The amounts of money being spent on the hardware and buildings behind the new AI boom may turn out to be enormously large by historical standards. How is it being financed?
Stijn Van Nieuwerburgh tells the story in “Financing the AI Buildout” (Brookings Papers on Economic Activity, Fall 2026, conference draft version). As he points out, the new AI technologies are often discussed as a new form of software, and how the capabilities of that software are developing. The software perspective is true, but partial. The new AI technologies are also a physical capital boom. Indeed, our public discourse on AI seems to have entered a period where one set of concerns is that Ai software will be so cheap and available that it will lead to mass unemployment, while another set of concerns is that building the physical capital behind the AI boom–including data centers, semiconductor plants, additional energy production, and more–will be extraordinarily expensive. I’m alwas loathe to make predictions, but I feel confident in saying that the new AI technologies will not be both cheap and expensive at the same time.
Van Nieuwenbergh walks through a set of calculations to forecast the size physical capital AI investment in the next few years, based on current plans and the recognition that not all the current plan will actually happen. He produces this historical comparison, which needs to be taken with a few grains of salt. The AI buildout is a prediction, while the others actually happened. Also, the time periods vary: the railroad investment, the second-highest number on the right-hand column, lasted much longer than these predictions for the AI buildout. Still, the estimates for the cost of the AI buildout are striking.
Where are firms getting the funding for this buildout? Remember, stock market wealth is not an immediate source of funding–unless you sell the stock to someone else. One of the economic puzzles of the last decade or so is that big US tech companies were holding huge amounts of cash, more than $1 trillion (for example, I wrote about this back in 2013, here and here). A common question has been: What investment opportunity were they waiting for? The answer appears to be the AI buildout. Van Nieuwenbergh writes:
Aggregate capital expenditures by Oracle, Microsoft, Amazon, Meta, and Alphabet rose from roughly $97 billion in 2020 to more than $400 billion in 2025 and are projected to exceed $800 billion in 2026. On current estimates, capex will surpass the five firms’ combined operating cash flow for the first time in 2026. The figure therefore marks an important transition: internal cash generation remains substantial, but it is no longer sufficient to finance the projected pace of investment without greater reliance on external capital or financing structures that shift assets and obligations away from the operating companies’ balance sheets.
The new outside financing is a combination of equity and debt finance. The equity finance is arriving in the form of “[s]pecialized data center developers, REITs [real estate investment trusts], infrastructure funds, and private equity investors.” The debt finance includes bank loans and bonds, but private credit funds provide a growing share.
For those with an interest in private credit in particular, Puriya Abbassi, Iñaki Aldasoro, and Sebastian Doerr provide an overview of “Financing the digital economy: the role of private credit “ (BIS Quarterly Review, September 2026, pp. 17-30). In contrast to bank loans, which often require that a borrower have physical collateral to back up the loan, private credit funds make loans based on expected cash flows and intangible assets. As banks have become more tightly regulated, private credit funds have expanded, and as of 2025 had loan amounts outstanding of $2.5 trillion. As the authors note: “Aggregate [private credit] borrowing by tech firms rose from around $22 billion (22% of total private credit) in 2010 to over $1 trillion (44%) by 2025.” These private credit funds typically operate by getting money from investors that is usually committed for the middle-term of five to eight years, and cannot be withdrawn. If a bank has many loans that go sour, the bank has been lending out money of depositors, which can cause a problem; if a private credit fund loan portfolio goes sour, the investor in that fund would suffer losses, but there is much less reason to think that those losses would propogate into banking or the rest of the financial system.
One of the strengths of a market economy is that some companies can succeed, while others fail, without endangering the health of the economy as a whole. But when the scale of the investment gets this big, and the risks are distributed across a web of interlocking companies, it’s natural to wonder about which investors will bear the losses if the AI boom doesn’t generate revenue as fast as expected–and whether losses from those investors might be substantial enough to affect the broader economy. Some readers will remember the dot-com boom-and-bust when the overall stock market fell by about 50% from March 2000 to October 2002, and many internet companies fell by even more. There was a recession for eight months in 2001. But although the economy took a hit and a number of companies vanished outright, the technology of the internet didn’t go away, and the physical capital boom behind the internet was put to productive use.
In some ways, the physical capital boom surrounding AI puts me in mind of a rhetorical question I remember hearing about the California gold rush in the mid-1800s: During the California gold rush, which group got richer: the gold miners, or the people who sold picks and shovels to the gold miners? Many of the gold miners were unsuccessful, so the average gain for a gold miner wasn’t all that high. But every gold miner, successful or not, needed to buy the tools. Right now, the stock market prices of AI software companies are sky-high. But some of those companies are likely to become unsuccessful goldminers–who nonetheless had to pay the costs for the physical investment in the new technologies.
