The Federal Reserve paper reveals the real gap behind the AI boom: productivity dividends have yet to be realized, and the cost of inflation is limited, but the impact of employment is difficult to quantify
The Zhitong Finance App learned that according to a research paper published by Federal Reserve staff, demand for artificial intelligence (AI) has driven a surge in US data center investment and raised the prices of some technology inputs, but its impact on overall productivity and employment is still limited. This means that the current AI boom is driving the US economy more reflected in the increase in investment brought about by data center construction and procurement of chips and related equipment, and whether AI technology can actually be transformed into a wide range of enterprise efficiency improvements and productivity dividends has yet to be tested over time.
According to reports, this study entitled “Artificial Intelligence and the U.S. Economy: Current Evidence” (Artificial Intelligence and the U.S. Economy: Current Evidence) was written by Nick Carollo, Leland Crane, Daniel Garcia Molina, and Joseph Nichols and released in November. The author points out that although AI is likely to have a huge and long-lasting productivity effect over a long period of time, only a few years have passed since the launch of ChatGPT, and the spread of technology, corporate organizational adjustments, and implementation of economic effects are still in the early stages. It should be noted that this is a discussion paper by the Federal Reserve staff. The analysis and conclusions represent the author's views, and do not mean that the Federal Reserve Board of Governors or other staff agree with it.
Data center investment plans exceed $1 trillion, but import spending is offsetting part of the economic impact
The paper found that the investment plan to build a new data center in 2025 exceeds 1 trillion US dollars, up from nearly 300 billion US dollars in 2024. According to projects announced as of May 2026, under a neutral scenario, the annualized scale of investment related to data centers is expected to exceed US$880 billion by the end of 2027, equivalent to more than 2% of GDP.
According to Morgan Stanley's estimates, the cash capital expenditure of hyperscalers (hyperscalers) and emerging cloud service providers (neoclouds) is expected to continue to grow, rising from $946 billion in 2026 to $1.66 trillion in 2027, and further reaching $1.77 trillion in 2028. Alphabet (GOOGL.US) and Amazon (AMZN.US) account for the largest share of projected spending. The combined cash capital expenditure of the two companies is expected to reach $730 billion in 2027 and approximately $765 billion in 2028.
However, the scale of data center investment does not mean an increase in domestic economic output on the same scale. The authors estimate that about half of related expenses go to imported goods, especially high-tech equipment, so the increase in imports will offset the boosting effect of some investment on US domestic output. At the same time, data center construction may also drive supporting investment such as power infrastructure, but rising prices of high-tech equipment may also squeeze the investment budgets of other companies.
There is also uncertainty about the project implementation itself. Data center construction requires land, electricity, water resources, labor, and a large amount of equipment. Any shortage of resources may drive up construction costs and delay project progress. Some local governments have limited data center projects due to electricity and water demand, and queuing power grids and shortages in the supply of gas turbines and chips also pose potential bottlenecks. The paper points out that the delivery cycle for some natural gas turbines has been extended from two years to more than five years.
The AI construction boom is driving up the prices of some technology products. The impact of inflation is noteworthy but limited
The boom in data center construction has also boosted the prices of products such as memory chips, data storage devices, bare circuit boards, and assembled circuit boards. In the 12 months up to March 2026, the price of personal consumption expenditure (PCE), which covers video, audio, photography, and information processing equipment and media, increased by nearly 8%; in the previous three years, the price of such products fell by an average of about 3% per year.
The paper points out that the rise in the price of software and accessories is also an important factor in the price increase in this category, and some software companies may be passing on the cost of adding AI functions to users. The rise in the price of data storage devices is also driving related consumer prices.
However, the author believes that up to now, the impact of these price increases on the overall PCE price index is still obvious but limited, and the impact may expand further in the future. Related research also found that since January 2025, the core PCE price category with the highest data center exposure had an inflation rate of about 5 percentage points higher than the median exposure category.
This means that AI infrastructure construction may drive up some costs in the short term through chips, electronic equipment, software, and electricity, and the cost reduction effects brought about by increased productivity have not yet been fully demonstrated. Federal Reserve researchers believe that if AI can significantly increase productivity in the future, it may indeed reduce inflation by reducing corporate production costs, but it is not yet possible to observe this effect.
The scope of AI applications continues to expand, but the depth of enterprise use is still insufficient
The scope of application of AI in consumers and enterprises continues to expand, but the degree of popularity is still uneven. A 2025 survey found that about 50% of adults use AI outside of work, mainly for information gathering and writing; while multiple surveys up to the beginning of 2026 showed that more than 40% of workers use AI at work. However, only about 10% of workers use AI at work every day, indicating that this technology has not yet become a core component of most employees' daily workflows.
The survey results at the enterprise level varied widely. According to the US Census Bureau's Business Trends and Prospects Survey, about 20% of companies use AI; other studies suggest 58% to 78%. This difference relates to the specific wording of the survey questions, sample size, and statistical weight. Overall, large enterprises and enterprises that have been established for a short time are more likely to use AI, and the degree of application in industries such as information services, financial activities, and professional and business services is also relatively leading.
However, enterprises currently tend to use AI in only a few business areas, such as IT, human resources, marketing, finance, and strategic management, rather than fully integrating it into the entire organization. It is particularly noteworthy that the proportion of AI applications in R&D is still relatively low, and R&D may be an important channel for AI to drive long-term innovation and productivity growth.
Federal Reserve researchers believe that to achieve continuous and widespread productivity improvement, enterprises may need to further incorporate AI into core production processes, redesign organizational structures, automate more tasks, and invest in supporting resources such as training and management. Simply buying AI tools or giving employees access to them is not enough to guarantee a significant increase in productivity.
The impact of AI on the job market is already showing, but the overall impact is still difficult to quantify
The impact of AI on the labor market is still mixed. A study cited in the paper found that after the launch of a large-scale language model, the number of workers between the ages of 22 and 25 employed in occupations with high AI exposure decreased by 16% compared to workers in the same occupation in other age groups and workers in occupations with low AI exposure. Other studies have come to similar conclusions.
The author points out that this impact is mainly reflected in reduced recruitment demand rather than large-scale layoffs. First-time workers are likely to be more vulnerable because their jobs often include more simple, standardized, and easy-to-code tasks, while experienced workers often have expertise and in-house experience that is more difficult to replace with AI.
At the same time, AI may also lead to job redistribution. Some companies prefer retraining affected employees rather than outright layoffs. A New York Federal Reserve survey found that when faced with jobs where AI may be affected, the proportion of companies reporting employee retraining is higher than layoffs.
As a result, the impact of AI on overall employment and unemployment remains difficult to discern. Currently, the most affected occupational groups account for a relatively limited share of total employment in the US, but as the scope and depth of AI adoption by enterprises continues to expand, this impact may be further evident in the future.