AI investment lifts prices but adds little to productivity so far
AI is raising U.S. investment and prices in exposed sectors, while aggregate employment and productivity effects remain limited, Federal Reserve researchers find. Much of the recent productivity acceleration predates generative AI.
Capital spending outruns proven returns
Federal Reserve authors Nick Carollo, Leland D. Crane, Daniel Garcia Molina and Joseph Nichols document an AI infrastructure expansion.
Data center plans exceeded $1 trillion in 2025, versus almost $300 billion in 2024.
A middle scenario using plans through May 2026 puts investment above $880 billion by end-2027, exceeding 2 percent of GDP, versus about $100 billion in 2024.
Imports account for about half of related equipment investment, limiting domestic GDP gains.
Power, equipment, labor, local opposition and debt financing constrain construction.
OpenAI, Anthropic and Google DeepMind had combined annualized revenue of at least $107 billion in Q3 2026, based on unaudited disclosures.
That equals about 19 percent of 2024 software gross output; consistent profitability remains elusive.
Broad adoption, narrow daily use
More than 40 percent of workers report workplace AI use, but only 10 percent use it daily; firm adoption estimates range from about 20 to 78 percent.
Highly data-center-exposed core PCE categories recorded inflation around 5 percentage points above the median-exposed category since January 2025.
One study finds employment among workers aged 22–25 in AI-exposed occupations fell 16 percent relative to comparison groups after LLMs emerged.
Reduced hiring dominates, but aggregate effects remain unclear; overhiring and higher interest rates complicate attribution.
Task gains are not GDP gains
The authors find that task-level efficiency gains have not yet translated into a clear aggregate productivity boost.
Broader gains depend on deeper adoption, organizational changes and complementary investment.
These preliminary staff findings do not represent the Federal Reserve Board's conclusions.