Artificial Intelligence Revenues Would Need to Grow by 80 Percent Annually

Stijn Van Nieuwerburgh of Columbia Business School calculates that AI investment in the United States from 2025 through 2032 will total around 9 trillion euros. For that to pay off, AI revenues, measured against today’s sales of OpenAI and Anthropic, would have to grow by around 80 percent year after year through 2032.

The scale of the buildout

Van Nieuwerburgh, a professor of real estate and finance, presented his paper on 25 September at the Brookings Institution’s fall conference. According to the paper, investment in AI infrastructure in 2025 and in the first quarter of 2026 made a noticeable contribution to U.S. GDP growth.

For the years 2025 through 2032, the study puts investment in data centers as well as power, grid, and chip infrastructure at around 9 trillion euros. That equals around 3.6 percent of U.S. GDP per year.

In the study’s central scenario, around 183 gigawatts of additional data-center capacity come online over that period. Around 57 gigawatts are currently installed in the United States.

The railroad boom averaged 2.24 percent of gross domestic product over two decades, the highest average among the investment waves examined. Canal building, electrification, highway construction, and the fiber-optic boom stayed well below that. Planned AI investment, at around 3.6 percent, sits well above it.

What is Private Credit? Private Credit refers to loans issued by specialized funds and financial investors rather than by banks. These financings run largely outside regulated capital markets and are hard for outsiders to follow. In the AI buildout, Private Credit firms finance, among other things, data centers that corporations spin out through joint ventures.

The industry’s revenue gap

The revenue side has so far been far smaller. By 2032, the industry would, on that math, need around 3.3 trillion euros in annual sales for the investment to pay off. OpenAI and Anthropic currently generate around 88 billion euros in combined revenue.

From that base, the investment plans imply annual revenue growth of around 80 percent. Measured against today’s rental prices for GPU compute, the study does not treat the required revenue as unrealistic. What remains open, in its view, is whether demand will be strong enough for those prices to stay at today’s level even as capacity grows sharply.

At the hyperscalers, the relationship between investment and operating cash flow is tipping. Through 2024, the combined operating cash flow of Oracle, Amazon, Alphabet, Microsoft, and Meta sat well above their capital expenditure. In 2026, the five companies’ investment is likely to exceed that cash flow for the first time, at a projected more than 700 billion euros.

A further difficulty lies in the maturities. The financings are structured for long horizons, while graphics cards and other data-center components can lose value much earlier.

Financing off the balance sheet

According to the paper, the risks are migrating from the transparent on-balance-sheet financing of large corporations into off-balance-sheet structures such as joint ventures, private credit, securitizations, special purpose vehicles, lease obligations and credit guarantees. Speaking to journalists, Van Nieuwerburgh called the constructions “freaking complicated.” The opacity of the special purpose vehicles, he said, in parts recalls the subprime mortgage crisis.

One example is Meta’s Hyperion data center. The group sold private-credit provider Blue Owl 80 percent of the project for around 2.2 billion euros. The associated joint venture Beignet took on the financing of the construction.

In October 2025, Beignet raised around 24 billion euros in debt and thereby financed around 90 percent of the project. Standard & Poor’s rated the debt A+, one notch below Meta’s own credit standing.

According to Morgan Stanley estimates cited in the paper, more than half of the around 2.6 trillion euros for additional compute capacity of the hyperscalers from 2025 to 2028 comes from external financiers. On the debt side, private credit accounts for the largest share, at around 700 billion euros.

Extensive lease obligations come on top. According to Moody’s figures, they add up at the hyperscalers to around 850 billion euros, of which around 580 billion euros do not yet appear on the balance sheet.

Risk assessment and transparency

The study sees in the combination of uncertain demand, rapid technological change, execution bottlenecks and high indebtedness a tangible downside risk if expectations shift. Financial distress, on this assessment, is nonetheless not imminent. The author also considers it premature to speak already of a systemic risk of the kind seen in earlier credit booms.

The paper names greater measurability and transparency in the industry’s capital structure as the most important contribution from policy. More clarity about the financing constructs would also help IT leaders assess the stability of their data-center and cloud providers.

Frequently Asked Questions

Image source: AI-generated (September 2026)

More from the MBF Media Network