How much of the AI boom is adding to U.S. growth?

Hyperscalers are on track to spend nearly $800bn on capex this year, roughly ten times their 2019 spend. Relative to GDP, investment in tech equipment and software has already surpassed peak dot-com levels, helping power corporate profits to their best quarter since 2021.

Yet, by our estimates, the impact on the economy has been relatively modest. A proxy for AI-related investment contributed 0.47 percentage points to the 2.1% pace of U.S. real GDP growth over the past year, roughly one-fifth of growth, after netting out imported hardware. Consumption contributed over three times as much.

The AI capex boom is a key uncertainty for investors, economists and above all the Federal Reserve. Is the spend simply not very growth additive, or are the national statistics not capturing it properly? For several reasons, AI capex is not a one-for-one boost to U.S. growth.

1. Hyperscaler capex is not all U.S. AI investment.

U.S. hyperscalers are multinational companies, so not all of their spending is domestic. Estimates suggest roughly 30% of their capex goes to overseas data centers.¹ Nor is all capex AI-related, since these companies were spending roughly $150bn a year before AI. The measure misses spending too. It leaves out private AI companies, neoclouds and foreign investment, along with adoption costs like workflow redesign or worker training.

2. Imported hardware is netted out of GDP. 

When a hyperscaler buys an imported server, the purchase adds to investment but is subtracted in trade, which should net to roughly zero. AI supply chains, however, make that treatment imperfect. Imported chips used to train a firm’s own model can be treated as intermediate inputs, meaning the import is subtracted from growth without a corresponding addition to investment.² 

And perhaps the biggest issue is U.S. chip design. Nvidia chips, for instance, owe much of their value to U.S. designs but are manufactured and assembled in Asia. That design value should appear as an export of intellectual property, but often does not because no physical good leaves the U.S. and no foreign buyer pays separately for the design. Epoch AI estimates this issue alone has understated annualized U.S. growth by about 0.3 percentage points in the last year.³

3. Some of the spending increase is price, not quantity. 

Surging memory prices have contributed to a record climb in the price index for business investment in computers, up 10.9% year-over-year. GDP measures real investment, so rising prices should not add to growth, and roughly a sixth of last year's rise in nominal tech investment came from price rather than new equipment.

AI’s payoff will unfold gradually

The true impact from this AI buildout will unfold gradually. An AI server assembled abroad may be neutral for GDP growth on the day it is purchased, but it still adds to the U.S. capital stock, humming away in a data center in Virginia or Ohio and providing U.S. workers with far more AI capabilities. That capacity should eventually show up as productivity as diffusion deepens.

Kevin Warsh has put the productivity question at the center of the Fed's agenda and set up a task force to study it. Clearly, capturing the AI boom in our national statistics has proven challenging, but while mismeasurement is directionally clear, the magnitude isn’t. Amid rapid economic transformation, policymakers may have to rely more on alternative data sources, equity market signals and judgment to discern the appropriate policy rate for an AI-powered economy.

¹ Goldman Sachs Global Investment Research, Joseph Briggs and Sarah Dong, August 2, 2026. Authors estimate that roughly 70% of U.S. hyperscaler capex goes to projects in the U.S., with 15% in Asia and 9% in Europe.

² The BEA treats semiconductors as intermediate inputs, counted toward final demand only when the products they enable, such as laptops or finished servers, are sold. Separately, chips used to train a firm's own model go into an intangible asset, which is not fully capitalized as investment. 

³ Epoch AI, “The Nvidia-sized hole in US GDP statistics”, August 24, 2026. Chips designed by Nvidia in the U.S. and fabricated abroad return in finished servers whose import price includes that U.S. design value-add, so GDP subtracts it in whole. The national accounts should capture this as an export of IP, but the data so far hasn’t. Epoch uses Nvidia's U.S.-booked operating income as a proxy and estimates the gap could approach two percentage points of GDP growth by the end of 2028.

83798eef-a61d-11f1-997c-1f8ba3a2f539