Global · AI value
What AI is worth so far: big investment, a large consumer surplus, small measured productivity
Investment in information-processing equipment and software added an average of 0.85 percentage points to US real GDP growth across the quarters of 2025, 41% of the 2.1% average (BEA, annualised quarterly rates); +0.52pp of 2.2% in 2026 Q2. Net of imported computers, our rough estimate is -0.15pp for 2025. Stanford puts US consumer surplus from generative AI at $183bn a year (Jul 2026).
Updated 11 October 2026 · data build 2026-10-10 · how we know · corrections
IT and software investment supplied 41% of US growth in 2025; on our rough estimate, computer imports offset about as much
Contributions to US real GDP growth, percentage points at annual rates. Gross: BEA's lines for information-processing equipment and for software investment. Offset: our estimate of the drag from imports of computers, peripherals and parts. Measured
| Quarter | Real GDP growth | IT equipment | of which computers | Software | IT + software | Computer imports (est.) | Net (est.) |
|---|---|---|---|---|---|---|---|
| 2026-Q2 | 2.2% | +0.19pp | -0.09pp | +0.33pp | +0.52pp | +0.39pp | +0.91pp |
| 2026-Q1 | 2.5% | +0.79pp | +0.62pp | +0.6pp | +1.39pp | -1.59pp | -0.2pp |
| 2025-Q4 | 0.2% | +0.67pp | +0.55pp | +0.05pp | +0.72pp | -1.6pp | -0.88pp |
| 2025-Q3 | 3.9% | +0.3pp | +0.39pp | +0.08pp | +0.38pp | -0.29pp | +0.09pp |
| 2025-Q2 | 4% | +0.37pp | +0.52pp | +0.6pp | +0.97pp | -0.36pp | +0.61pp |
| 2025-Q1 | 0.1% | +0.87pp | +0.47pp | +0.45pp | +1.32pp | -1.75pp | -0.43pp |
| 2024-Q4 | 2.4% | -0.06pp | -0.11pp | +0.13pp | +0.07pp | +0.21pp | +0.28pp |
| 2024-Q3 | 3.2% | +0.37pp | +0.27pp | +0.05pp | +0.42pp | -0.58pp | -0.16pp |
| 2024-Q2 | 3.6% | +0.16pp | +0.15pp | +0.09pp | +0.25pp | -0.46pp | -0.21pp |
| 2024-Q1 | 1.4% | -0.1pp | +0.11pp | +0.24pp | +0.14pp | -0.35pp | -0.21pp |
| 2023-Q4 | 3.6% | +0.21pp | +0.15pp | +0.21pp | +0.42pp | -0.19pp | +0.23pp |
| 2023-Q3 | 4.5% | +0.02pp | -0.02pp | +0.12pp | +0.14pp | -0.05pp | +0.09pp |
| 2023-Q2 | 2.2% | -0.06pp | 0pp | +0.07pp | +0.01pp | -0.04pp | -0.03pp |
| 2023-Q1 | 3.4% | -0.02pp | -0.13pp | +0.02pp | 0pp | +0.13pp | +0.13pp |
| 2022-Q4 | 2.7% | -0.23pp | -0.15pp | +0.27pp | +0.04pp | +0.23pp | +0.27pp |
| 2022-Q3 | 2.7% | +0.14pp | +0.1pp | +0.19pp | +0.33pp | -0.04pp | +0.29pp |
| 2022-Q2 | 0.6% | -0.05pp | -0.08pp | +0.3pp | +0.25pp | +0.2pp | +0.45pp |
| 2022-Q1 | -1% | +0.2pp | +0.07pp | +0.47pp | +0.67pp | -0.21pp | +0.46pp |
Source: BEA NIPA Table 1.5.2 (contributions to real GDP growth) · BEA NIPA Tables 4.2.1 and 4.2.5 (computer imports) · BEA NIPA flat file · as of 2026-10-11 · series Y034RY, B935RY, B985RY, A191RL, A191RC, B852RC, B852RL; next: Q3 advance estimate, 29 Oct 2026 · Compiled by StackPath
In 2025 Q1 and Q4 GDP barely grew while computer imports surged (in Q1 probably ahead of tariffs: our reading), which is why the net turns negative there. Quarterly figures are revised twice after the advance estimate.
Consumers: about $183bn a year of value that GDP doesn't count
Consumer surplus: the gap between what people would need to be paid to give up generative AI and what they pay for it. US only. Estimate
| Estimate | Value | Period | Method |
|---|---|---|---|
| US consumer surplus from generative AI | $183bn a year | Jul 2026 | Survey: what users would need to be paid to give up generative AI tools for a month, times the estimated number of US users, annualised. |
| US consumer surplus from generative AI | $116bn a year | Jul 2025 | Same tracker, a year earlier. |
Source: Stanford Digital Economy Lab, GDP-B surplus observer · as of 2026-10-11 · survey estimate · Compiled by StackPath
Stated-preference surveys ask about hypothetical payments, and answers of this kind tend to run high. The $183bn is 12 × $120 × 127 million users.
The official price of hosting rose 0.3% in a year
US producer price index for data processing, hosting and related services, monthly. Index 125.2 in 2026-08 (preliminary). Measured
Source: BLS PPI PCU518210518210 · as of 2026-10-11 · PPI: data processing, hosting and related services (NAICS 518210); 8.6% over five years · Compiled by StackPath
AI's price falls show up per token, not in this index, which covers all hosting and data processing. If AI were cutting what firms pay for computing overall, this index would be falling; it isn't.
Productivity: strong in task trials, small or zero in whole-economy data
7 studies, one row each: what was measured, how, and the effect as the authors state it. 3 positive, 1 mixed, 1 negative, 1 null, 1 small. Reported
| Study | Design | Sample | Effect | Level |
|---|---|---|---|---|
| Generative AI at Work (customer support) | Staggered roll-out at one firm (natural experiment) · Brynjolfsson, Li, Raymond, Quarterly Journal of Economics, 2025 | 5,179 agents | +14% issues resolved per hour on average; +34% for novice and low-skilled workers, little effect for the most experienced | task / firm |
| Experimental evidence on the productivity effects of generative AI (writing tasks) | Randomised online experiment · Noy, Zhang, Science, 14 Jul 2023 | 453 college-educated professionals | Time taken down 40%; output quality up 18% | task |
| Navigating the Jagged Technological Frontier (BCG consultants) | Randomised field experiment · Dell'Acqua, McFowland, Mollick and others, HBS working paper 24-013 | 758 consultants | Inside the AI's competence: 12.2% more tasks, 25.1% faster, about 40% higher quality. On a task outside it, AI users were more often wrong | task |
| Early-2025 AI and experienced open-source developers | Randomised controlled trial on real issues · METR, METR report, 10 Jul 2025 | 16 developers, 246 issues | Developers took 19% longer with AI tools. METR said in Feb 2026 the result is out of date for current tools | task |
| Still Waters, Rapid Currents (Danish workers) | Survey linked to Danish administrative earnings and hours records; difference-in-differences · Humlum, Vestergaard, NBER w33777, revised Mar 2026 | Danish workers and workplaces (sample size in the paper, not the abstract) | Precise null effects on earnings and recorded hours; rules out effects above 2% two years after ChatGPT's launch | economy (admin data) |
| Real-Time Population Survey: time saved | Nationally representative online survey, US adults 18-64 · Bick, Blandin, Deming, Management Science 2026; tracker May 2026 | not stated on tracker | 6.3% of all work hours involve generative AI; time saved equals 2.2% of all work hours | economy (survey) |
| The Simple Macroeconomics of AI | Task-based model calibrated to exposure and cost-saving estimates · Acemoglu, NBER w32487; Economic Policy 2025 | model | No more than 0.66% total factor productivity over 10 years; under 0.53% after adjusting for hard-to-learn tasks | economy (model) |
Source: Brynjolfsson, Li, Raymond · Noy, Zhang · Dell'Acqua, McFowland, Mollick and others · METR · Humlum, Vestergaard · Bick, Blandin, Deming · Acemoglu · as of 2026-10-11 · Compiled by StackPath
Our reading. The money is real and the value to users is large, but the economy-wide productivity gain is not yet visible in the data. Much of the measured GDP boost is spending on equipment that is largely imported, so it adds less to US output than headlines suggest. Europe has no comparable quarterly line: its AI spending shows up mainly as imported servers (see server imports) and in usage (see adoption). Ireland's GDP is heavily shaped by multinational accounting, so this reading wouldn't transfer there.
BEA contributions measure investment spending on computers, other information-processing equipment and software, whatever it is for. They are not 'AI's contribution': they include ordinary IT, and miss AI spending booked elsewhere (data-centre buildings, power equipment). Much of the equipment is imported. Imports subtract from GDP, so the gross contribution overstates what AI investment adds to US output. Our import adjustment covers computers, peripherals and parts only (not chips or networking gear), using a first-order formula; it is an estimate, not a BEA figure. Quarterly contributions are at annual rates and swing with timing of deliveries and revisions; the advance estimate for each quarter is revised twice. Consumer surplus is a stated-preference survey estimate (what people say they would need to be paid to give up the tools). It is not in GDP, and survey answers about hypothetical payments are known to run high. Study effects are from different settings and tasks; task-level gains do not add up to economy-wide gains.
Questions people ask
- How much is AI adding to US GDP growth?
- BEA doesn't measure AI separately. Investment in information-processing equipment and software, the closest line, added 0.85 percentage points to US growth on average in 2025 (41% of growth) and +0.52pp in 2026 Q2. Much of the hardware is imported, which subtracts from GDP: our rough estimate of the net effect in 2025 is -0.15pp, with a wide margin either way (see the page for why).
- What is the consumer surplus from generative AI?
- About $183bn a year in the US in Jul 2026, from 127 million users who would need about $120 a month to give it up (Stanford Digital Economy Lab). That is about 0.56% of US GDP, and it is not counted in GDP.
- Does AI raise productivity?
- In controlled studies of specific tasks, often yes: 14% more customer-support issues resolved per hour, 40% less time on writing tasks. Across whole economies the measured effect is still small: Danish administrative data show no change in earnings or hours, and US workers report time savings equal to about 2% of work hours. One trial found experienced developers 19% slower with early-2025 tools.
- Is cloud computing getting cheaper because of AI?
- Not in the official US price index for data processing and hosting: it was 0.3% higher in 2026-08 than a year earlier and 8.6% higher than five years before (BLS). Prices per AI token have fallen fast, but they are a small part of that index.
- What do sceptics say about AI's economic value?
- Daron Acemoglu's model puts the total factor productivity gain from AI at no more than 0.66% over ten years. Humlum and Vestergaard rule out effects on Danish earnings or hours above 2% two years after ChatGPT. The GDP contribution from investment is spending, not proof of returns.