
- Semiconductor roadmap transitions, rack density, networking architecture, and memory intensity.

## MVP Deliverable

The first public result should contain:

- One-page investment read: bullish, neutral, bearish, or mixed by time horizon.
- Five-indicator scorecard with month-over-month or quarter-over-quarter changes.
- AI-ready MW funnel by stage.
- Power-to-silicon bridge table.
- Bottleneck and divergence notes.
- Source pack with confidence levels.
- Update checklist for the next refresh.

## Scoring

Each indicator is scored from -2 to +2:

- +2: clear acceleration.
- +1: modest acceleration.
- 0: neutral or mixed.
- -1: modest deterioration.
- -2: clear deterioration.

Every score must include:

- source references,
- basis,
- horizon,
- confidence,
- change since prior update,
- reason the signal matters for semiconductor revenue.

## First Build Sequence

1. Define the scoring rubric and conversion assumptions.
2. Pull current capex, purchase obligation, and semiconductor supply-chain signals.
3. Pull utility and power-commitment signals for the US core markets.
4. Add construction and leasing conversion checks for selected markets.
5. Build the first scorecard and output preview.
6. Review contradictions and decide where project-level data is actually needed.
7. Automate monthly refreshes only after the first manual pass proves the signal works.

## What To Avoid In The MVP

- Do not build a full global site-by-site database first.
- Do not sum announced MW, grid queue MW, under-construction MW, and operational MW.
- Do not treat ordinary data-center capacity as equivalent to dense AI training or inference capacity.
- Do not infer semiconductor demand from data-center square footage alone.
- Do not ignore demand absorption: idle GPUs, weak rental prices, or capex guide-downs can offset build progress.