A report published August 24, 2026 by Epoch AI researchers Isabel Juniewicz, Daniel Carey, Phil Trammell and Anson Ho. It argues that official US GDP statistics do not capture most of the value that Nvidia and other US fabless chipmakers create by designing chips that are manufactured, assembled and sold abroad, so that the AI hardware boom has been recorded as adding roughly nothing to the US economy on net. Using Nvidia's US-booked operating income as a proxy for the missing value-add, it estimates that "US GDP growth over the last year has been underestimated by about 0.3 percentage points." The report states that it confirmed its analysis with the Bureau of Economic Analysis (BEA). A companion post by Carey appeared on Epoch AI's Substack the same day.
Starting observation
The report opens from the scale of AI hardware spending: US investment in computing equipment of roughly $400 billion per year, "nearly triple its 2023 level," with a modest measured effect on GDP growth. It accepts the usual explanation that much of this investment buys imported goods, which are subtracted from GDP, but calls that explanation incomplete. Citing BEA data via FRED, it shows US private fixed investment in computers and peripherals and net imports of the same category rising nearly in lockstep, both to around $400 billion in 2026, so that the accounting "registers as adding roughly zero value to the US economy." The authors say this near-exact cancellation is what first drew their attention to the issue.
How the value-add goes missing
The report describes three steps in getting Nvidia chips into US data centers: Nvidia sends chip designs to foreign contractors such as TSMC, which manufacture and package the chips; Nvidia sells the chips to foreign server makers (OEMs or ODMs); and those server makers sell finished servers to US data-center owners such as Microsoft.
Its hypothetical example has Nvidia paying contractors $10,000 per chip, server makers paying Nvidia $40,000 per chip, and US data-center owners paying $50,000 per chip installed in a server. Nvidia's design adds $30,000 and foreign firms add $20,000 ($10,000 each for the contractor and the server maker). The report sets out what the accounts should and do record:
| Should record | Actually record | |
|---|---|---|
| Domestic spending (US data-center owners) | +$50k | +$50k |
| Exports (Nvidia's value-add) | +$30k | $0 |
| Imports (Nvidia plus foreign value-add) | −$50k | −$50k |
| Total domestic value-add | +$30k | $0 |
In the report's summary: "Nvidia's value-add to imports is counted, but its value-add to exports is not." Nvidia's design spending is recorded as domestic R&D investment when incurred; what is missing is the ongoing income that investment produces. The report quotes the BEA as stating that information on sales of transformed goods sent from the contractor's country to a third country "is not collected on US customs documentation, as the goods do not cross the US border."
Size of the gap
The proxy is Nvidia's US-booked operating income: worldwide operating income, excluding one-off H20 inventory charges and reversals, multiplied by the US-booked share for each year, with the previous year's share extended to the current fiscal year and fiscal quarters calendarized by days of overlap. The authors state the proxy is not exact and that the direction of error is unclear: some chips may enter the US without the profit margin embedded (for example if Nvidia imports them at cost), while operating income is slightly conservative because it also deducts US-based coordination and shipping costs.
On this basis, annualized US GDP growth estimates for recent quarters have been underestimated by about 0.3 percentage points; the report sets this against an estimated 1.5% growth rate for the second quarter of 2026. Extrapolating the exponential growth of Nvidia's operating income, it projects that the growth-rate underestimate could reach almost two percentage points by the end of 2028 (1.78 percentage points by the fourth quarter of 2028 in its chart). The report notes that the size of the underestimate has risen less consistently than Nvidia's profits, because it depends on the absolute growth of those profits relative to the absolute growth of GDP.
For scale, it states that Nvidia made around $120 billion in US pre-tax income between January 2025 and January 2026, and attributes a measurement gap to Nvidia averaging about $30 billion per quarter throughout 2025.
Which firms are affected
The report states that the gap arises only when a US factoryless manufacturer (1) has its product manufactured abroad, (2) sells it abroad, (3) after transforming it, and (4) without explicit compensation for its IP. Apple's iPhone is given as a contrast: phones sold in the US are marked up after import, so only the payment to Foxconn counts as an import, and foreign sales typically go to foreign Apple affiliates that return the markup as a recorded IP export. The conditions apply to other fabless chipmakers, but the report says Nvidia's share of the chip design market and its growth are large enough that extending the analysis to the top five US fabless chipmakers (with Qualcomm, AMD, Marvell and Broadcom charted) would barely change the estimate.
As a preliminary check across the BEA's 414 detail-level industries, the authors used a "ChatGPT 5.6 Sol agent" to analyze each category and spot-checked its estimates, which attributed a $140 billion gap to semiconductor and related device manufacturing against roughly $4 billion across all other industries combined. The report states that the $140 billion is consistent with its own estimates, and that the $4 billion "should not be taken as a reliable quantitative estimate" but offers some evidence the pattern is currently concentrated in semiconductors.
GDP and GDI
The report notes that Nvidia's US value-add does appear in gross domestic income (GDI), through US wages, profits and taxes, but that GDI is arguably mismeasured in larger ways, for example through underreported income, and was 0.9% lower than GDP in the first quarter of 2026. It describes a persistent 0.3% underestimate of the GDP level as a large issue, and in its view probably the largest measurement issue attributable to a single firm, but within the range national accounts already contend with. Because the gap has grown quickly, it affects measured growth, not only the level: published GDP growth over the past year has been about 0.3 percentage points too low, and will remain biased downward as long as the share of fabless chipmakers in US output keeps rising.
Categories checked
The authors reviewed each national-accounts category in which the value-add might plausibly be recorded and report that it appears in none:
- General goods exports. Census Bureau trade data are built from customs records of goods crossing the US border, and Nvidia sells chips while they are still overseas. Exports of computers and peripherals did spike, from around $50 billion in 2023 to over $150 billion by 2026, but imports rose from around $150 billion before 2024 to nearly $600 billion in 2026; the report attributes most export growth to re-exports of foreign-made components sent to countries such as Mexico for server assembly and reimported at higher value, citing Foxconn's work for OpenAI's Stargate project.
- IP exports. Nvidia does not sell or license its designs to contractors; it pays TSMC to manufacture. A $45 billion rise in US IP exports from 2023 to 2025 is not attributable to Nvidia, since around 65% of it came from tax havens such as Ireland, Switzerland and Luxembourg rather than Taiwan.
- Other service exports. Nvidia's overseas sales are goods, not services, and computer-related service exports show no spike around 2024.
- Merchanting. Goods under merchanting must be resold in the same condition as purchased; Nvidia's dies go through assembly, testing and packaging abroad. The BEA records under $1 billion a year in net exports of goods under merchanting.
- Taiwanese accounts. Taiwan follows the IMF's BPM-6 ownership-transfer rules, so only payments to TSMC and other Taiwanese firms enter Taiwanese GDP, and Nvidia's sale to a foreign ODM is recorded as a transaction between non-residents. The value-add therefore appears in neither country's GDP, nor in gross world product. A footnote argues it is highly unlikely a Taiwanese Nvidia branch is capturing the markup, since total direct-investment income from Taiwan was $12 billion in 2025, against the roughly $100 billion that would be required.
International guidelines and proposed fixes
The BEA documented in 2014 that, under the then-updated BPM-6, products made abroad by US factoryless producers and sold abroad should be counted as US goods exports, but it made "no plans to introduce changes to core US economic accounts," citing data-collection difficulties and following objections to the Office of Management and Budget that counting overseas-made goods as US exports would interfere with efforts to promote manufacturing on US soil. As of June 2026, the report states, the BEA's International Economic Accounts: Concepts and Methods manual still does not classify such sales as exports.
The IMF's BPM-7, released in 2025 with a parallel UN national-accounts update, states more explicitly that factoryless producers such as Nvidia should be classified as US manufacturers and their Taiwan-based sales as US merchandise exports. The report outlines an alternative it calls conceptually cleaner: recording fabless chipmakers' markups on foreign sales as IP exports to the country of manufacture. It acknowledges this would depart from international standards and from the logic of an expenditure-side GDP measure, and would be closer to trade-in-value-added accounting, but would avoid the political objection to counting outsourced production as US manufacturing. Either change could take years; the report notes the BEA took five years to implement the last UN guideline update.
Relation to other estimates
The report positions itself against analyses that found AI capital spending large but its net GDP contribution small because of imports, quoting a 2025 J.P. Morgan Asset Management note (AI-driven capital spending giving "an impressive 1 percentage point (ppt) boost to GDP growth in the second quarter of 2025 alone" while imported technology goods subtract from GDP) and an EY analysis describing generative AI's net contribution as modest after accounting for trade in computer equipment. Its argument is that much of the measured surge in computer-equipment imports is re-imported US value-add. This bears on other estimates of AI's macroeconomic contribution: a St. Louis Fed estimate that AI-related investment added 0.97 percentage points to growth in the first three quarters of 2025 (Anecdotes Everywhere, Evidence Almost Nowhere (Steve Newman, July 2026)), and a Goldman Sachs executive's remark, quoted by Luiza Jarovsky, that AI investment made "basically zero" contribution to US GDP growth in 2025 (Lack of Public Trust Could Kill AI (Jarovsky, February 2026)). The Epoch report does not dispute the import-subtraction arithmetic those estimates rely on; it argues that the export side is missing.
The report's own caveats are that the proxy is approximate, that the BEA cannot measure the relevant transactions either, and that the 2028 figure is an extrapolation of Nvidia's operating-income growth rather than a forecast of chip demand.
Provenance
Captured 2026-09-27 from Epoch AI's publication page, including all nine footnotes and author notes; chart values are recorded only where the report's text or image descriptions state them. Single source with no independent replication of the estimate; confidence is medium. Corroboration of the report's authenticity and circulation (Epoch AI's X account, a J.P. Morgan Asset Management note citing the 0.3 percentage point estimate, and The Signal newsletter of August 26, 2026) was recorded at queue time.
Relationships
- supports: AI Infrastructure Capex — argues measured GDP understates the domestic value of the hardware buildout
- related: AI and Productivity — a measurement argument bearing on the macro-skeptic reading of AI's GDP contribution
- related: Anecdotes Everywhere, Evidence Almost Nowhere (Steve Newman, July 2026), Lack of Public Trust Could Kill AI (Jarovsky, February 2026) — earlier estimates of AI's GDP contribution
- related: Nvidia & TSMC — AI Compute Infrastructure — the firm and supply chain the gap concerns
- related: Epoch AI — publisher
- related: Phil Trammell, Anson Ho — authors; Erik Brynjolfsson — Trammell's Stanford Digital Economy Lab collaborator
- related: Semiconductor Supply Chain, AI Bubble Debate