The AI Buildout and the Economy: Publicly Available Data to Assess AI's Impact

July 17, 2026

The AI Buildout and the Economy: Publicly Available Data to Assess AI's Impact¹

Paul E. Soto, Mason Thieu, and Jeffrey S. Allen

Introduction

This note presents publicly available indicators that can help researchers and policymakers track the evolution of the generative AI buildout and its potential impact on the economy on a timely basis. We organize the indicators into three categories: capabilities and costs; firm investment and adoption; and productivity and labor. These categories map closely to the sequence often associated with general-purpose technologies (GPT) and the economy: capability improvements and cost declines precede broad firm adoption and investment, which in turn precede measurable aggregate productivity gains and labor market outcomes (Bresnahan and Trajtenberg, 1995; Lipsey et al., 2005). Whether AI qualifies as a GPT is examined in Baily et al. (2025). Here, we propose a set of public, often official, indicators that researchers can draw on to assess economic developments resulting from this technology as they unfold.

A growing literature has examined the impact of generative AI throughout the economy, focusing on adoption (Bick et al., 2026; Bonney et al., 2024; Yotzov et al., 2026), investment (Brandsaas et al., 2025), employment (Brynjolfsson et al., 2025; Crane and Soto, 2026), and productivity (Baily et al., 2025; Acemoglu, 2025). Our goal is to provide a roadmap for monitoring whether the economic effects of AI remain concentrated in investment-led growth or whether transformative effects are starting to materialize in the labor market and aggregate productivity. We also expect that monitoring these indicators could shed light on whether the current gap between financial markets, which have been highly responsive to changes in the AI narrative, and aggregate output and labor market data, which currently show limited signs of broad-based transformation, narrows. As of 2026, much of the evidence points to an economy that is reorganizing around this new technology, with real effects concentrated on certain areas of the economy.

For each of the current topics, we summarize recent trends in the indicators below, and the data file attached to this note compiles the relevant series with their public sources. As new data become available, researchers can use and extend this framework to monitor developments in the AI buildout.

Capabilities and Costs

A prerequisite for any transformative technology to generate broad-based economic impacts is that it can facilitate or perform economically meaningful work in a cost-effective manner. Measuring what AI can actually do has proven quite difficult. Early benchmarks based on standardized tests (the SAT, Bar Exam, MMLU, etc.) were quickly saturated, with subsequent expert level reasoning benchmarks (e.g. GPQA and ARC-AGI) following a similar pattern. Some benchmarks might not map cleanly onto real-world work, since strong benchmark performance does not necessarily translate into the ability to complete actual on-the-job tasks. Model Evaluation & Threat Research (METR) measures agentic task-completion horizons (Figure 1), which indicates the time it takes a human expert to perform a task that a respective AI model could perform at a specified success rate. Currently, this benchmark focuses on software engineering and machine learning tasks, as these are the domains frontier labs are automating first, where impacts should appear earliest. As of mid 2026, this horizon has been doubling roughly every several months (Kwa et al., 2025). At this pace, a full workweek of agentic tasks could be achieved within years, although whether this pace is sustained and whether it expands to other work remains to be seen.

It is important to note that measures such as task-completion horizon measure technical feasibility, but not cost-effective deployment. Whether deployment is cost-effective hinges on two distinct costs. The first is a fixed cost of adjustment. The jump from completing an isolated task to substituting for complete workflows requires reliable and often costly integration into firm-specific systems. Such adjustment costs are not directly observable and are difficult to quantify. The second is the marginal cost of inference, which is more readily observable. The cost of graphical processing units (GPUs), important inputs for training and running AI models at scale, has declined precipitously in recent years (Figure 2), following a longer running trend resulting from algorithmic and hardware efficiency gains.

GPUs, however, are only one input. Memory is another, with high-bandwidth memory in particular becoming a binding constraint as models scale. Korean semiconductor export prices (Figure 3) serve as a useful proxy for such input costs, given that Samsung and SK Hynix account for roughly 70 percent of global DRAM supply and are the dominant share for high-bandwidth memory production. Whether expanding semiconductor supply, discussed in the next section, can relieve demand-constrained cost pressures will be an important determinant of the pace of how quickly AI becomes cost-competitive with labor.

Input costs ultimately are reflected in the prices firms pay to use AI, typically quoted per token. Such prices embed not only compute and memory costs but also include the providers markups. Meaningful token price comparisons require quality adjustments since more capable models may cost more per token but require fewer tokens per task, making raw per-token comparisons misleading. Because pricing is largely proprietary and posted rates might not reflect enterprise contracts, we do not publish a series in this note, but the margin is worth monitoring as the binding constraint on adoption shifts from building capacity to paying for this intermediate input.

The accompanying data file includes current readings on METR task horizons; unit costs of compute for select chips (EpochAI, 2025); and Korean export prices for semiconductors and DRAM.

Firm Investment and Adoption

High levels of investment preceding measurable productivity gains are characteristic of GPTs. Measuring the scale of investment activity in the AI buildout requires inferring AI-related spending from various sources that each have certain attribution and scoping challenges. Among these, hyperscaler capital expenditure (Figure 4), which includes investments in data center structures, land, servers, and networking equipment, has arguably become one of the main demand-side proxies for AI infrastructure. While the trend clearly reflects the AI investment boom, the level is hard to interpret because it includes non-AI capex.² Additionally, because the hyperscalers increasingly lease data center capacity, the headline capex figures could progressively understate total AI-related infrastructure investment.

The Census Bureau's data center construction series and the BEA's equipment investment data from the National Income and Product Accounts (NIPA) capture complementary layers of the AI buildout. The former (Figure 5a) is a leading indicator of anticipated computing demand because physical structures are built before they are filled with computing equipment. A major limitation of the series on its own is that physical structures are not the dominant source of value in data centers. The NIPA equipment investment data, especially "computers and peripheral equipment" (Figure 5b), help fill this gap by capturing spending on the servers and other computing equipment that fill data centers, but they include substantial non-AI activity as well, such as normal office computers. An AI-specific component can be approximated by extrapolating a pre-2023 baseline and treating the deviation as the AI buildout (see, for example: Brandsaas et al., 2025), but such estimates may become less reliable over time as other trends contaminate the counterfactual. Triangulating the three investment indicators can help track the trajectory of the AI buildout, which is still growing rapidly. Meaningful deceleration could signal infrastructure demand having been met or a downward revision in expected return on investment, though other factors such as financing conditions could also play a role.

Rising production of semiconductors (Figure 6) alongside capex would be an important indicator of supply expanding to meet AI-driven demand, consistent with continued cost relief in compute terms. Any divergence between investment and capacity expansion, for example, with capex rising while construction starts plateauing, might be an early sign of supply constraints or a shift towards software and operating expenditures as the buildout matures.

Measuring precisely how much recent GDP growth has been boosted by AI-related investment poses a greater challenge given the lack of a dedicated line item in the national accounts and the high import-content of the equipment underlying the buildout. Figure 7 shows one approximation, combining contributions from software, data centers, power facilities, and computer and peripheral equipment while adjusting for net exports. The Appendix provides details on the calculation methodology. Though the series makes strong assumptions about which categories to include and how to account for import shares, it could serve as a proxy for the AI buildout's contribution to GDP growth. From 2025 through the first quarter of 2026, the proposed set of AI-related components have contributed meaningfully to quarterly GDP growth, with software and computer and peripheral equipment the largest positive contributors. The net effect (black line) varies considerably across quarters, as the drag from net exports of computer, peripherals, and parts offsets much of the gross investment in quarters where imports rose sharply.

Adoption trends provide a mixed picture of the ubiquity of AI, since survey methods of uptake can lead to differing conclusions (Crane et al. 2025). The Census Bureau's Business Trends and Outlook Survey (BTOS) provides a relevant firm-level measure of AI uptake, with biweekly updates. Figure 8 shows uptake trending upwards, with a generally positive association between adoption and firm size. Such headline figures do not reflect usage intensity, which surveys suggest remains shallow even where reported adoption is broad (Yotzov et al., 2026; Baslandze et al., 2026). Nonetheless, tracking BTOS biweekly provides a representative and early signal of firm-level AI deployment, a necessary step before aggregate effects become apparent.

The accompanying data file includes data on capital expenditure of major AI-related technology companies; Census Private Construction of Office Data Centers; BEA NIPA Computers and Peripheral Equipment; Industrial Production of Semiconductor and Other Electronic Components; GDP contributions from AI-related spending components; and BTOS adoption rates by firm size.

Productivity and Labor Impacts

Micro-level experiments consistently find productivity gains from AI tools (Brynjolfsson, Li, and Raymond, 2025; Cui et al. 2026). The broader question is whether these gains translate into aggregate productivity growth.

If AI is driving aggregate productivity, it most likely appears first and foremost in high-exposure sectors. Figure 9 tracks sectoral productivity by AI exposure level. While sectors with more exposure appear to have higher labor productivity growth, we see that productivity trends across all three levels have been relatively consistent over time, suggestive of micro-level productivity gains not adding up in aggregate.

One possible reason for the divergence is that micro-level experiments typically measure task-level productivity, like the speed a programmer generates code, rather than job-level or firm-level output. A 10% improvement on a task does not necessarily lead to proportional gains for a firm if adjustment costs or other bottlenecks lie elsewhere in the production process and erode the upstream productivity gains. Another reason could be that experiments finding gains for information-sector roles leave open the question what AI could mean for productivity in other sectors where work looks very different. As Figure 9 suggests, productivity growth is not homogenous across industries, and AI uptake in low exposure industries such as construction will not necessarily lead to the same productivity gains as in high exposure sectors such as information.

A widening divergence between high and low exposure sectors could be early evidence of AI-driven productivity gains, though under the caveat that AI-exposed sectors may have ex-ante different productivity trends unrelated to AI. The absence of a large aggregate signal as of 2026 would not necessarily invalidate future productivity impacts, given the shallow adoption outlined in the previous section. Historically, measured productivity gains from GPTs lag investment by years (Brynjolfsson et al, 2021), and the realized gains from genAI may take time to appear even if its long-run potential is significant (Baily et al., 2025).

Beyond these timing dynamics, AI's productivity effects are inherently hard to measure. The gains may be misattributed between capital deepening and total factor productivity, and complementary intangibles are often not completely captured in the data (Corrado et al., 2021; Byrne, 2022). In services, where output is often inferred from revenue, price declines can distort measured output. Attribution can be further complicated since AI's contribution to output may be spread across sectors that use it as an intermediate input rather than appearing as final output.

The labor market shows more immediate but also uncertain impacts of AI. Aggregate unemployment has remained moderate by historical standards, but this headline number masks compositional shifts that might be related to AI. Figure 10a shows unemployment by age group over time. Youth unemployment is particularly important to track. If AI substitutes for a set of tasks typically done for entry-level workers, it might not only displace junior workers, but it can impair on-the-job learning that younger workers acquire, potentially impacting labor market outcomes in the longer run. We also track labor force participation rates across younger cohorts against prime-age workers to capture whether individuals are withdrawing from the labor market entirely rather than moving between jobs (Figure 10b).

Eloundou et al. (2024) identify information processing employment as most exposed to AI based on task descriptions. The Anthropic Economic Index (Handa et al., 2025), which maps actual Claude usage to job tasks, points in a similar direction with AI usage concentrated in software development, technical writing, and analytical work. Crane and Soto (2026) find that programming-intensive occupations, which involve highly LLM-exposed tasks, are concentrated in professional and technical services (NAICS 54) rather than the information sector (NAICS 51) alone. Both sectors report the highest AI adoption rates across sectors in the BTOS and are likely to reflect materialized effects first. As seen in Figure 11, layoffs and discharges within this sector can provide a signal of AI-driven displacement in white collar work, while job openings capture potential reinstatements to offset displacement elsewhere (Acemoglu and Restrepo, 2019). Early evidence suggests that AI is impacting younger workers via slower hiring rather than outright layoffs (Lichtinger and Hosseini Maasoum, 2025). Wage growth and job separation rates across sectors can provide further information on whether labor market tightness or slack is emerging more broadly.

The accompanying data file tracks: sectoral productivity across High vs Medium vs Low exposure industries; aggregate unemployment and labor force participation rates alongside rates for young workers (20-24 years old) and prime-age workers (25-54); layoffs and discharges as well as job openings in aggregate and within the Information sector.

Conclusion

The indicators presented in this research note, a snapshot of which are available in the accompanying data file, can provide a starting point for assessing the AI transition using publicly available and timely data. Overall, the evidence as of the publication of this note is consistent with a buildout phase rather than the onset of broad-based displacement. Capabilities are advancing rapidly, investment continues to boom, and adoption is rising. While some highly exposed sectors show relatively strong productivity, labor market impacts remain concentrated and have not yet broadened in the aggregate. A sustained change in labor market dynamics among AI-exposed groups, capex being associated with productivity gains, or a widening gap between high and low AI exposed sectoral productivity growth might signal a shift towards broader macroeconomic change. Tracking these indicators can serve as one avenue for determining where we are in such a transition.

References

Acemoglu, D. (2025). The simple macroeconomics of AI. Economic Policy, 40(121), 13-58.

Acemoglu, D., & Restrepo, P. (2019). Automation and new tasks: How technology displaces and reinstates labor. Journal of economic perspectives, 33(2), 3-30.

Baily, M., Byrne, D., Kane, A., & Soto, P. (2025). Generative AI at the Crossroads: Light Bulb, Dynamo, or Microscope?. arXiv preprint arXiv:2505.14588.

Baslandze, S., Edwards, Z., Graham, J., McClure, T., Meyer, B. H., Sparks, M., Waddell, S. R., & Weitz, D. (2026). Artificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executives. NBER Working Paper 34984.

Bick, A., Blandin, A., & Deming, D. J. (2026). The rapid adoption of generative AI. Management Science.

Bonney, K., Breaux, C., Buffington, C., Dinlersoz, E., Foster, L. S., Goldschlag, N., ... & Savage, K. (2024). Tracking firm use of AI in real time: A snapshot from the Business Trends and Outlook Survey (No. w32319). National Bureau of Economic Research.

Bontadini, F., Corrado, C., Haskel, J., & Jona-Lasinio, C. (2026). Is Software Eating the World? Measuring the Progress and Diffusion of AI. NBER Working Paper.

Brandsaas, E. E., Garcia, D., Kurtzman, R. J., Nichols, J., & Zytek, A. (2025). Estimating Aggregate Data Center Investment with Project-level Data. FEDS Working Paper. Washington: Board of Governors of the Federal Reserve System.

Bresnahan, T. F., & Trajtenberg, M. (1995). General purpose technologies 'Engines of growth'?. Journal of econometrics, 65(1), 83-108.

Brynjolfsson, E., Chandar, B., & Chen, R. (2025). Canaries in the Coal Mine?: Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab.

Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889-942.

Brynjolfsson, E., Rock, D., & Syverson, C. (2021). The productivity J-curve: How intangibles complement general purpose technologies. American Economic Journal: Macroeconomics, 13(1), 333-372.

Byrne, D. (2022). The Digital Economy and Productivity. FEDS Working Paper. Washington: Board of Governors of the Federal Reserve System.

Carpinelli, L., Natoli, F., & Taboga, M. (2026), Artificial Intelligence and the US Economy: An Accounting Perspective on Investment and Production. SSRN Working Paper.

Corrado, C., Haskel, J., & Jona-Lasinio, C. (2021). Artificial intelligence and productivity: an intangible assets approach. Oxford Review of Economic Policy, 37(3), 435-458.

Coyle, D., & Poquiz, J. L. S. (2025). Making AI Count: The Next Measurement Frontier, NBER Working Paper 34330.

Crane, L., Green, M., & Soto, P. (2025). Measuring AI Uptake in the Workplace, FEDS Notes. Washington: Board of Governors of the Federal Reserve System.

Crane, L. D., & Soto, P. E. (2026). AI and Coder Employment: Compiling the Evidence. FEDS Working Paper. Washington: Board of Governors of the Federal Reserve System.

Cui, K. Z., Demirer, M., Jaffe, S., Musolff, L., Peng, S., & Salz, T. (2026). The effects of generative AI on high-skilled work: Evidence from three field experiments with software developers. Management Science.

Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702), 1306-1308.

Epoch AI, 'Data on AI Chip Sales'. Retrieved from 'https://epoch.ai/data/ai-chip-sales'.

Handa, K., Tamkin, A., McCain, M., Huang, S., Durmus, E., Heck, S., ... & Ganguli, D. (2025). Which economic tasks are performed with AI? Evidence from millions of Claude conversations. arXiv preprint arXiv:2503.04761.

Hobijn, B., Mestieri, M., Werquin, N., & Zhang, J. (2025). Quarterly industry-level labor productivity data for the U.S. Economic Perspectives, No. 1. Federal Reserve Bank of Chicago.

Korinek, A., & McKelvey, P. (2026). Measuring the AI economy. Peterson Institute for International Economics Working Paper 26-9.

Kwa, T., West, B., Becker, J., Deng, A., Garcia, K., Hasin, M., ... & Chan, L. (2025). Measuring AI ability to complete long tasks. arXiv preprint arXiv:2503.14499.

Lichtinger, G., & Hosseini Maasoum, S. M. (2025). Generative AI as seniority-biased technological change: Evidence from US resume and job posting data. Available at SSRN.

Lipsey, R. G., Carlaw, K. I., & Bekar, C. T. (2005). Economic transformations: general purpose technologies and long-term economic growth. Oxford University Press.

METR. "Measuring AI Ability to Complete Long Tasks." METR (blog). https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/.

Yotzov, I., Barrero, J. M., Bloom, N., Bunn, P., Davis, S. J., Foster, K. M., ... & Wang, B. Z. (2026). Firm Data on AI (No. w34836). National Bureau of Economic Research.

Appendix

Measuring the effect of the AI buildout on GDP growth faces many challenges and is an ongoing area of work (for a detailed analysis, see Korinek and McKelvey, 2026; Bontadini et al., 2026; and Coyle and Poquiz, 2025). In particular, while AI-related investments are clearly rising, much of the underlying equipment is imported, so an analysis that focuses on gross AI investment and fails to account for net exports would overestimate the precise contribution of the AI buildout to GDP accounts. Additionally, while BEA NIPA tables describe net exports, they do not parse out imports/exports that go into consumption from imports/exports that go into investment, so factoring in the effect of net exports of AI-related goods risks capturing net exports of consumption goods.

Figure 7 offers a way, albeit imperfect, to address these concerns and to track the impact of AI investment on aggregate output growth. We show contributions from capital expenditure on software, data centers, power facilities, and computer and peripherals to annualized quarter-over-quarter GDP growth. To account for the import shares for computer and peripherals, we include the contribution from net-exports of computer, peripherals, and parts provided by the NIPA tables. To account for the share of net exports that go into investment, we also down-weight the nominal shares of imports and exports of nominal GDP in the calculation of net export's contribution to GDP growth.

Formally, we approximate contributions to annualized quarter-over-quarter GDP growth as

$$$$ C_t=\ \frac{N_{t-1}}{GDP_{t-1}^N}\cdot4\cdot\left(\frac{R_t}{R_{t-1}}-1\right)\cdot100\ +\frac{N_t}{GDP_t^N}\cdot\left[100\cdot\left(\left(\frac{GDP_t^R}{GDP_{t-1}^R}\right)^4-1\right)-4\cdot\left(\frac{GDP_t^R}{GDP_{t-1}^R}-1\right)\cdot100\right], $$$$

where for a given quarter $$t$$, $$C_t$$ is the category contribution, $$N_t$$ is the nominal series, $$R_t$$ is the real series, $$GDP_t^N$$ is nominal GDP, and $$GDP_t^R$$ is real GDP. For net exports, which is comprised of the contributions of imports/exports, we down-weight the lagged nominal share of GDP, $$\frac{N_{t-1}}{GDP_{t-1}^N}$$ , by the share of capital goods $$w_t$$ in quarter $$t$$ which is computed as

$$$$ w_t=\frac{CG_t}{CG_t+CS_t}, $$$$

where $$CG_t$$ is total imports/exports of capital goods and related parts, except automotive and $$CS_t$$ is total imports/exports of consumer goods, except food and automotive.

While this calculation methodology provides a rough picture of how AI investment may contribute to GDP growth, it is nevertheless limited. In the first place, there is no established consensus for how AI investment can be identified in aggregate statistics. Given the lack of a line item on the BEA NIPA tables, one must infer from broad investment categories, which may not adequately capture AI activities. For instance, the power facilities component proxies for data-driven grid and electrical spending, but it also includes investment in electrical infrastructure across non-AI manufacturing, construction, and utilities. Investment in other aggregated categories, such as communication equipment (not included in the current calculation), faces a similar tradeoff. While communication equipment captures relevant networking and communication infrastructure present in data centers, the category also includes spending on non-AI investments such as mobile phone and broadcast transmission equipment. Moreover, AI investment may have important downstream effects on GDP growth beyond the initial investment, e.g. by transforming productivity and labor composition (as discussed in the note), that cannot be captured through a simple investment accounting approach (for an alternative approach and further analysis, see Carpinelli et al., 2026).

Data File: Figure data (XLSX)