The Bureau of Economic Analysis (BEA) has now published its third estimate of second-quarter 2026 GDP, and it tells a different story from the one markets read in July. The advance estimate showed real GDP growing at a 1.5 percent annual rate. After BEA's annual update, the third estimate puts the quarter at 2.2 percent, according to BEA's current release page. EY's macroeconomics team, summarizing the revisions, attributes much of the upgrade to stronger business investment led by AI-related spending.

A 0.7-point swing for a quarter that has already ended is a reminder that investment in AI hardware, software and data centers is measured with real uncertainty. It also raises a question that headline GDP does not answer: what does a historic capital-spending cycle show up as in the official statistics, and what does it cost the grid and ratepayers to power it? This piece follows that chain through federal data on investment, adoption, productivity and electricity.

What the investment data actually show

The most careful public attempt to isolate AI in the national accounts comes from the St. Louis Fed. Economist Hannah Rubinton and research associate Bontu Ankit Patro, in a January 12, 2026 post, Tracking AI's Contribution to GDP Growth, combine three BEA series (information processing equipment, software, and R&D) with Census Bureau data on data-center construction. Their results are striking. Together, those four categories contributed 1.3 percentage points to real GDP growth in the first quarter of 2025 and 1.16 points in the second quarter. That was about 30 percent of second-quarter growth. The combined contribution then dropped to 0.48 points in the third quarter.

Over the first nine months of 2025, the authors calculate, these AI-related categories added 0.97 percentage points to growth. That compares with 0.81 points during 2000, the peak of the dot-com boom. They made up 39 percent of total growth, against 28 percent in 2000. Software drove much of the difference, contributing 0.35 points versus 0.11 in 2000. The authors also flag a detail that is easy to miss in coverage of the boom. Investment levels stayed high, but the contribution to growth faded because the pace of increase slowed. In their words, "Contributions to GDP growth depend on growth, not high levels alone."

The dollar figures from BEA, as republished by FRED, show the level story. Private fixed investment in information processing equipment and software ran at a seasonally adjusted annual rate of about $1.227 trillion in the fourth quarter of 2024. By the fourth quarter of 2025 it was about $1.466 trillion, a nominal increase of roughly 19.5 percent in a year. The advance estimate for the second quarter of 2026 showed the same pattern: investment growth was led by industrial, transportation and information processing equipment, plus software and R&D. EY's summary of the third estimate puts equipment spending up 13.4 percent and intellectual property products up 9.2 percent. Information processing equipment rose a more modest 7.6 percent in that reading, and structures were nearly flat.

Three reasons the headline number can mislead

First, imports. A large share of AI hardware is built abroad, so a data-center server shows up twice in the national accounts. It adds to investment when a firm buys it and subtracts through imports. BEA's own release notes that a rise in imports produces a negative contribution to GDP. In the third estimate, EY reports, net trade subtracted 1.1 percentage points from second-quarter growth as imports surged on demand for technology equipment. For a country buying much of its AI equipment from overseas, gross investment overstates the boost to domestic production.

Second, the early estimates of the software and R&D components are partly modeled. BEA's advance release says the second-quarter increase in intellectual property products was based primarily on a judgmental trend and Bureau of Labor Statistics employment data. That does not mean the software boom is imaginary. It does mean that some of what looks like an observed surge in AI software spending is an estimate that later source data will revise, which is one reason GDP numbers keep moving after release.

Third, timing. National accounts record equipment as investment when it is acquired, not when it begins producing anything. A cluster of accelerators that sits waiting for a grid connection still counts as investment, but it adds no output until power arrives. That gap between spending and productive use is the core of the productivity question below.

Adoption is broad in headlines, narrow in the data

If AI spending is large, adoption should be showing up in firms. The Census Bureau's Business Trends and Outlook Survey (BTOS), which is nationally representative and biweekly, gives the clearest answer. Census statisticians Adam Grundy, Cory Breaux and Dhan Khatiwoda analyzed data from December 14, 2025 to May 3, 2026. As of May 3, 19.8 percent of firms reported using AI. Rates were 37 percent for firms with 250 or more employees and 32 percent for firms with 100 to 249 employees. The Information sector stood at 39.7 percent and Finance and Insurance at 33.9 percent. Use rose among firms with at least 20 employees but did not change significantly among smaller firms.

One caution for anyone reading trend charts: Census changed the survey wording last November. The question moved from AI use in producing goods or services to AI use "in any business function." Part of any jump across that break reflects the definition, not new adopters. Census's spring 2026 AI supplement adds texture. It put adoption at roughly 18 percent in April and expected about 22 percent by fall. Most firms use AI in only one to three business functions, and two-thirds use it solely to augment tasks rather than replace them.

The mismatch matters. Spending is concentrated among a small group of very large buyers, while most American businesses are still at an early stage of use. Productivity gains come from using the technology, not from buying the chips, so the GDP boost and any productivity payoff are on different clocks.

The productivity evidence is real but thin

Productivity growth is the number that would validate the spending, and the official data are encouraging but ambiguous. The BLS reports that nonfarm business labor productivity rose 2.2 percent from the second quarter of 2025 to the second quarter of 2026, with output up 2.5 percent and hours worked up only 0.2 percent. On a quarterly annualized basis, productivity increased 1.4 percent. BLS also reported the labor share of income at 52.8 percent, the lowest on record. Those figures predate BEA's annual revision to GDP, and BLS is due to publish third-quarter productivity on November 5, so they may move.

Federal Reserve regional research suggests why caution is warranted. A working paper from the Federal Reserve Banks of Atlanta and Richmond and Duke University's CFO Survey, by Ty McClure, Brent Meyer, Michael Sparks, Daniel Weitz, Zachary Edwards, Sonya Ravindranath Waddell and John R. Graham, draws on responses from nearly 750 executives. It finds that more than half of firms have invested in AI, with the largest expected gains in high-skill services and finance. The authors document a productivity paradox in which "perceived productivity gains are larger than measured productivity gains." They attribute this to a delay before revenue gains materialize. The same research finds little evidence of near-term aggregate job losses, although large firms expect reductions of about 0.8 percent in 2026. This fits a pattern economists such as Erik Brynjolfsson, Daniel Rock and Chad Syverson described as the productivity J-curve, in which intangible investments depress measured productivity before they lift it. The St. Louis Fed post cites that framework as well.

The electricity bottleneck

The physical constraint on all this is power. The Department of Energy's Lawrence Berkeley National Laboratory estimated in its 2024 data center report that data centers used about 4.4 percent of U.S. electricity in 2023, or 176 terawatt-hours, up from 58 terawatt-hours in 2014. The same report projects 325 to 580 terawatt-hours by 2028, or 6.7 to 12 percent of national consumption. The range is wide because nobody knows how many announced projects will actually be built.

Regulators are now deciding how those loads connect. In October 2025, the Department of Energy directed the Federal Energy Regulatory Commission (FERC) to consider a rule on interconnecting large loads, with a 20-megawatt threshold. In the advance notice, large loads would bear 100 percent of network upgrade costs assigned through interconnection studies. FERC did not issue a single national rule. On June 18, 2026, it issued show-cause orders to six regional transmission organizations (PJM, MISO, SPP, CAISO, ISO-NE and NYISO), giving each 60 days to justify its current tariff or file changes. FERC left open the option of more direct action if the responses fall short. That is a tailored, region-by-region approach, which will likely produce different interconnection timelines in different parts of the country.

Where GDP and the utility bill meet

Transmission upgrades are only part of the cost. Generation, substations and distribution run through state rate cases, and those filings are climbing. The nonprofit PowerLines reports that utilities requested about $31 billion in rate increases in 2025, more than double the $15 billion requested in 2024. Its newest report, published October 8, 2026, counts $23 billion in requests across 28 states so far in 2026. That includes $4.5 billion in the third quarter, the highest third-quarter total in a decade, and investor-owned utilities have about $1.4 trillion of capital investment planned through 2030. EIA retail sales data show the average residential price at 18.31 cents per kilowatt-hour in July 2026, up 4.9 percent from a year earlier while sales rose just 1.0 percent. Price is rising faster than volume.

It would be a mistake to blame all of that on AI. PowerLines found that in the first quarter of 2026, utilities in the West requested the most ($4.4 billion, versus $2.7 billion in the South). It attributes much of the Western total to wildfire costs and risk rather than data centers. Regional differences are the point. Virginia, Texas and the PJM states face data-center-driven cost allocation fights, while California's rate pressure is mostly about fire.

This is my interpretation, not a BEA finding, but the accounting here is awkward in a way that rarely gets discussed. When a utility builds a plant or line to serve new load, that spending counts as investment and adds to GDP. When households pay higher bills to fund it, nominal consumer spending rises, but real consumption of electricity services barely moves if usage is flat. A rate-financed buildout can therefore lift measured GDP while squeezing households' purchasing power. Nonresidential structures were not a strong contributor in the second quarter, with BEA citing decreases in that category and EY showing near-zero growth. That suggests the physical infrastructure behind the AI boom has not yet shown up in the national accounts to the same degree as the equipment sitting inside it.

What to watch next

Several data releases in the coming weeks will test this story. BEA's advance estimate of third-quarter GDP is scheduled for October 29, and the details on information processing equipment, software and structures will show whether the spending cycle is still accelerating or has plateaued as it did in the St. Louis Fed's 2025 numbers. BLS third-quarter productivity on November 5 will show whether the 2.2 percent year-over-year gain holds after BEA's revisions. The FERC show-cause responses and the next rate-case filings will indicate whether the cost of connecting large loads lands mostly on developers or on households. Together they will give a clearer answer than any single number on whether AI spending is turning into lasting output or just into higher bills.