Published July 17, 2026 by Paul E. Soto, Mason Thieu, and Jeffrey S. Allen of the Board of Governors of the Federal Reserve System, with an accompanying data file compiling the series discussed.
Purpose and framework
The note assembles "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," organized into three categories that "map closely to the sequence often associated with general-purpose technologies": capabilities and costs, then firm investment and adoption, then productivity and labor.
Its stated diagnostic purpose is to distinguish two possible states: "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." It also frames a gap it expects the indicators to speak to — that 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."
Its summary of the position 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."
Capabilities and costs
The note treats benchmark saturation as a measurement problem: early standardized-test benchmarks "were quickly saturated, with subsequent expert level reasoning benchmarks (e.g. GPQA and ARC-AGI) following a similar pattern," and strong benchmark performance "does not necessarily translate into the ability to complete actual on-the-job tasks."
It adopts METR's agentic task-completion horizon as its capability indicator, noting the focus on software engineering and machine learning "as these are the domains frontier labs are automating first, where impacts should appear earliest." As of mid-2026 the horizon "has been doubling roughly every several months," at which 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."
The distinction the note draws next is the analytically load-bearing one: task-completion horizon measures "technical feasibility, but not cost-effective deployment." Deployment turns on two costs. The 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" — is "not directly observable and difficult to quantify." The marginal cost of inference is observable, tracked through GPU cost per unit of bandwidth and compute (Epoch AI) and through Korean semiconductor export prices, used as a memory-cost proxy because "Samsung and SK Hynix account for roughly 70 percent of global DRAM supply and are the dominant share for high-bandwidth memory production."
The note declines to publish a token-price series, explaining why: 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," and posted rates "might not reflect enterprise contracts." It flags the margin as worth watching "as the binding constraint on adoption shifts from building capacity to paying for this intermediate input."
Investment and adoption
Each investment proxy is presented with its attribution problem. Hyperscaler capital expenditure "has arguably become one of the main demand-side proxies for AI infrastructure," but "the level is hard to interpret because it includes non-AI capex," and increasing reliance on leased capacity means headline figures "could progressively understate total AI-related infrastructure investment." Census data-center construction is "a leading indicator of anticipated computing demand because physical structures are built before they are filled with computing equipment," limited because "physical structures are not the dominant source of value in data centers." BEA equipment investment fills that gap but "include[s] substantial non-AI activity as well, such as normal office computers"; extrapolating a pre-2023 baseline to isolate an AI component "may become less reliable over time as other trends contaminate the counterfactual."
The note reads divergences between the series as diagnostic: "with capex rising while construction starts plateauing" would be "an early sign of supply constraints or a shift towards software and operating expenditures as the buildout matures," while meaningful deceleration "could signal infrastructure demand having been met or a downward revision in expected return on investment."
Productivity and labor
The central finding is the gap between micro and macro evidence. "Micro-level experiments consistently find productivity gains from AI tools. The broader question is whether these gains translate into aggregate productivity growth." Tracking sectoral productivity by AI exposure — high (Information, Finance, Professional and Business Services), medium (Wholesale Trade, Manufacturing, Education and Health, Retail), low (Utilities, Construction, Leisure and Hospitality, Transportation) — the note reports that "productivity trends across all three levels have been relatively consistent over time, suggestive of micro-level productivity gains not adding up in aggregate."
Four explanations are offered rather than one. Micro experiments "typically measure task-level productivity, like the speed a programmer generates code, rather than job-level or firm-level output," and "a 10% improvement on a task does not necessarily lead to proportional gains for a firm if adjustment costs or other bottlenecks lie elsewhere." Gains found in information-sector roles may not generalize to sectors "where work looks very different." Historically, "measured productivity gains from GPTs lag investment by years." And measurement itself is hard: gains "may be misattributed between capital deepening and total factor productivity," intangibles are incompletely captured, and in services "where output is often inferred from revenue, price declines can distort measured output."
On labor, the note reports that "aggregate unemployment has remained moderate by historical standards, but this headline number masks compositional shifts that might be related to AI." Youth unemployment is singled out for a mechanism beyond displacement: if AI substitutes for entry-level tasks, "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." Labor-force participation is tracked separately "to capture whether individuals are withdrawing from the labor market entirely rather than moving between jobs."
Citing Crane and Soto (2026), the note locates programming-intensive occupations in professional and technical services (NAICS 54) rather than the information sector alone, and reports early evidence "that AI is impacting younger workers via slower hiring rather than outright layoffs." It also cites the Anthropic Economic Index as mapping actual usage to tasks concentrated in "software development, technical writing, and analytical work."
Conclusion
"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."
The note names three signals that would indicate a shift: "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."
Relationships
- supports: AI and Productivity — supplies official-statistics evidence that task-level gains have not aggregated
- related: AI Labor Disruption — finds effects concentrated in slower hiring for younger workers rather than layoffs
- related: AI Bubble vs. Buildout — Synthesis — proposes the indicators by which the two would be distinguished
- related: METR, AI Data Centers, Financial Services — AI Deployment, Enterprise AI Deployment Gap