The takeaway

Firms have intensified their research and development during the AI boom. They’ve reduced their cash holdings, too, likely to pay for R&D. In this post, we use FRED data to explore these trends and the future of AI-related investment.

Background research

Two recent posts from the St. Louis Fed offer insights into AI-related investment:

• A relationship seems likely between rising R&D expenditures and declining cash reserves relative to total assets.
• Significant AI-related investment boosted real GDP growth in 2025 beyond even what IT components contributed during the dot-com boom.

Another analysis projects the AI boom will require $6.7 trillion worldwide in capital expenditures by 2030 to keep pace with the demand for computing power.*

What the data show us

In our FRED graph above, the green dashed line in effect shows the cash holdings of US nonfinancial firms in the corporate business sector as a share of the firm’s assets. We define cash holdings as the sum of checkable deposits and currency, total time and savings deposits, and money market fund shares. (For more about the data, see the Z.1 tables from the Board of Governors’ US Financial Accounts.)

Cash holdings increased from close to 2% in 1990 to close to 4% in the year before the Global Financial Crisis. It increased during that crisis but remained just above 4% in the 2010s. It increased again during the COVID-19 recession, reaching more than 6%. Then it decreased and has fluctuated around 5.5% recently.

With data from the Bureau of Economic Analysis, the solid blue line shows the increase in capital expenditures typically associated with the AI boom. It’s the ratio of private fixed investment in information processing equipment and software to GDP.

The path of this ratio has three distinct phases.

• This ratio surged through the 1990s and peaked at the dot-com crash, followed by a steep decline from 2000 to 2002.
• The ratio remained relatively stable from 2002 through the Great Financial Crisis and into 2023.
• It has jumped sharply from 3.9% in the third quarter of 2023 to 4.7% in the fourth quarter of 2024. It has now surpassed the fourth quarter 2000 peak for the first time.

Future AI investment

As AI-related investment expands further, cash holdings won’t be sufficient to fund it and firms will be more dependent on external financing. Stijn Van Nieuwerburgh argues that the AI buildout has been changing who owns and finances AI infrastructure, as hyperscalers are moving away from fully self-funding data centers and are increasingly combining owned capacity with leased facilities, joint ventures, and partnerships with specialized third-party developers. Monitoring both the adequacy of internal funding and the availability of external finance will be critical for assessing the health of the AI boom.

*“The cost of compute: A $7 trillion race to scale data centers,” April 28, 2025, McKinsey Quarterly, McKinsey & Co.

How this graph was created: Search FRED for and select “TABSNNCB.” Click “Edit Graph”: Use the “Customize” field to search for and select “BOGZ1FL103020005Q,” “TSDABSNNCB,” and “BOGZ1FL103034000Q,” and add the series. Insert (b+c+d)/a in the formula field. Use the “Add Line” tab to search for and select “GDP.” Then add series “A679RC1Q027SBEA” and insert b/a in the formula field. Use the “Format” tab to change the line styles.

Suggested by Masataka Mori and Juan Sanchez.

Data providers may use the same labels for their data even if their methods of collecting the data differ. FRED can help you compare and understand these differences.

Our FRED graph above displays two types of mortgage rates from two different sources: the weekly 30-year and 15-year fixed mortgage rates reported by Freddie Mac (solid lines) and Bankrate (dashed lines). Both data sets show similar interest rates, even though the sources use different methodologies to collect their data. Freddie Mac calculates the average rate on “thousands of mortgage loan applications” from lenders across the country when a borrower applies for a loan. Bankrate reports data from a survey of the “10 largest banks and thrifts in 10 large US markets.”

Use FRED’s graphing features to compare data series visually: This FRED graph shows the entirety of both series, and this FRED graph adds a formula to reveal the differences between the two series, which at times were notably pronounced. You can also download data from FRED to compare series quantitatively and read the FRED series notes to better understand the collection methods used.

We provide another interest rate data comparison for researchers and sleuths with the set of benchmark data from the G.19 Consumer Credit release reported by the Board of Governors of the Federal Reserve System:

• Auto loans at 48 months and 60 months
• Credit cards
• Personal loans

Here are all the recently added series from the Bankrate Monitor National Index and a link to an earlier post discussing recent patterns in interest rates on bank accounts.

How this graph was created: Search FRED for and select “30-Year Fixed Rate Mortgage Average in the United States.” Click on the “Edit Graph” button and select the “Add Line” tab to search for “Bankrate Monitor (BRM): Fixed Mortgage Rate – 30 Year Fixed.” Don’t forget to click “Add data series.” Repeat the last two steps to add data on “15-Year Fixed Rate Mortgage Average in the United States” and “Bankrate Monitor (BRM): Fixed Mortgage Rate – 15 Year Fixed.”

Suggested by Diego Mendez-Carbajo.

Our FRED graph above shows that unemployment is almost always doing one of two things: (1) declining slowly during expansions or (2) rising rapidly during recessions. Friedman (1964, 1993) compared this behavior to playing a musical instrument, describing it as a “plucking model” of unemployment.

Over the past 3 years, however, the unemployment rate has done something it almost never does: It has risen slowly from a low level, but there has been no sharp rise accompanied by a recession.

Our non-FRED graph below* illustrates how unusual this behavior is.

• The jagged red line is the US civilian unemployment rate.
• The smooth green line is a 25-month moving average of the unemployment rate that smooths out small movements in the rate.
• The green triangles show the minima of this moving average.
• The blue circles show the unemployment rate 36 months after the minimum.

In nearly every case —except 2020 (COVID) and 2023 — the unemployment rate climbs significantly in the 36 months after a local minimum. 2020 was exceptional because the huge COVID-related spike in unemployment rose and fell within 36 months. The past three years (2023-2026) have also been exceptional in that unemployment has risen very modestly and slowly from a very low level, but there has been no recession and no sharp uptick in unemployment.

It’s not obvious how to explain this very unusual behavior, but recent economic activity has been extraordinary. The 2020 recession associated with COVID was unprecedented, and the accompanying fiscal and monetary stimuli were quite large, leaving many consumers flush with cash. Unemployment rose to record levels, and then it declined at a record rate to a very low level. The 2020 recession was followed by supply shocks from the Russian invasion of Ukraine in 2022. In short, strange inputs tend to produce strange outcomes.

How these graphs were created: First graph: Search FRED for and select “unemployment rate.” *Second graph: FRED’s a great graphing tool, but can’t do everything: You can import data from FRED to your favorite application and create custom graphs like this one, which displays moving averages and only selected data points.

Suggested by Christopher Neely.