A new, reproducible mapping of semiconductor policy finds support skewed across countries and hampered by long implementation delays; the resulting dataset and classification framework make it possible to track where AI‑hardware subsidies, tax credits and targeted financing are announced — though not yet their economic effects.

Main Finding

The paper builds a transparent, value-chain aware, text-based pipeline to identify and partially quantify semiconductor industrial policies (2010–2025) from Global Trade Alert and related documents. It shows that careful measurement—combining dictionary/HS-code screening, manual verification, monetary extraction, and package de-duplication—is essential to construct plausible country-year exposure measures and composition indicators (inputs, fab/design, assembly). Using event-style dynamics and continuous-exposure panel designs, the study emphasizes diagnostics (pre-trends, timing, sensitivity) rather than claiming definitive causal effects, and stresses that uneven observability, reporting lags, and implementation frictions materially shape inference about AI-hardware-related policy impacts.

Key Points

• Motivation: Rising AI-hardware demand makes semiconductor policy (subsidies, tax credits, directed finance) economically consequential and policy measurement nontrivial.
• Measurement pipeline: Hybrid identification using HS codes (when present) plus an iteratively refined domain dictionary applied to policy texts; manual verification and contextual checks for relevance and value-chain targeting.
• Reliability procedures: Stratified double-coding, documented reconciliation rules, ambiguity flags, and conservative de-duplication by policy package to avoid mechanical spikes.
• Quantification: Extract monetary commitments (ceilings vs realized disbursements flagged), convert to common currency and horizon; tax instruments converted to tax-equivalents only when defensible; otherwise conservative qualitative intensity scores.
• Outcomes / triangulation: Multi-dimensional outcomes—capacity proxies, trade/sourcing flows, innovation indicators (patents), and price measures—used jointly to reduce reliance on any single noisy proxy.
• Empirical design: Country-year panel with continuous exposure and composition measures; event-study diagnostics emphasize pre-trends, anticipation, and staged implementation.
• Uncertainty treatment: Versioned ETL scripts, explicit missingness reporting, winsorization sensitivity, limited imputation, and restricted-sample checks to address non-random missingness and observability bias.
• Mechanism focus: Recognizes operational margins (yield, uptime, downtime, logistics, congestion) that can mute or delay observable policy effects; constructs foreign-exposure measures to capture cross-border supply-chain spillovers.
• Limitations highlighted: Under-recording (opaque credit channels, subnational programs), reporting lags, possible systematic under-capture in some jurisdictions, subjectivity in classification, and residual measurement error that can bias causal claims.

Data & Methods

• Primary data source: Global Trade Alert (GTA), baseline window 2010–2025; pragmatic choice acknowledging under-recording risks.
• Identification strategy:
  • Use HS codes when available; supplement with a domain-specific keyword dictionary applied to policy descriptions.
  • Iteratively refine dictionary to balance recall (catch broad “advanced manufacturing” packages) and precision (avoid weakly linked ICT measures).
• Manual validation and coding:
  • Manual verification of candidates; map each policy to primary value-chain target (inputs, design/manufacture, assembly/packaging/testing) and instrument type (direct spending, tax incentives, preferential lending, equity, guarantees, procurement).
  • Stratified double-coding, logged disagreements, rule-based reconciliation prioritizing explicit eligibility/implementable clauses, ambiguity flags for sensitivity checks.
• Quantification procedures:
  • Extract monetary commitments and horizons; retain both ceilings and implementation-status flags where ceilings are reported.
  • Assign package identifiers and code incremental changes only when amendments alter effective support.
  • Conservative treatment of tax instruments and indirect measures; avoid aggressive imputation when documentation is weak.
• Outcome linkage:
  • Merge policy indices to panels of capacity proxies, trade flows/sourcing patterns, innovation metrics, and price measures.
  • Preprocessing: skew transformations, winsorization sensitivity, report coverage/frequency/missingness; restrict analyses to better-documented subsamples for robustness.
• Empirical designs:
  • Event-study around policy surge years (surge year defined by changes in exposure + program starts), controlling for country and year fixed effects; emphasis on timing diagnostics and anticipation.
  • Continuous-exposure panel regressions and heterogeneity analysis by value-chain composition and instrument type.
• Robustness & transparency:
  • Versioned ETL scripts, explicit missingness reporting, sensitivity to dictionary choices, alternative exposure definitions, and secondary-tagging for multi-target interventions.

Implications for AI Economics

• Measurement matters for policy evaluation: For AI-hardware economics, the paper shows that how you identify, quantify, and time semiconductor policies materially affects conclusions about their effectiveness and spillovers.
• Value-chain targeting is crucial: The same aggregate support can have very different effects depending on whether it targets upstream inputs/equipment, fabrication capacity, or downstream assembly/testing. Evaluations and policy design must consider composition, not only magnitude.
• Operational constraints can mask policy effects: Yield, uptime, equipment lead times, and logistics can delay or blunt observable capacity/trade responses; economists should incorporate operational margins when interpreting outcomes in the AI-hardware context.
• Anticipation and implementation lags: Firms respond to expected future support; researchers should use dynamic diagnostics and be cautious with event timing—policy announcements, legislative debate, and disbursement dates can imply different behavioral responses.
• Cross-border spillovers and supply-chain linkages: Domestic-only evaluations risk conflating national efficiency gains with international reallocation; constructing foreign-exposure measures tied to supply chains is important for understanding global AI-hardware resilience and strategic interactions.
• Policy monitoring and research infrastructure: The pipeline demonstrates the value of open, versioned ETL, systematic validation, and multi-source triangulation—recommendations for improving transparency and enabling better cross-country comparisons.
• Caution on causal claims, but actionable: While not definitive about policy efficacy, the framework narrows plausible narratives and provides a replicable basis for future firm-level triangulation, structural modeling, or coordinated international analyses—useful for policymakers and economists assessing industrial policy in the AI era.