The AI-GDP measurement gap

The claim that AI spending accounted for 39-92% of U.S. GDP growth in 2025 is being substantially revised downward by economists at Goldman Sachs, Morgan Stanley, JPMorgan Chase, and the St. Louis Fed. The core issue: GDP measures domestic production, and a large share of AI hardware (chips, servers, networking equipment) is manufactured overseas. Once import-adjusted, AI's net contribution drops to somewhere between 0 and 0.5 percentage points of the 2.2% total growth.

For enterprise technical leadership, the practical takeaway is not that AI investment is wasted -- it is that the macro narrative inflated AI's measurable economic footprint relative to consumer spending, which remained the actual primary growth driver. This matters for budgeting, vendor negotiations, and board-level discussions about AI ROI expectations.

The range is wide but the direction is consistent: every import-adjusted analysis finds AI's net GDP contribution materially lower than the unadjusted figures that dominated 2025 financial media coverage. The Q1 2025 distortion -- when AI hardware investment spiked while the broader economy contracted at -0.3% -- created a misleading "AI is all that matters" narrative that persisted through the year.

The accounting dispute centres on a structural feature of the AI supply chain: chip design is concentrated in the U.S., but fabrication, packaging, and assembly occur almost entirely overseas -- principally in Taiwan, Mexico, and Vietnam. When a U.S. company purchases a $4.5M AI server rack, a substantial share of the expenditure flows to foreign producers and is subtracted from GDP as imports.

This differs from most categories of business investment, where the import share is lower and analysts rarely strip it out. AI hardware is unusual because the import fraction is so large. As Politano notes, it is rare to perform this adjustment for individual sectors, but AI's import intensity makes it a special case.

A further complication: AI's GDP contribution appears larger on the income side than investment alone would suggest. Once data centres become operational, they generate service revenue (cloud compute, API subscriptions) captured in sectoral value-added accounts. The CEPR/Bank of Italy analysis emphasises this distinction -- the investment channel and the service-flow channel are both real, but they operate on different timescales and are measured differently.

1. ROI expectations should be calibrated to micro, not macro, evidence. The macro GDP debate is largely irrelevant to whether a specific organisation's AI investment generates returns. The confusion between "AI is boosting the economy" and "our AI investment is delivering value" conflates two different claims. Enterprise leaders should focus on internal productivity metrics, cost displacement, and time-to-decision improvements rather than citing macro GDP figures in business cases.

2. The productivity J-curve is the more useful frame. Multiple sources in this debate reference Brynjolfsson, Rock, and Syverson's productivity J-curve hypothesis: general-purpose technologies show delayed productivity gains because complementary investments in organisational change, workflow redesign, and skills development must accumulate before the payoff materialises. The St. Louis Fed analysis notes that AI-related investment has already surpassed dot-com era levels as a share of GDP. Whether that investment converts to productivity depends on the People/Processes axis -- the organisational complement.

3. Vendor narratives will selectively cite the convenient numbers. The existence of a 0-to-92% range for "AI's share of GDP growth" means both AI optimists and sceptics can find a credible economist to support their preferred story. Enterprise buyers should expect vendors to cite the higher, unadjusted figures. The import-adjusted range (0-0.5 pp, or roughly 0-25% of growth) is the more defensible reference point.

4. Consumer spending, not AI capex, remains the primary growth driver. The U.S. economy grew 2.2% in 2025, slowing from 2.4% in 2024. Consumer spending was the dominant contributor across the full year. For organisations whose revenue depends on the broader economy, this is the more relevant macro signal: the expansion has diversified support, and is not solely dependent on continued AI capex.

This signal directly informs the Q1 2026 research agenda. The macro-level measurement debate mirrors what we observe at the enterprise level: organisations struggle to attribute value to AI investments because the benefits are diffuse, delayed, and entangled with complementary changes.

The CEPR framework distinguishing between the investment channel (capital spending) and the service-flow channel (operational productivity) provides a useful structuring device for our practitioner interviews. The ethnographic research should probe how organisations are (or are not) measuring the service-flow side -- the ongoing operational value -- rather than fixating on the upfront spend.

The St. Louis Fed's comparison to the dot-com era is also instructive: in 2000, IT-related categories contributed 0.81 pp to GDP growth; in the first three quarters of 2025, AI-related categories contributed 0.97 pp (gross). The scale is comparable, and the historical precedent suggests a multi-year lag before productivity effects become measurable at the macro level.