On this Labor Day, we look at an important issue: In recent years, women have been faring better than men in the job market. The following figure shows that, for workers 20 years and older, men still hold more jobs than women do, but the gap has been closing. For example, as measured by the household survey conducted by the Bureau of Labor Statistics (BLS), between January 2022 and August 2026, there was a net increase of 5,829,000 jobs in the United States. More than two thirds of those jobs were filled by women.
In recent months, the business press has begun to focus on this issue. Here are some recent headlines: “A Changing Job Market Leans Against Men,” “In This Job Market, Women Have the Upper Hand,” and “Young Men Are Abandoning the Workforce.” In Macroeconomics, Chapter 9 (Economics, Chapter 19), we discuss the employment-population ratio, which measures the fraction of the working-age population of a particular segment of the population that is employed. The following figure shows that the employment-population ratio for prime-age men—those aged 25 to 54—has been slowly trending downward for decades (the blue line), while that ratio has generally been increasing for women (the orange line).
In March 1953, the employment-population ratio for prime-age males reached a peak of 96.0 percent. In August 2026, the ratio was 85.8 percent. If prime-age males were working in 2026 at the rate that they did in 1953, 10 million more men would be working today than actually are.
The following figure makes clearer the differing trends in men and women’s employment-population ratio in recent years. In this figure, the values for both ratios are set equal to 100 in January 2000. Since that time the employment-population ratio for prime-age women (the orange line) has increased by 1.1 percent, while the ratio for men (the blue line) has declined by 4.1 percent.
Why do a smaller fraction of prime-age men have jobs today than in the past? A large number of explanations have been offered, both in the business media and by academic economists. One key factor, as shown in the following figure, is that women (the orange line) are now more likely to earn a college degree than are men (the blue line).
The fraction of jobs requiring a four-year degree has been increasing over time, a trend that the BLS projects will continue. As the following figure shows, men with a bachelor’s degree or more have a higher employment-population ratio than do men with only a high school degree. (Note that the data in this figure are for all men 25 years and older, not just for prime-age men. The average age of men has been rising, which lowers the employment-population ratio as an increasing fraction of men become of retirement age. These data are not available on a seasonally-adjusted basis, which accounts for the choppiness in the figure.) As men have fallen behind in earning college degrees, more men have found themselves unqualified to be hired in some jobs.
An article in the Wall Street Journal used BLS data to divide jobs primarily held by women and those primarily held by men. As the following figure from the article shows, jobs help primarily by women have been increasing faster than those held by men.
Image generated by ChatGPT
As we noted in a blog post earlier this year, health care jobs have come to dominate U.S. employment growth. The following figure shows monthly changes in health care and social assistance jobs (the blue bars) and monthly changes in total employment (the red bars) for each month since January 2025. During this time period, net employment in health care and social assitance increased by 1,027,300 jobs. All other job categories experienced a decrease of 268,300 jobs. Women account for 77.9 percent of health care and social assistance workers. In other words, the number of jobs in industries dominated by men have been declining.
Image generated by ChatGPT
If you look again at the graph showing changes in the employment-population ratio for prime-age men (the second graph in this blog post), you’ll notice that there seems to be a ratchet effect in the data: The employment-population ratio declines during each recession (shown by the gray bars in the figure) and then struggles to return to its pre-recession level. It’s unsurprising that the male employment-population falls sharply during recessions, because, as we discuss in Macroeconomics, Chapter 13 (Economics, Chapter 23) spending on residential construction and consumer durables, such as automobiles and appliances, falls sharply during a recession.In 2025, men were 86.8 percent of workers in construction and 77.9 percent of workers in manufacturing. (In fact, as we note in that chapter, the late Edward Leamer of the University of California, Los Angeles, went so far as to argue that “housing is the business cycle.”)
Image generated by ChatGPT
Just before the Great Recession and Global Financial Crisis of 2007–2009, the prime-age male employment-population ratio was 88.0 percent, a level it hasn’t attained since. (In a recent blog post, we discuss the role the bankruptcy of the Lehman Brothers investment bank played in the financial crisis.) The prolonged unemployment experienced by some male workers in construction and manufacturing may have led to their skills deteriorating, making it more difficult for them to find employment during the following economic recovery. Some of these workers may have dropped out of the labor force resulting in a decline in the employment-population ratio.
One explanation for the declining employment-population ratio for prime-age males that has received significant attention in the media is the increased appeal of video games. Or, as the headline of an article in the New York Times put it: “Why Some Men Don’t Work: Video Games Have Gotten Really Good.” The U.S. Census Bureau annually conducts the American Time Use Survey, which is published by the BLS. The following figure shows that young adult men have increased the time they spend playing games. In 2003, men aged 21 to 30 spent an average of 2.23 hours per week. In 2025, they spent an average of 7.75 hours per week, down from a peak of 8.56 hours per week in 2022.
Mark Aguiar, of Princeton University, and colleagues argue that the increase in time young men devote to playing video games and engaging in other “recreational computer activities” has significantly reduced the amount of hours that some young men work. There has, however, been an academic debate over this contention. First, it’s unclear which way the causality runs: Do young men work less because they find playing video games particularly attractive or has the ability of young men to find jobs declined, so they spend time playing video games that they would rather spend working? Second, older prime-age males, who have not increased their time playing video games by as much, have also experienced a falling employment-population ratio.
There have been a number of other changes in labor markets and in American society that may have contributed to the decline in employment of prime-age males. ChatGPT offers the following summary of the various factors:
“I would rank the explanations this way:
Most important: the disappearance of stable, comparatively well-paid routine and manual jobs available to men without college degrees, together with slow occupational and geographic adjustment.
Closely related: educational and skills differences, the concentration of new employment in female-heavy service sectors, and the difficulty men face moving into those jobs.
Important amplifiers: chronic health problems, mental illness, pain, opioids and other substance abuse, and the long-term effects of recessions and prolonged joblessness.
Important for particular groups: criminal records, incarceration, geographic isolation, and weak local labor markets.
Reinforcing social mechanisms: delayed marriage and parenthood, living with relatives, weaker social expectations concerning steady work, and reduced connection to employers and communities.
Real but often overstated: disability benefits, other public assistance, and video games.
The central academic message is therefore different from the most sensational press version. It is not principally that millions of otherwise successful men suddenly preferred video games or welfare to jobs. The decline began with a weakening of the kinds of labor-market opportunities historically available to noncollege men. Health, addiction, criminal records, family change, geographic immobility, and more attractive leisure then made the resulting withdrawal from employment more persistent.”
A recent article in the Wall Street Journal discussed the surprising fact that some colleges are sending letters of acceptance to students who haven’t actually applied for admission:
“Hundreds of colleges are sending students letters of admission—without even requiring an application. … Known as ‘direct admissions,’ this expedited process is free and omits required essays, questions about extracurriculars and mandated standardized tests.”
The following figure from the article shows the increase in the number of colleges among the 1,100 colleges that accept the Common Application (or Common App) that use direct admissions.
The rise in the use of direct admissions reflects a decline in students’ demand for admission to these schools. Part of the reason for this decline in demand is the falling number of people in the United States who are in the prime college attending ages of 18 to 24. The following figure shows projections from the Census Bureau of the number of U.S. residents in this age group from the present to the year 2100. The numbers on the vertical axis are thousands of persons. From 2022 to 2026, the number of people in this age group declined by about 1 million. The number is projected to have declined by another 2 million in 2040.
Another factor that may be affecting the demand for college admissions is stagnation in the college wage premium, which is the amount by which wages earned by college graduates exceed wages earned by high school graduates. The following figure from a publication of the Federal Reserve Bank of Minneapolis shows values for the college wage premium from 1961 to 2023. The figure uses data from a working paper by economists at the Federal Reserve Bank of San Francisco that adjusts the college wage premium to take into account several factors, including differences in the ages of high school and college graduates.
The college wage premium has fluctuated, but from 1980 to 2000 it was generally increasing. Since 2000, however, the premium has stagnated. Several explanations have been offered for this stagnation. Lisa Camner McKay of the Minneapolis Fed notes that the relative supply of workers with college degrees has been increasing: “In 2000, workers with a bachelor’s degree or higher were 31 percent of the civilian labor force. In January 2025, they were 45 percent.”
The labor market demand for college graduates may also have declined relative to the demand for high school graduates. The following figure, based on data in the working paper from the San Francisco Fed referred to earlier, shows the ratio of public job postings that require applicants to have a college degree relative to job posting that don’t require a college degree. The ratio has steadily declined since 2010.
We’ve identified two factors that may account for a decline in the demand for a college degree that’s led some colleges to rely on direct admissions to recruit students. Media stories have also noted that some smaller colleges have been forced to close in recent years as they were unable to recruit enough students to cover their costs. These closings have reduced the supply of college degrees. However, only about 46 traditional nonprofit private colleges closed between 2023 and 2025. While these closures have been a hardship for the students, faculty, and administrators involved, they have been a very small fraction of the more than 3,000 public and private colleges in the United States. But some observers have forecast that closures of small private colleges may sharply increase in the coming years. For example, an article in the Wall Street Journal cited a study by Huron Consulting that found that 442 of the 1,700 private nonprofit colleges have experienced shrinking enrollments and are at risk of closing at some point in the next 10 years.
How might declines in the demand for and supply of college degrees affect the tuition that students will pay in the future? First, it’s worth noting that, corrected for the effects of inflation, college tuition has not increased significantly in recent years. The following figure, using data from the College Board, shows that, when measured in 2025 dollars, college tuition at public and private colleges has been roughly flat over the past 10 years, particularly if we look at net tuition charged, which subtract grants the colleges have awarded to students from the colleges’ published tuition amounts.
We can use the model of demand and supply to analyze how tuition might change in the future. In Microeconomics, Chapter 3, Section 3, we show that whether the price in a market rises over time depends on the direction in which demand and supply curves shift and on the relative magnitudes of the shifts. In this case, our discussion indicates that both the demand for college degrees and the supply of college degrees are likely to continue shifting to the left. Whether tuition rises or falls depends on the magnitude of the shifts. If the shift in demand is greater than the shift in supply, tuition will fall. If the shift in supply is greater than the shift in demand, tuition will rise. The following figure illustrates the situation in which the demand for college degrees shifts by more than the supply of college degrees, causing tuition to fall.
Image created by ChatGPT, as are the other images in this post.
Technological breakthroughs are often embodied in machinery, equipment, or, as in the case of artificial intelligence (AI), software. For example, the Industrial Revolution of the late 1700s, which involved mass production of low-priced cotton textiles, was made possible by the development in England of large cotton spinning machines that were often powered by steam engines.
Other technological breakthroughs are disembodied because they involve reorganizing production rather than employing new machinery or software. For example, the just-in-time (JIT) system that Toyota developed beginning in the 1940s, lowered the costs of assembling automobiles by, among other things, having parts arrive at a factory just when needed.
Across U.S. history, there have been a number of technological revolutions that required massive capital investments. For instance, the transportation revolution of the early 1800s that significantly reduced the costs of shipping freight, relied on substantial investments in canals. Later in the 1800s, transportation costs were reduced further as firms invested in the largest network of railroads in the world. Electrification of businesses and private homes in the late 1800s and early 1900s required the building of power plants and transmission lines. In the late 1990s, many firms competed to build networks of fiber-optic cable to carry the rapid increase in digital traffic resulting from the development of the internet.
Each of these technological revolutions succeeded in providing consumers with new or lower-cost goods or services. But building the physical capital that embodied these technologies could result in substantial losses or bankruptcy. For example, in the 1830s, several states, including Pennsylvania, Illinois, and Michigan, defaulted on the bonds they had issued to finance canal instruction.
Similarly, in the 1890s, some of the largest private railroad companies in the United States, including the Union Pacific, the Northern Pacific, and the Atchison, Topeka & Santa Fe, went bankrupt. In the early 2000, investors realized that internet commerce wouldn’t expand as rapidly as they had expected. As a result, several prominent internet firms, such as Pets.com and eToys.com, saw their stock prices collapse in what was called the dot com crash, and eventually declared bankruptcy. Several firms that had invested in building large fiber-optic networks, including Worldcom and Global Crossing, also failed.
Why did the companies making large capital investments in new technologies sometimes fail even though the technologies themselves proved to be revolutionary? The answer in part is that forecasting accurately the demand for a new product can be difficult. For example, ultimately the level of internet traffic was more than great enough to justify the investment in fiber-optic networks, but the increase came too late for some companies to avoid bankruptcy. Competition among firms to take advantage of a new technology can lead to overbuilding, as when railroad companies built more miles of track than were needed to meet the demand to transport freight and passengers.
A number of economists and investment analysts have noted that the very heavy investment in data centers used to power AI models resembles the large capital investments firms made during earlier technological revolutions. As the following figure shows, beginning in early 2023, shortly after the public release of ChatGPT, the first widely used AI chatbot, investment in information processing equipment and software as a percentage of all nonresidential fixed investment has increased dramatically.
Do the firms building data centers run the risk of overestimating demand for AI and overbuilding data centers? More generally, how does the build-out of AI data centers compare with the earlier build-outs of railroads and fiber-optic networks? Who better to answer this question than ChatGPT!
The following (very long!) answer was ChatGPT’s response to the following prompt: “In what ways is the current build-out of AI data centers similar to the build-out of the railroad network in the United States in the late nineteenth century and the build-out of the fiber-optic network in the United States in the 1990s and in what ways is it different?” (A quick summary is given in the brief “Bottom Line” section at the very end.)
AI Data Centers, Railroads, and Fiber-Optic Networks
The best way to frame the comparison is this:
The railroad analogy is most useful for understanding the physical scale, bottlenecks, public subsidies, and geographic consequences of the AI buildout. The fiber-optic analogy is most useful for understanding its financial risks, uncertain demand, rapid technological improvement, and potential for overcapacity.
AI data centers nevertheless differ importantly from both. They are not simply links in a transportation or communications network. They are more like factories that produce computing services, and the extremely expensive machinery inside them becomes obsolete much faster than either railroad rights-of-way or buried fiber.
The scale is already extraordinary. The International Energy Agency estimates that the largest technology companies spent more than $400 billion on capital investment in 2025 and expects that figure to rise about 75 percent in 2026. U.S. data centers consumed approximately 176 terawatt-hours of electricity in 2023—4.4 percent of U.S. electricity use—and Berkeley Lab projects consumption of 325–580 terawatt-hours by 2028.
Ways the three buildouts are similar
1. All three are investments in enabling infrastructure
Railroads did much more than generate revenue from tickets and freight charges. They brought previously isolated regions into a national market, encouraged settlement, changed the location of factories and cities, and made complementary investments in farms, mines, warehouses, and factories more valuable. The United States added roughly 170,000 miles of railroad track between 1871 and 1900.
Fiber-optic builders similarly expected abundant bandwidth to generate applications that could not yet be foreseen. More bandwidth would encourage new internet services, which would increase bandwidth demand, which would justify still more fiber.
The AI argument is nearly identical: companies are building computing capacity before all its profitable applications have been discovered. The value of the data centers will depend partly on complementary developments—better models, proprietary data, new software, organizational changes, and the willingness of companies and consumers to pay for AI services.
In all three cases, therefore, infrastructure and applications are jointly determined. Waiting until demand is plainly visible can mean waiting too long, because the infrastructure itself helps create the demand.
2. High fixed costs encourage companies to build ahead of demand
A railroad requires land, grading, bridges, tunnels, stations, track, and rolling stock before it carries its first paying customer. A fiber network requires rights-of-way, trenches, cable, switching equipment, and connections before it can sell substantial bandwidth.
AI data centers likewise require sites, grid connections, substations, transformers, cooling systems, generators, networking equipment, and large orders for advanced chips. Many of those expenditures are sunk: once a company has built a substation or a specialized data-center shell in a particular location, much of the investment cannot readily be recovered.
Companies also face a strategic incentive to build early. An individual hyperscaler may reasonably conclude that the loss from having insufficient computing capacity—lost customers, inferior models, or technological dependence on a competitor—would be greater than the loss from temporarily having too much. But when every major company reasons this way, the industry can collectively build more capacity than it needs. The IEA reports increasingly severe bottlenecks involving electricity, grid connections, chips, memory, transformers, and capital, even while acknowledging that many announced projects will never be completed.
3. Each boom rests on highly uncertain projections of future demand
Railroad promoters projected the future settlement, agricultural production, and mineral output of regions that were sometimes scarcely populated. Eventually, railroad projects began to outpace the demand for their capacity, reducing returns and contributing to widespread defaults.
Fiber builders extrapolated the extraordinary early growth of internet traffic. They assumed that new applications and fiber connections to homes and businesses would arrive rapidly enough to absorb the new long-distance capacity. Those assumptions proved too optimistic, at least in the short run.
AI companies are making similarly difficult forecasts. No one knows with much confidence:
how much computing future models will require;
how quickly businesses will reorganize around AI;
how much customers will pay for AI services;
whether most AI revenue will accrue to model developers, cloud companies, application providers, or users; or
how much improved chips and algorithms will reduce the computing required for a given task.
The uncertainty does not mean that the investment is irrational. It means that large forecast errors—in both directions—are likely.
4. Rapid technological progress can itself produce overcapacity
This was especially important during the fiber boom. In 1996, one fiber strand could carry approximately 2.5 gigabits per second. By 2000, improved multiplexing equipment could raise the capacity of the same strand to approximately 100 gigabits per second. Thus, companies were not merely laying more fiber; technology was multiplying the capacity of fiber already in the ground. Communications-equipment investment rose from about $62 billion at an annual rate in early 1996 to more than $135 billion in late 2000, before falling sharply.
AI faces an analogous risk. The cost of running a model with roughly GPT-3.5-level performance fell more than 280-fold between November 2022 and October 2024, according to Stanford’s AI Index. Hardware costs were declining by about 30 percent annually and energy efficiency was improving by approximately 40 percent annually.
Consequently, today’s estimates of the number of data centers needed in 2030 may prove too high if chips, models, and software become much more efficient. That is the closest parallel to the fiber glut.
But there is an important countervailing effect: cheaper computation can generate much more use. The IEA estimates that energy consumption per basic AI task has recently fallen by at least an order of magnitude annually, yet more demanding reasoning, video, and agentic applications are expanding so quickly that total data-center electricity consumption is still projected to roughly double between 2025 and 2030. Efficiency may therefore lower the cost of AI without lowering total demand for computing or electricity.
5. A private investment failure can leave a valuable social legacy
Many railroad investors lost money, and many railroad companies entered receivership, even though the resulting network transformed the American economy. The Panic of 1873 began in substantial part because railroad construction had outrun demand and heavily indebted companies could not refinance, but the tracks themselves did not cease to be useful.
The fiber experience was similar. Long-distance fiber companies failed, telecommunications shares collapsed, and equipment investment fell, but the underlying fiber remained available to carry later increases in internet traffic. The glut helped drive down the price of communications capacity and made subsequent internet businesses less expensive to operate. The Richmond Fed concluded that excessive investment and rational responses to genuinely revolutionary—but highly uncertain—technology were not mutually exclusive explanations.
AI could follow the same pattern. A future decline in data-center rents or computing prices might be bad for data-center owners and chip suppliers but good for application developers and users. AI could be economically transformative while a substantial fraction of current AI investment earns a disappointing return.
Important differences
1. A data center is not a network in quite the same sense
A railroad line moves something between two locations. A fiber-optic cable transmits information between locations. Their usefulness depends strongly on connections to the rest of the network.
A data center primarily produces computation. It resembles an electricity-generating plant, steel mill, or semiconductor fabrication plant more than a railroad line. It needs fiber connections to communicate with users and other data centers, but its principal economic output is processing rather than transportation.
This matters because railroad and fiber networks have especially strong physical network effects: a route becomes more valuable when it connects to additional routes. AI has powerful scale economies and ecosystem effects, but much of that advantage resides in models, data, software, cloud platforms, and proprietary chip architectures rather than merely in connecting one data-center building to another.
2. Technological progress affects the installed assets differently
With fiber, new electronics could greatly increase the capacity of fiber that was already buried. Technological improvement therefore often enhanced the usefulness of the old physical asset.
With AI, a new generation of accelerators may make the previous generation significantly less economical. The site, building, electrical connection, cooling system, and fiber links can remain useful, but the chips and servers may have to be replaced.
Meta’s 2025 annual report estimated useful lives of five to five-and-a-half years for servers and network assets, compared with 25–30 years for its buildings. Even five years may overstate the competitive life of some frontier AI equipment if newer chips provide much better performance per dollar or per watt.
This makes unused AI capacity more perishable than “dark fiber.” A buried fiber strand can sit unused for years and later be equipped with improved electronics. A warehouse full of unused accelerators loses value simply as better chips arrive.
3. AI data centers have unusually large continuing resource costs
After fiber has been installed, transmitting another unit of information has a very low marginal cost. Railroads have significant continuing costs—labor, fuel, maintenance, and rolling stock—but the track itself does not require a massive continuous energy input merely to remain economically relevant.
AI computation consumes large amounts of electricity whenever it is used and requires cooling, maintenance, networking, and periodic hardware replacement. The IEA expects global data-center electricity consumption to rise from approximately 485 terawatt-hours in 2025 to about 950 terawatt-hours in 2030. AI-related servers are expected to account for much of the increase.
Thus, an overbuilt AI center can continue generating costs even after the initial construction bill has been paid. Its economics depend not merely on filling the building but on earning enough revenue to cover electricity, maintenance, hardware depreciation, and financing.
4. Today’s leading builders are financially stronger than many railroad and fiber promoters were
The railroad boom was heavily financed through bonds. When railroad returns declined and European investors withdrew, defaults spread to banks and securities firms and helped produce the Panic of 1873.
The 1990s fiber boom included numerous new or rapidly expanding companies—Global Crossing, Level 3, Qwest and others—whose survival depended on continued access to capital and rapid growth in wholesale bandwidth demand. When demand disappointed, many failed or were reorganized.
The principal AI builders—Amazon, Microsoft, Alphabet and Meta—began with enormous revenues, cash flows, and valuable existing businesses. That makes a wave of immediate hyperscaler bankruptcies considerably less likely than nineteenth-century railroad failures or the failures of many fiber entrants.
The distinction is narrowing, however. Amazon, Alphabet, Meta and Oracle had issued about $194 billion of bonds during 2026 through July 7, 79 percent more than they issued in all of 2025. Goldman Sachs estimated that debt would finance roughly one-third of 2026 hyperscaler capital spending.
Companies are also shifting projects into leases and joint ventures. A $14 billion Meta–BlackRock data-center project in El Paso, for example, will be 80 percent owned by BlackRock-managed funds and partly financed with $12.5 billion of debt, while Meta leases the capacity.
My inference is that a downturn would be more likely to produce project cancellations, asset write-downs, lower returns, and losses for developers, lenders, lessors, utilities, and equipment suppliers than the wholesale disappearance of the leading technology companies. But increasing reliance on debt and outside capital means the financial consequences would no longer be confined to their own shareholders.
5. The return on AI investment is harder to observe
A railroad could measure freight and passenger revenue against the cost of operating a route. A wholesale fiber company could measure the bandwidth it leased and the price per unit of capacity.
Much AI infrastructure is used internally. AI may improve advertising, search, recommendation systems, programming, logistics, customer service, or the defensive position of an existing business. The return may not appear as separately identified “AI revenue.”
That gives the builders greater flexibility, but it also makes investment discipline more difficult. A company may continue spending because it fears falling behind, even when it cannot demonstrate that the next data center will earn an adequate direct return. Meta, for example, derives most of its revenue from advertising while making enormous infrastructure investments intended both to support that business and to develop new AI products.
6. The principal bottleneck is electricity rather than merely construction capital
Railroads competed for land and rights-of-way. Fiber companies needed trenches, conduits, poles, and access to local networks.
AI data centers compete above all for dependable electricity and rapid grid connections. A data center can be completed in two or three years, while new transmission, generation, and other energy infrastructure can require much longer. In some regions, expected data-center loads already exceed available capacity, causing projects to be delayed or paired with on-site natural-gas generation.
This creates a distinctive public-policy problem. Utilities may build generation and transmission for projected data-center demand, only to be left with underused assets if projects are cancelled or use less power than anticipated. Depending on rate design, households and other businesses could bear some of those costs.
7. The form of government assistance is different
Four of the five transcontinental railroads received federal land grants, giving them both rights-of-way and land they could sell to finance construction.
AI data centers have not received an equivalent nationwide transfer of federal land. Assistance has instead generally taken the form of state and local tax concessions, discounted or specially designed electricity tariffs, accelerated permitting, semiconductor subsidies, and publicly regulated grid investment. That creates similar questions about who receives the gains and who bears the costs, but through less visible mechanisms
Which historical analogy is the more useful one?
For the next several years, the fiber analogy is probably more informative. Both involve rapidly improving technology, uncertain demand for applications that do not yet fully exist, extrapolation of very high early growth rates, and the possibility that efficiency improvements will create more usable capacity than planners expected.
For the longer-run physical and economic consequences, the railroad analogy may be more informative. AI is driving investment not only in servers but also in power plants, transmission lines, substations, gas pipelines, batteries, cooling technology, chip factories, and fiber connections. Those investments may alter the economic geography of the United States and remain useful long after the first generation of AI chips has been replaced.
My tentative expectation is therefore a fiber-style shakeout within a railroad-scale infrastructure transformation:
Some announced data centers will never be completed.
Older accelerators will depreciate faster than expected.
Computing prices and profit margins may fall.
The strongest cloud and technology companies will absorb or outlast weaker developers.
Much of the underlying electrical, building, and communications infrastructure will survive and support applications that are not yet apparent.
That is an inference rather than a confident forecast. AI demand might grow quickly enough to absorb most of the planned capacity, especially as lower costs generate more intensive applications. But the historical comparisons show why “AI will transform the economy” and “too much AI capacity is currently being built” are entirely compatible propositions.
Bottom line
The three buildouts share a common economic pattern:
A genuinely revolutionary technology creates a race to build costly infrastructure ahead of uncertain demand. Competition and optimism produce duplication and sometimes overinvestment. Investors may suffer when expectations are revised, yet the surplus infrastructure can lower costs and accelerate the technology’s eventual diffusion.
The largest difference is that railroad rights-of-way and buried fiber are unusually durable, while the most expensive part of an AI data center—the computing equipment—is rapidly depreciating and must be continuously powered. The crucial question is therefore not simply whether demand for AI will grow. It almost certainly will. The question is whether monetizable demand will grow quickly enough to cover the extraordinary depreciation, electricity, and financing costs of the capacity now being installed.
This morning (May 8), the Bureau of Labor Statistics (BLS) released its “Employment Situation” report (often called the “jobs report”) for April. The report showed a stronger than expected increase in employment.
The jobs report has two estimates of the change in employment during the month: one estimate from the establishment survey, often referred to as the payroll survey, and one from the household survey. As we discuss in Macroeconomics, Chapter 9, Section 9.1 (Economics, Chapter 19, Section 19.1), many economists and Federal Reserve policymakers believe that employment data from the establishment survey provide a more accurate indicator of the state of the labor market than do the household survey’s employment and unemployment data. (The groups included in the employment estimates from the two surveys are somewhat different, as we discuss in this post.)
According to the establishment survey, there was a net increase of 115,000 nonfarm jobs during April. Economists surveyed by the Wall Street Journal had forecast an increase of only 55,000 jobs. Economists surveyed by Bloomberg had a slightly higher forecast of a net increase of 62,000 jobs. The BLS revised downward its previous estimates of employment in February and March by a combined 16,000 jobs. (The BLS notes that: “Monthly revisions result from additional reports received from businesses and government agencies since the last published estimates and from the recalculation of seasonal factors.”)
The following figure from the jobs report shows the net change in nonfarm payroll employment for each month in the last two years. The figure shows an unusual pattern in the job market since the middle of 2025 in which months of declining employment and months of increasing employment have been alternating. March and April of 2026 are the first back-to-back months of increasing net employment since March and April of 2025.
These fluctuations of net employment gains around roughly zero are consistent with a recent analysis from economists at the Federal Reserve Bank of Dallas that estimates the break-even rate of employment growth—the rate of employment growth at which the unemployment rate remains constant. They note that “continued net outflows of unauthorized immigrants, together with shifts in labor force participation, have pushed the monthly break-even employment growth lower than previously thought.” They conclude that: “The break-even rate [of employment growth] peaked at about 250,000 jobs per month in 2023, fell to roughly 10,000 by July 2025, and declined to near zero thereafter, averaging about –3,000 jobs per month from August to December 2025, indicating, if anything, a modest net jobs loss over this period.” In other words, in the current labor market, the break-even rate of employment growth may actually be negative.
The unemployment rate, which is calculated from data in the household survey, was 4.3 percent in April, unchanged from March. As the following figure shows, the unemployment rate has been remarkably stable over the past year and a half, staying between 4.0 percent and 4.4 percent in each month since May 2024. The Federal Open Market Committee’s current estimate of the natural rate of unemployment—the normal rate of unemployment over the long run—is 4.2 percent. So, unemployment is slightly above that estimate of the natural rate. (We discuss the natural rate of unemployment in Macroeconomics, Chapter 9 and Economics, Chapter 19.)
As the following figure shows, the monthly net change in jobs from the household survey moves much more erratically than does the net change in jobs from the establishment survey. As measured by the household survey, there was a net decrease of 226,000 in April, the fourth consecutive month of decreases. (Note that because of last year’s shutdown of the federal government, there are no data for October or November.) In any particular month, the story told by the two surveys can be inconsistent. In this case, the establishment survey shows a strong increase in net employment, while the household survey shows a decline.
The household survey has another important labor market indicator: the employment-population ratio forprime age workers—those workers aged 25 to 54. In April the ratio was 80.7 percent, the same as in February and March. The prime-age population ratio remains above its value for most of the period since 2001. The continued high levels of the prime-age employment-population ratio indicate continuing strength in the labor market.
There have been media reports of firms, including Salesforce, Cloudflare, Coinbase, and Freshworks, laying off workers in information systems. The following figure shows net employment changes in the BLS employment category of “computing infrastructure providers, data processing, web hosting, and related services.” Employment in this sector declined for the sixth straight month in April. Since November 2025, the sector has experienced a net decline of 23,000 jobs.
The establishment survey also includes data on average hourly earnings (AHE). As we noted in this post, many economists and policymakers believe the employment cost index (ECI) is a better measure of wage pressures in the economy than is the AHE. The AHE does have the important advantage of being available monthly, whereas the ECI is only available quarterly. The following figure shows the percentage change in the AHE from the same month in the previous year. The AHE increased 3.6 percent in April, up from 3.4 percent in March.
What effect is this jobs report likely to have on the decisions of the Federal Reserve’s policymaking Federal Open Market Committee at its next meeting on June 16–17, the first meeting with Kevin Warsh as chair? Although employment growth has been relatively slow in recent months, as noted earlier, even that slow rate may be close to the break-even rate of employment growth. So, it’s unlikely that the FOMC will see current conditions in the job market as warranting a cut in the committee’s target range for the federal funds rate. In addition, disruptions to the world oil market as a result of the conflict in Iran have caused oil prices to rise, putting upward pressure on the price level. And the effects of tariff increases have likely not yet fully passed through to increases in prices. These factors make it likely that the committee will keep its target range for the federal funds rate unchanged at its next meeting and may even begin considering future increases in the target range.
The probability that investors in the federal funds futures market assign to the FOMC keeping its target rate unchanged at its June meeting decreased slightly this afternoon to 93.9 percent, from 96.4 percent yesterday. Investors no longer assign a greater than a 50 percent probability to a rate cut occuring at any meeting through the end of 2027.
Today (February 5), the Bureau of Labor Statistics (BLS) released its “Job Openings and Labor Turnover” (JOLTS) report for December 2025. The report indicated that labor market conditions may be weakening. The following figure shows that the rate of job openings fell to 3.9 percent in December from 4.2 percent in November. The rate was 4.5 percent in October. The job openings rate is the lowest since April 2020, at the start of the Covid pandemic. We should note the usual caveat that the monthly JOLTS data is subject to potentially large revisions as the BLS receives more complete data.
(The BLS defines a job opening as a full-time or part-time job that a firm is advertising and that will start within 30 days. The rate of job openings is the number of job openings divided by the number of job openings plus the number employed workers, multiplied by 100.)
In the following figure, we show a measure of the state of the labor market that economists frequently use: the total number of job openings to the total number of people unemployed. In December there were 0.87 job openings per unemployed person, the lowest value for that measure since March 2021, during the recovery from the pandemic. The value was 1.0 in September. (Note that data for October and November are unavailable because the data weren’t collected during the shutdown of the federal government from October 1 to November 12 last year.) The value for December is well below the 1.21 job openings per employed person in February 2020, just before the pandemic. (Note that, as we discussed in this blog post, the employment-population ratio for prime age workers, which many economists consider a key measure of the state of the labor market, rose in December, putting it above what the ratio was in any month during the period from January 2008 to February 2020.)
The rate at which workers are willing to quit their jobs is an indication of how they perceive the ease of finding a new job. As the following figure shows, the quit rate declined slowly from a peak of 3 percent in late 2021 and early 2022 to 2.0 percent in August 2024, the same value as in December 2025. That rate is below the rate during 2019 and early 2020. By this measure, workers’ perceptions of the state of the labor market have remained remarkably stable over the last year and a half.
Overall, this JOLTS report is consistent with what some economists have labeled a “slow hire, slow fire” labor market. Fed Chair Jerome Powell’s remarks at his press conference following the last meeting of the Federal Open Market Committee (FOMC) indicates that Fed policymakers share this view, which Powell believes complicates monetary policymaking:
“So there are lots of … little places that suggest that the labor market has softened, but part of … payroll job softening is that both the supply and demand for labor has come down … growth in those two have come down. So that makes it a difficult time to read the labor market. So, imagine they both came down a lot, to the point where there is no job growth. Is that full employment? In a sense it is. If demand and supply are … in balance, you could say that’s full employment. At the same time, is it—do we really feel like … that’s a maximum employment economy? It’s a challenging—it’s very challenging and quite unusual situation.”
The BLS was scheduled to release its monthly “Employment Situation” report (often called the “jobs report”) for January 2026 tomorrow. Because of the temporary lapse in funding that began Saturday, the report will instead be released next Wednesday, February 11. That report will provide additional data on the state of the labor market. (Note that the data in the JOLTS report lag the data in the “Employment Situation” report by one month.)
Image generated by ChatGPT 5 of a 1981 IBM personal computer.
The modern era of information technology began in the 1980s with the spread of personal computers. A key development was the introduction of the IBM personal computer in 1981. The Apple II, designed by Steve Jobs and Steve Wozniak and introduced in 1977, was the first widely used personal computer, but the IBM personal computer had several advantages over the Apple II. For decades, IBM had been the dominant firm in information technology worldwide. The IBM System/360, introduced in 1964, was by far the most successful mainframe computer in the world. Many large U.S. firms depended on IBM to meet their needs for processing payroll, general accounting services, managing inventories, and billing.
Because these firms were often reliant on IBM for installing, maintaining, and servicing their computers, they were reluctant to shift to performing key tasks with personal computers like the Apple II. This reluctance was reinforced by the fact that few managers were familiar with Apple or other early personal computer firms like Commodore or Tandy, which sold the TRS-80 through Radio Shack stores. In addition, many firms lacked the technical staffs to install, maintain, and repair personal computers. Initially, it was easier for firms to rely on IBM to perform these tasks, just as they had long been performing the same tasks for firms’ mainframe computers.
By 1983, the IBM PC had overtaken the Apple II as the best-selling personal computer in the United States. In addition, IBM had decided to rely on other firms to supply its computer chips (Intel) and operating system (Microsoft) rather than develop its own proprietary computer chips and operating system. This so-called open architecture made it possible for other firms, such as Dell and Gateway, to produce personal computers that were similar to IBM’s. The result was to give an incentive for firms to produce software that would run on both the IBM PC and the “clones” produced by other firms, rather than produce software for Apple personal computers. Key software such as the spreadsheet program Lotus 1-2-3 and word processing programs, such as WordPerfect, cemented the dominance of the IBM PC and the IBM clones over Apple, which was largely shut out of the market for business computers.
As personal computers began to be widely used in business, there was a general expectation among economists and policymakers that business productivity would increase. Productivity, measured as output per hour of work, had grown at a fairly rapid average annual rate of 2.8 percent between 1948 and 1972. As we discuss in Macroeconomics, Chapter 10 (Economics, Chapter 20 and Essentials of Economics, Chapter 14) rising productivity is the key to an economy achieving a rising standard of living. Unless output per hour worked increases over time, consumption per person will stagnate. An annual growth rate of 2.8 percent will lead to noticeable increases in the standard of living.
Economists and policymakers were concerned when productivity growth slowed beginning in 1973. From 1973 to 198o, productivity grew at an annual rate of only 1.3 percent—less than half the growth rate from 1948 to 1972. Despite the widespread adoption of personal computers by businesses, during the 1980s, the growth rate of productivity increased only to 1.5 percent. In 1987, Nobel laureate Robert Solow of MIT famously remarked: “You can see the computer age everywhere but in the productivity statistics.” Economists labeled Solow’s observation the “productivity paradox.” With hindsight, it’s now clear that it takes time for businesses to adapt to a new technology, such as personal computers. In addition, the development of the internet, increases in the computing power of personal computers, and the introduction of innovative software were necessary before a significant increase in productivity growth rates occurred in the mid-1990s.
Result when ChatGPT 5 is asked to create an image illustrating ChatGPT
The release of ChatGPT in November 2022 is likely to be seen in the future as at least as important an event in the evolution of information technology as the introduction of the IBM PC in August 1981. Just as with personal computers, many people have been predicting that generative AI programs will have a substantial effect on the labor market and on productivity.
In this recent blog post, we discussed the conflicting evidence as to whether generative AI has been eliminating jobs in some occupations, such as software coding. Has AI had an effect on productivity growth? The following figure shows the rate of productivity growth in each quarter since the fourth quarter of 2022. The figure shows an acceleration in productivity growth beginning in the fourth quarter of 2023. From the fourth quarter of 2023 through the fourth quarter of 2024, productivity grew at an annual rate of 3.1 percent—higher than during the period from 1948 to 1972. Some commentators attributed this surge in productivity to the effects of AI.
However, the increase in productivity growth wasn’t sustained, with the growth rate in the first half of 2025 being only 1.3 percent. That slowdown makes it more likely that the surge in productivity growth was attributable to the recovery from the 2020 Covid recession or was simply an example of the wide fluctuations that can occur in productivity growth. The following figure, showing the entire period since 1948, illustrates how volatile quarterly rates of productivity growth are.
How large an effect will AI ultimately have on the labor market? If many current jobs are replaced by AI is it likely that the unemployment rate will soar? That’s a prediction that has often been made in the media. For instance, Dario Amodei, the CEO of generative AI firm Anthropic, predicted during an interview on CNN that AI will wipe out half of all entry level jobs in the U.S. and cause the unemployment rate to rise to between 10% and 20%.
Although Amodei is likely correct that AI will wipe out many existing jobs, it’s unlikely that the result will be a large increase in the unemployment rate. As we discuss in Macroeconomics, Chapter 9 (Economics, Chapter 19 and Essentials of Economics, Chapter 13) the U.S. economy creates and destroys millions of jobs every year. Consider, for instance, the following table from the most recent “Job Openings and Labor Turnover” (JOLTS) report from the Bureau of Labor Statistics (BLS). In June 2025, 5.2 million people were hired and 5.1 million left (were “separated” from) their jobs as a result of quitting, being laid off, or being fired.
Most economists believe that one of the strengths of the U.S. economy is the flexibility of the U.S. labor market. With a few exceptions, “employment at will” holds in every state, which means that a business can lay off or fire a worker without having to provide a cause. Unionization rates are also lower in the United States than in many other countries. U.S. workers have less job security than in many other countries, but—crucially—U.S. firms are more willing to hire workers because they can more easily lay them off or fire them if they need to. (We discuss the greater flexibility of U.S. labor markets in Macroeconomics, Chapter 11 (Economics, Chapter 21).)
The flexibility of the U.S. labor market means that it has shrugged off many waves of technological change. AI will have a substantial effect on the economy and on the mix of jobs available. But will the effect be greater than that of electrification in the late nineteenth century or the effect of the automobile in the early twentieth century or the effect of the internet and personal computing in the 1980s and 1990s? The introduction of automobiles wiped out jobs in the horse-drawn vehicle industry, just as the internet has wiped out jobs in brick-and-mortar retailing. People unemployed by technology find other jobs; sometimes the jobs are better than the ones they had and sometimes the jobs are worse. But economic historians have shown that technological change has never caused a spike in the U.S. unemployment rate. It seems likely—but not certain!—that the same will be true of the effects of the AI revolution.
Which jobs will AI destroy and which new jobs will it create? Except in a rough sense, the truth is that it is very difficult to tell. Attempts to forecast technological change have a dismal history. To take one of many examples, in 1998, Paul Krugman, later to win the Nobel Prize, cast doubt on the importance of the internet: “By 2005 or so, it will become clear that the Internet’s impact on the economy has been no greater than the fax machine’s.” Krugman, Amodei and other prognosticators of the effects of technological change simply lack the knowledge to make an informed prediction because the required knowledge is spread across millions of people.
That knowledge only becomes available over time. The actions of consumers and firms interacting in markets mobilize information that is initially known only partially to any one person. In 1945, Friedrich Hayek made this argument in “The Use of Knowledge in Society,” which is one of the most influential economics articles ever written. One of Hayek’s examples is an unexpected decrease in the supply of tin. How will this development affect the economy? We find out only by observing how people adapt to a rising price of tin: “The marvel is that … without an order being issued, without more than perhaps a handful of people knowing the cause, tens of thousands of people whose identity could not be ascertained by months of investigation are made [by the increase in the price of tin] to use the material or its products more sparingly.” People adjust to changing conditions in ways that we lack sufficient information to reliably forecast. (We discuss Hayek’s view of how the market system mobilizes the knowledge of workers, consumers, and firms in Microeconomics, Chapter 2.)
It’s up to millions of engineers, workers, and managers across the economy, often through trial and error, to discover how AI can best reduce the cost of producing goods and services or improve their quality. Competition among firms drives them to make the best use of AI. In the end, AI may result in more people or fewer people being employed in any particular occupation. At this point, there is no way to know.
“Artificial intelligence is profoundly limiting some young Americans’ employment prospects, new research shows.” That’s the opening sentence of a recent opinion column in the Wall Street Journal. The columnist was reacting to a new academic paper by economists Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen of Stanford University. (See also this Substack post by Chandar that summarizes the results of their paper.) The authors find that:
“[S]ince the widespread adoption of generative AI, early-career workers (ages 22-25) in the most AI-exposed occupations have experienced a 13 percent relative decline in employment … In contrast, employment for workers in less exposed fields and more experienced workers in the same occupations has remained stable or continued to grow. Furthermore, employment declines are concentrated in occupations where AI is more likely to automate, rather than augment, human labor.”
The authors conclude that “our results are consistent with the hypothesis that generative AI has begun to significantly affect entry-level employment.”
About a month ago, we wrote a blog post looking at whether unemployment among young college graduates has been abnormally high in recent months. The following figure from that post shows that over time, the unemployment rates for the youngest college graduates (the red line) is nearly always above the unemployment rate for the population as a whole (the green line), while the unemployment rate for college graduates 25 to 34 years old (the blue line) is nearly always below the unemployment rate for the population as a whole. In July of this year, the unemployment rate for the population as a whole was 4.2 percent, while the unemployment for college graduates 20 to 24 years old was 8.5 percent, and the unemployment rate for college graduates 25 to 34 years old was 3.8 percent.
As the following figure (also reproduced from that blog post) shows, the increase in unemployment among young college graduates has been concentrated among males. Does higher male unemployment indicate that AI is eliminating jobs, such as software coding, that are disproportionately male? Data journalist John Burn-Murdoch argues against this conclusion, noting that data shows that “early-career coding employment is now tracking ahead of the [U.S.] economy.”
Another recent paper written by Sarah Eckhardt and Nathan Goldschlag of the Economic Innovation Group is also skeptical of the view that firms adopting generative AI programs is reducing employment in certain types of jobs. They use a measure developed by Edward Felton on Princeton University, and Manav Raj and Robert Seamans of New York University of how exposed particular jobs are to AI (AI Occupational Exposure (AIOE)). The following table from Eckhardt and Goldschlag’s paper shows the five most AI exposed jobs and the five least AI exposed jobs.
They divide all occupations into quintiles based on the exposure of the occupations to AI. Their key results are given in the following table, which shows that the occupations that are most exposed to the effects of AI—quintiles 4 and 5—have lower unemployment rates and higher wages than do the occupations that are least exposed to AI.
The Brynjolfsson, Chandar, and Chen paper mentioned at the beginning of this post uses a larger data set of workers by occupation from ADP, a private firm that processes payroll data for about 25 percent of U.S. workers. Figure 1 from their paper, reproduced here, shows that employment of workers in two occupations—software developers and customer service—representative of those occupations most exposted to AI declined sharply after generative AI programs became widely available in late 2022.
They don’t find this pattern for all occupations, as shown in the following figure from their paper.
Finally, they show results by occupational quintiles, with workers ages aged 22 to 25 being hard hit in the two occupational quintiles (4 and 5) most exposted to AI. The data show total employment growth from October 2022 to July 2025 by age group and exposure to AI.
Economics blogger Noah Smith has raised an interesting issue about Brynjolfsson, Chandar, and Chen’s results. Why would we expect that the negative effect of AI on employment to be so highly concentrated among younger workers? Why would employment in the most AI exposed occupations be growing rapidly among workers aged 35 and above? Smith wonders “why companies would be rushing to hire new 40-year-old workers in those AI-exposed occupations.” He continues:
“Think about it. Suppose you’re a manager at a software company, and you realize that the coming of AI coding tools means that you don’t need as many software engineers. Yes, you would probably decide to hire fewer 22-year-old engineers. But would you run out and hire a ton of new 40-year-old engineers?“
Both the papers discussed here are worth reading for their insights on how the labor market is evolving in the generative AI era. But taken together, they indicate that it is probably too early to arrive at firm conclusions about the effects of generative AI on the job market for young college graduates or other groups.
Supports: Macroeconomics, Chapter 9,Economics, Chapter 19, and Essentials of Economics, Chapter 13.
Image generated by ChatGPT
A recent article on axios.com notes that from April 2023 to July 2024, the U.S. economy generated an average net increase of 177,000 jobs per month. Despite that job growth, the unemployment rate during that period increased by 0.8 percentage point. The article observes that: “At first glance, the combination of a rising unemployment rate and strong jobs growth simply does not compute.” How is it possible during a given period for both total employment and the unemployment rate to increase?
Solving the Problem Step 1: Review the chapter material. This problem is about calculating the unemployment rate, so you may want to review Chapter 9, Section 9.1, “Measuring the Unemployment Rate, the Labor Force Participation Rate, and the Employment-Population Ratio.”
Step 2: Answer the question by explaining how it’s possible for both the total number of people employed and the unemployment rate to both increase during the same period. The unemployment rate is equal to the number of people unemployed divided by the number of people in the labor force (multiplied by 100). The labor force equals the sum of the number of people employed and the number of people unemployed.
Let’s consider the situation in a particular month. Suppose that the unemployment rate in the previous month was 4 percent. If, during the current month, both the number of people employed and the number of people unemployed increase, the unemployment rate will increase if the increase in the number of people unemployed as a percentage of the increase in the labor force is greater than 4 percent. The unemployment rate will decrease if the increase in the number of people unemployed as a percentage of the increase in the labor force is less than 4 percent.
Consider a simple numerical example. Suppose that in the previous month there were 96 people employed and 4 people unemployed. In that case, the unemployment rate was (4/(96 + 4)) x 100 = 4.0%.
Suppose that during the month the number of people employed increases by 30 and the number of people unemployed increases by 1. In that case, there are now 126 people employed and 5 people unemployed. The unemployment rate will have fallen from 4.0% to (5/(126 + 5)) x 100 = 3.8%.
Now suppose that the number of people employed increased by 30 and the number of people unemployed increases by 3. The unemployment will have risen from 4.0% to (7/(126 + 7)) x 100 = 5.3%.
We can conclude that if both the total number of people employed and the total number of people unemployed increase during a during a period of time, it’s possible for the unemployment rate to also increase.
As we noted in yesterday’s blog post, the latest “Employment Situation” report from the Bureau of Labor Statistics (BLS) included very substantial downward revisions of the preliminary estimates of net employment increases for May and June. The previous estimates of net employment increases in these months were reduced by a combined 258,000 jobs. As a result, the BLS now estimates that employment increases for May and June totaled only 33,000, rather than the initially reported 291,000. According to Ernie Tedeschi, director of economics at the Budget Lab at Yale University, apart from April 2020, these were the largest downward revisions since at least 1979.
The size of the revisions combined with the estimate of an unexpectedly low net increase of only 73,000 jobs in June prompted President Donald Trump to take the unprecedented step of firing BLS Commissioner Erika McEntarfer. It’s worth noting that the BLS employment estimates are prepared by professional statisticians and economists and are presented to the commissioner only after they have been finalized. There is no evidence that political bias affects the employment estimates or other economic data prepared by federal statistical agencies.
Why were the revisions to the intial May and June estimates so large? The BLS states in each jobs report that: “Monthly revisions result from additional reports received from businesses and government agencies since the last published estimates and from the recalculation of seasonal factors.” An article in the Wall Street Journal notes that: “Much of the revision to May and June payroll numbers was due to public schools, which employed 109,100 fewer people in June than BLS believed at the time.” The article also quotes Claire Mersol, an economist at the BLS as stating that: “Typically, the monthly revisions have offsetting movements within industries—one goes up, one goes down. In June, most revisions were negative.” In other words, the size of the revisions may have been due to chance.
Is it possible, though, that there was a more systematic error? As a number of people have commented, the initial response rate to the Current Employment Statistics (CES) survey has been declining over time. Can the declining response rate be the cause of larger errors in the preliminary job estimates?
In an article published earlier this year, economists Sylvain Leduc, Luiz Oliveira, and Caroline Paulson of the Federal Reserve Bank of San Francisco assessed this possibility. Figure 1 from their article illustrates the declining response rate by firms to the CES monthly survey. The figure shows that the response rate, which had been about 64 percent during 2013–2015, fell significantly during Covid, and has yet to return to its earlier levels. In March 2025, the response rate was only 42.6 percent.
The authors find, however, that at least through the end of 2024, the falling response rate doesn’t seem to have resulted in larger than normal revisions of the preliminary employment estimates. The following figure shows their calculation of the average monthly revision for each year beginning with 1990. (It’s important to note that they are showing the absolute values of the changes; that is, negative change are shown as positive changes.) Depite lower response rates, the revisions for the years 2022, 2023, and 2024 were close to the average for the earlier period from 1990 to 2019 when response rates to the CES were higher.
The weak employment numbers correspond to the period after the Trump administration announced large tariff increases on April 2. Larger firms tend to respond to the CES in a timely manner, while responses from smaller firms lag. We might expect that smaller firms would have been more likely to hesitate to expand employment following the tariff announcement. In that sense, it may be unsurprising that we have seen downward revisions of the prelimanary employment estimates for May and June as the BLS received more survey responses. In addition, as noted earlier, an overestimate of employment in local public schools alone accounts for about 40 percent of the downward revisions for those months. Finally, to consider another possibility, downward revisions of employment estimates are more likely when the economy is heading into, or has already entered, a recession. The following figure shows the very large revisisons to the establishment survey employment estimates during the 2007–2010 period.
At this point, we don’t fully know the reasons for the downward employment revisions announced yesterday, although it’s fair to say that they may have been politically the most consequential revisions in the history of the establishment survey.
HubbardOBrienEconomics.com Blog 