Fed paper: official data are too broad and too slow to track US chip fabs
A Federal Reserve working paper published on 29 September asks a practical question: how would a policymaker know, in close to real time, whether the United States is actually building chip factories? The answer from David Byrne and Robert Kurtzman of the Fed Board and Heidi Williams of the Congressional Budget Office and Dartmouth College is that official statistics alone cannot tell them.
The authors examine four government sources: the national accounts and fixed asset accounts from the Bureau of Economic Analysis, the Census Bureau's Annual Integrated Economic Survey, and the two most timely monthly surveys, manufacturers' shipments and orders and the value of construction put in place. All of them, the paper finds, use categories that are too broad or come out too late. The construction series covers computer, electronic and electrical manufacturing together, so a boom in battery plants landed in the same line as the fab buildout; that series rose tenfold from 2021 to 2023, while the annual survey's estimate for the same sectors did not rise nearly as much.
The alternative is project-level data from SEMI, the industry association, which tracks individual fabs worldwide each quarter with a lag of about ten weeks. Comparing it with official figures through 2023, the authors find broad consistency in investment levels and conclude that SEMI is a credible supplement. SEMI's forecasts include projects it judges at least 50% likely to go ahead.
The motivation is macroeconomic. Chips account for more than half the cost of AI-ready data centres, the paper notes, and because most of those chips are imported, the data-centre boom adds less to US growth than its size suggests.

Why it matters
For anyone reading US growth, AI investment is now large enough to move GDP, and this paper says the official data cannot yet show how much of it is being built at home. The authors call official and alternative data "complements, not substitutes".