AI Productivity Arrives After the Inflation Bill

Executive summary

The conventional macroeconomic case for artificial intelligence is disinflationary. Better software should let firms produce more with the same labor, lower marginal costs and expand capacity. That case may ultimately be right. It is incomplete on timing.

The Federal Reserve's July meeting minutes, released August 19, describe an economy in which the AI buildout is already supporting investment and demand while its productivity benefits remain uncertain in timing and magnitude. The staff attributed part of higher core inflation to AI-related price pressure. Participants pointed to increases in the prices of chips, steel, smartphones, computer equipment, software and electricity, as well as strong wage demand for electricians, machinists and engineers. Several saw broader demand effects arriving before productivity-led supply relief.

Our contrarian thesis is that AI will be inflationary before it is disinflationary. The current phase requires physical capacity: servers, semiconductors, electrical equipment, grid connections, buildings, cooling, skilled trades and financing. That investment can raise measured output and future productive capacity, but it also competes for scarce inputs now. The productivity dividend arrives only after systems are deployed, workflows change and organizations convert technical capability into lower unit cost.

This does not make AI investment a mistake. It changes the operating hurdle. Leaders should evaluate AI programs as capital projects with an explicit construction period, utilization ramp and cost-of-capital exposure. Savings promised at maturity should not be used to ignore the inflation bill paid during deployment.

Market context

The July 28-29 Federal Open Market Committee meeting ended with the federal funds target range unchanged at 3.5% to 3.75%. Three members preferred a 25-basis-point increase. The minutes show that many participants thought tightening would likely be necessary if inflation failed to decline, and some questioned whether financial conditions were restrictive enough.

The disagreement is important because the investment boom and the inflation problem are connected. The Fed staff estimated that 12-month PCE inflation was 4.1% in May and core PCE inflation 3.4%. It estimated June readings of 3.7% and 3.3%, respectively. At the same time, real private domestic final purchases appeared to be growing faster than GDP, with consumer spending firming and AI investment supporting business demand.

The Bureau of Economic Analysis later estimated that real GDP grew 1.5% at an annual rate in the second quarter, while real final sales to private domestic purchasers increased 3.9%. Private investment in intellectual-property products increased 8.8% at an annual rate after 13.8% in the first quarter. Those data are broader than AI, but they confirm that software, research and related investment remain a material source of demand.

The inflation signal is not confined to a single monthly headline. The Bureau of Labor Statistics reported that July's Producer Price Index for final demand was flat because falling energy goods offset increases elsewhere. Final demand excluding food, energy and trade services rose 0.4% in July and 4.7% over 12 months. Final-demand construction prices rose 2.2% in the month, while commercial electric power also increased. The detailed release does not identify AI as the cause. It does show that the input environment facing a physical compute buildout is not broadly deflationary.

Exhibit 1: The buildout is running ahead of the measured productivity dividend

Indicator	Period	Reported change or level	What it says	Limitation
Real private domestic final purchases	Q2 2026	+3.9% annualized	Private demand was stronger than the 1.5% GDP headline	Includes all consumer spending and private fixed investment, not AI alone
Real intellectual-property investment	Q2 2026	+8.8% annualized	Software, R&D and related investment remained strong	Category is broader than AI and is an advance estimate
Nonfarm business labor productivity	Q2 2026	+1.4% annualized; +2.2% year over year	Productivity is positive but not yet an economy-wide step change	Aggregate data cannot isolate AI contribution
Nonfarm business unit labor costs	Q2 2026	+1.3% annualized; +1.4% year over year	Productivity did not eliminate labor-cost growth	Preliminary and subject to revision
PPI final demand less food, energy and trade	July 2026	+0.4% month over month; +4.7% year over year	Underlying producer-price pressure remained firm	Not an AI-specific price basket

Sources and methodology: BEA advance Q2 GDP release and intellectual-property investment series; BLS Productivity and Costs, August 6, 2026; BLS July PPI detailed report, August 13, 2026. Units are seasonally adjusted annual rates for quarterly growth, seasonally adjusted monthly change where stated, and unadjusted 12-month change where stated. No values are added or interpolated. The table compares timing signals; it does not estimate AI's causal contribution.

Findings

Finding 1

The first-order AI economy is a physical-capacity economy.

The product is software, but the constraint set is industrial. Training and serving models require chips, servers, networking, power conversion, cooling, buildings and interconnection. The buildout also consumes engineering, construction and financing capacity. When many firms pursue the same inputs at once, costs can rise before any user sees a cheaper workflow.

The Fed minutes make this sequence unusually explicit. Participants noted price increases in chips and steel used for data centers and in electricity, computer equipment and software. They also described strong demand for electricians, machinists and engineers in AI-related sectors. A few participants noted that more capital spending was being financed with borrowing from nonbank investors or regional banks. The buildout therefore reaches prices through goods, wages, utilities and the cost of capital.

Energy is the clearest measurable bottleneck. The Energy Information Administration says U.S. electricity demand grew about 1.7% annually from 2020 to 2025 after growing only 0.1% annually from 2005 to 2019. Its February 2026 outlook analysis forecast national load growth of 1.9% in 2026 and 2.5% in 2027, with data centers a principal driver. In a scenario where load growth in data-center regions ran 50% above baseline, modeled wholesale prices in New York and New England increased about $3 per megawatt-hour, or 5%, compared with the baseline.

That is a scenario, not an observed national surcharge. Its value is operational: the cost risk is local and nonlinear. A national average can look manageable while a specific grid, interconnection queue or labor market becomes the binding constraint.

Exhibit 2: Electricity demand has changed regime

Measure	Earlier period or baseline	Recent period or forecast	Unit
Average annual U.S. electricity-load growth	0.1% (2005-2019)	1.7% (2020-2025)	Average annual percent change
Forecast U.S. load growth	—	1.9% in 2026; 2.5% in 2027	Annual percent change
Forecast average load growth in ERCOT	—	10% (2025-2027)	Average annual percent change
Forecast average load growth in PJM	—	3% (2025-2027)	Average annual percent change
High-demand scenario price effect in New York/New England	Baseline	About +$3/MWh, or +5%	Average wholesale price difference

Source and methodology: U.S. Energy Information Administration, March 12, 2026, using the February 2026 Short-Term Energy Outlook and a high-demand sensitivity in which 2026-2027 load growth was 50% above baseline in regions with significant data-center development. Period: 2005-2027. Units are annual load growth and dollars per megawatt-hour. Limitations: ERCOT and PJM figures are regional forecasts; the price effect is a modeled scenario, assumes generating capacity fixed at the baseline, and is not a prediction for every customer bill.

Finding 2

Productivity is real, but deployment has a longer clock than procurement.

The optimistic view can point to improving aggregate productivity. BLS reported nonfarm business productivity growth of 1.4% at an annual rate in the second quarter and 2.2% over four quarters. Productivity growth since the fourth quarter of 2019 has averaged 2.1%, above the 1.5% pace of the previous business cycle.

Those are encouraging facts. They do not establish that AI has already generated savings large enough to offset the buildout. Aggregate productivity blends technology, capital deepening, business formation, labor reallocation and ordinary operating improvement. It also records successful output, not abandoned pilots, duplicate systems or implementation labor.

Procurement happens first. A company commits to models, cloud capacity, data engineering, security, integration and change management before it knows which tasks will improve. Benefits arrive later and unevenly. A model can reduce the time to produce a draft while increasing review, governance or exception-handling work. An internal tool can improve one team's throughput without reducing headcount or external spend. Revenue may grow before unit cost falls, which is economically valuable but not disinflationary.

The better test is realized unit economics. For each workflow, operators need a pre-deployment baseline, total loaded cost, adoption rate, output quality, exception cost and labor or revenue outcome. Time saved is an input to the calculation, not the final benefit.

Finding 3

AI can keep interest rates and hurdle rates higher even if its long-run supply effect is favorable.

The Fed's minutes describe a two-sided economy: strong AI-related investment supports output, while supply shocks and investment demand keep inflation risks tilted upward. Market rates had already moved higher in the intermeeting period. Treasury yields rose 25 to 30 basis points, markets priced a full 25-basis-point policy increase by September, and credit spreads for hyperscalers widened relative to investment-grade issuers.

These are market conditions at one point in time, not a durable forecast. The mechanism is durable. A capital boom can raise demand for funds, labor and equipment. If the central bank believes demand is outrunning near-term supply, financing stays tighter. That raises the discount rate applied to every AI project, including projects intended to reduce costs.

The distributional effect matters. The Fed staff said financing remained accommodative for larger companies but somewhat restrictive for small businesses and households. Credit was tight for new card applicants, mortgage activity remained depressed, and small-business credit remained constrained. Large firms can fund infrastructure and benefit from asset-price gains while smaller firms pay higher borrowing and input costs before they can access the same productivity tools.

This creates a strategic trap. A company may accelerate AI spending because competitors are investing and because the technology could be transformative. At the same time, the boom raises its cloud prices, talent costs and hurdle rate. The right response is not to stop. It is to stage commitments so evidence, not narrative momentum, unlocks the next tranche.

Implications for operators

Treat AI programs as investments with a construction period. Build a cash-flow model that separates upfront integration, recurring inference, data, governance, security and change costs from later savings or revenue. State the utilization level required for payback.

Index major contracts to the real bottleneck. Cloud, colocation, electricity and model contracts can shift input risk back to the buyer. Procurement teams should model price, capacity and minimum-spend scenarios rather than negotiate only the headline unit rate.

Measure avoided cost, not theoretical time saved. If employees complete work faster but headcount, contractor spend, cycle time and error costs do not change, the productivity gain has not reached the income statement. That may be acceptable during learning, but it should be labeled correctly.

Protect smaller suppliers and partners from the financing gap. Faster payment, milestone funding and shared tools can prevent a buyer's AI program from imposing working-capital stress on the firms doing integration, data preparation or field installation.

Use stage gates tied to adoption and quality. Release additional capital when a workflow reaches defined utilization, accuracy, exception and contribution thresholds. The relevant portfolio choice is not "AI or no AI." It is which deployment earns scarce capital first.

Risks & open questions

The thesis would weaken if productivity growth accelerates broadly while unit labor costs, commercial power prices and AI-related equipment prices decelerate. It would also weaken if new generation, transmission and semiconductor supply arrives faster than demand, or if model efficiency sharply reduces compute required per useful task.

Attribution remains the largest evidence problem. None of the official macro series isolates AI's causal effect. Intellectual-property investment includes non-AI software and research. Electricity demand includes industrial uses. Producer prices reflect tariffs, energy shocks and other forces. The Fed minutes record informed judgments, not a structural model with a published AI coefficient.

There is also a compositional possibility. AI may be inflationary in regions and inputs tied to the buildout while disinflationary in services where adoption is fast. The aggregate outcome will depend on the size and timing of each channel.

Finally, a failed investment boom could reverse the result abruptly. The Fed minutes noted elevated valuations, increased borrowing and the risk that disappointment could trigger asset repricing and tighter financial conditions. That would reduce demand, but through capital losses and retrenchment rather than a healthy productivity dividend.

Appendix / methodology notes

This report reviewed primary materials available through August 19, 2026: the July FOMC minutes; BEA's advance second-quarter GDP release and intellectual-property investment series; BLS second-quarter Productivity and Costs and July PPI releases; and EIA's 2026 electricity-demand analyses.

Reported values retain source definitions. Quarterly BEA and BLS growth rates are seasonally adjusted annual rates unless noted. Twelve-month PPI changes are unadjusted. EIA regional electricity figures are forecasts, and the $3/MWh result is a sensitivity scenario rather than an observed price increase.

No attempt is made to estimate AI's share of inflation or productivity because the required causal data are not public. A stronger future test would combine, by region and quarter, data-center construction spending, interconnection requests, commercial electricity prices, semiconductor and electrical-equipment prices, AI-related occupational wages, borrowing spreads, compute utilization and realized workflow savings. The core falsification chart would compare the cumulative buildout-cost index with cumulative verified unit-cost savings for a consistent set of adopters.