How American Is America's AI Boom?

The boom is American in its centre of gravity: capital, hyperscalers, software, data centres and final demand. Its industrial chain is far less American. According to Federal Reserve staff estimates, roughly 90% of high-tech equipment used in the United States is sourced abroad. This dependence does not make investment illusory. It requires three figures that are often conflated to be kept separate: spending committed, output produced on US territory and value captured by US firms.

Capex is not a measure of US production

The Bureau of Economic Analysis provided the latest clue in its advance estimate for the second quarter. Investment rose, led by equipment and intellectual-property products. Within equipment, the BEA reported broad increases, including information-processing hardware. It also specified that the estimate was based primarily on import data. Capital-goods imports were led by telecommunications equipment, semiconductors and industrial equipment.

This composition helps explain the gap between record capex and its domestic contribution. A US company may buy a server assembled in Mexico with processors fabricated in Taiwan and South Korean memory, install it in Virginia, then sell a global software service. The full bill enters the company’s investment. Only the stages produced in the United States enter US GDP.

The accounting does not say that imports destroy demand or mechanically “subtract growth”. In the identity C + I + G + X - M, consumption and investment already contain imported goods. The BEA subtracts imports to avoid crediting the United States with production performed elsewhere. An import can therefore signal strong demand while widening the gap between gross investment and domestic output. Our balance-of-payments guide develops the other side of the relationship: financing the external deficit.

The external adjustment changes the result

Because the national accounts contain no AI line item, three Fed economists built a proxy combining software, data centres, power facilities, computers and peripherals. They then apply an adjustment based on net exports of computers, peripherals and parts. Their method is not an official AI statistic. It reveals what disappears when analysis stops at gross spending.

In the first quarter of 2026, the proxy’s four investment components together added 1.18 percentage points to annualised quarterly growth. Associated net exports subtracted 0.45 point, leaving a 0.73-point net effect. In the fourth quarter of 2025, the external adjustment absorbed 0.61 point of a 0.75-point gross contribution, leaving only 0.14 point net. The calculations come from the data file attached to the Fed note.

A supply chain with several geographies

The answer depends on which layer is being observed.

Demand and financing are largely American. The major platforms decide the investments, order computing capacity and bear the financial risk. US territory also hosts the buildings, grid connections, installation, part of the maintenance and associated jobs. This layer is why data-centre construction and software spending do create domestic value added.

Hardware is far more international. The 14 July Fed note estimates that roughly 90% of high-tech equipment is sourced abroad. Another Fed note on AI trade traces the surge in exports of servers, graphics cards and parts notably to Taiwan, Mexico and Vietnam. It uses three customs codes, 8471.50, 8471.80 and 8473.30. Their precision is imperfect: some AI hardware is missing, while some included products serve other uses.

Software and revenue can flow back to the United States. Foreign production of a server does not prevent a US company from later capturing cloud, software or intellectual-property margins. Conversely, US incorporation does not automatically turn all worldwide sales into US production. The BEA explains that value added excludes intermediate inputs and remunerates capital and labour. Its industry accounts show that information-sector capital growth was among the leading contributors to real GDP growth from 2021 to 2024, a result consistent with an AI effect but unable to isolate it.

A transfer of demand, not a total loss

Importing equipment does not mean that all project value leaves the United States. Land, the building, part of the power grid, installation, software, operations and financing may be domestic. Nor does it mean that foreign suppliers capture the full final price: a value chain distributes revenue, wages and profits across several jurisdictions.

The macro point is narrower. The boom immediately stimulates demand for foreign capital goods before any productivity gains diffuse. Using earlier technology episodes, the 14 July Fed note estimates that an investment shock can widen the current-account deficit by about 0.2 percentage point of GDP for several quarters. This is a model response, not a measurement of the deficit caused by AI in 2026. The authors also find stronger spillovers to Taiwan, Korea and Mexico than their aggregate trade links alone would suggest.

This asymmetry may persist: the United States mainly imports capital-intensive goods and exports more knowledge-intensive services. Software revenue may offset part of the hardware bill, but not necessarily at the same time or in the same account. The semiconductor cycle and the power constraint on data centres therefore become macro variables, not merely technology topics.

The answer in three propositions

The boom is American by decision: capital, risk, demand and a major share of software assets are concentrated in the United States.

It is international by hardware production: high-tech equipment depends heavily on Asian and Mexican supply chains. The net-export adjustment shows that this layer materially reduces gross capex’s contribution to US GDP in some quarters.

It remains unsettled by future outcomes: productivity gains cannot be inferred from spending. Our review of the available evidence on AI and productivity finds task-level gains but incomplete macroeconomic diffusion. The debt financing the infrastructure adds a second test: who retains the return if revenue arrives more slowly than liabilities mature?

A falsifiable diagnosis

The thesis of a US boom with high import content would weaken under three observable developments: sustained growth in US semiconductor and electronic-equipment production faster than capex, a decline in the net import share of high-tech equipment and faster US computer-service exports sufficient to offset hardware.

It would strengthen if investment, equipment imports and the current-account deficit rose together while domestic industrial production plateaued. The next step is therefore not to total capex announcements. It is to track production, net imports, industry value added and productivity at the same time.

Sources

1. Bureau of Economic Analysis, GDP, Advance Estimate, Second Quarter 2026, 30 July 2026. Investment, equipment, intellectual property and import composition.
2. Bureau of Economic Analysis, The Expenditures Approach to Measuring GDP, 3 June 2025. Accounting treatment of imports in GDP.
3. Federal Reserve, Soto, Thieu and Allen, The AI Buildout and the Economy, 17 July 2026, and data file. Growth-contribution proxy and methodological limits.
4. Federal Reserve, Fiori, Lipa and Nuenninghoff, Technology Shocks, the AI Boom, and the U.S. Current Account, 14 July 2026. Import share, historical model and international spillovers.
5. Federal Reserve, de Soyres, Haag, Liu and Van Leemput, The Global Trade Effects of the AI Infrastructure Boom, 13 February 2026. Customs codes, supplier economies and measurement limits.
6. Bureau of Economic Analysis, How Does AI Drive Growth?, 8 June 2026. KLEMS framework, value added and information-sector capital.

Limitations

There is no isolated AI sector or exhaustive AI-capex series in the national accounts. The Fed proxy components include non-AI spending. Its external adjustment assigns goods between investment and consumption using a weight because BEA accounts cannot directly trace the destination of each import. The 90% estimate concerns high-tech equipment as defined by the authors, not an entire data centre. FEDS Notes express their authors’ analysis, not an official position of the Board.

Data cut off at 10:35 CEST on 4 August 2026. The June 2026 US trade report, scheduled for 14:30 CEST, is not yet public and is not incorporated into this analysis. This is not investment advice.