The Bubble of Artificial Intelligence and Its Illusions

Carl Benedikt Frey: An Associate Professor of Artificial Intelligence and Work at the Oxford Internet Institute and Director of the Future of Work Programme at the Oxford Martin School, his latest book is titled “How Progress Ends: Technology, Creativity, and the Fate of Nations” (Princeton University Press, 2025).

When OpenAI recently allocated $1.4 trillion to secure future computing power, it was merely the latest indicator of the reckless abundance in 2025. According to some estimates, nearly all of the growth in U.S. GDP in the first half of this year came from data centers, prompting a flood of commentary about when the bubble might burst and what consequences this explosion might leave behind. While the dot-com boom ended in the late 1990s with ugly repercussions on Wall Street, efforts were underway on Main Street to sustain what truly mattered: infrastructure. Productivity soared, and the groundwork laid during the boom years continues to be effective today. Former U.S. President Bill Clinton’s pledge to build a “bridge to the 21st century” was one of those rare campaign promises that was actually fulfilled.

The reality is that today’s investments in artificial intelligence could bear fruit similar to that of internet investments. However, current gains appear modest, while macroeconomic negatives loom larger than in the case of the dot-com bubble. Consider the potential benefits: in the late 1990s, internet gains emerged while the bubble was still inflating, with average labor productivity growth in the U.S. at around 2.8% from 1995 to 2004—almost double the pace of the previous two decades—before tapering off in the mid-2000s. You could see gains in national accounts even while Pets.com was buying its infamous Super Bowl ads.

This time, U.S. labor productivity growth has revived after two decades of stagnation, reaching around 2.7% last year. However, it is too early to conclude that artificial intelligence is the cause. In fact, AI adoption appears to be declining, as a recent survey by the U.S. Census Bureau showed a drop in AI usage among large companies. If the recent increase in productivity is primarily linked to AI, we should expect it to fade as its adoption wanes—once again reminding us of how rapidly these waves can recede. Just as the tech boom of the 1990s became apparent almost overnight, it faded within a decade or so.

It is tempting to imagine that large language models (LLMs) will accelerate the pace of creativity and discovery by uncovering hidden links in academic literature, writing code, and formulating protocols. New tools—from Robert Hooke’s microscope to the Galileo telescope—have facilitated such leaps before. However, this time we already have the perfect research tool in the form of an internet-connected personal computer. Yet, even with instant access to accumulated knowledge and exceptional global talent, measures of research productivity and breakthrough creativity are declining. Sustaining Moore’s Law—which states that computing power doubles approximately every two years—now requires more researchers than were needed in the early 1970s.

It is also unclear whether the current boom in capital expenditures will leave behind a substantial amount of lasting digital infrastructure. Much like the railways of the 19th century, the dot-com era poured money into long-term assets—especially fiber optics and core networks—that could be “lit up” and re-lit as electronics improved. A significant portion of this fiber still carries traffic today, supporting generations of technology and business models.

In contrast, AI does not lay down tracks but operates a tool that walks in place. Electronic chips and memory become outdated within years, not decades. Each server used to train a large language model now requires 120 kilowatts of power, up from around 5-10 kilowatts a decade ago. While each new generation of graphics processors lowers the cost per watt, this means that super cloud services are expanding faster while older equipment becomes economically obsolete. Whereas fiber lasts as we swap endpoints, the stack of AI technologies is rapidly shrinking, necessitating relentless reinvestment.

Controlling this situation might be possible if the macroeconomic picture mirrored that of 1999. However, it does not. Though real interest rates were higher back then, budget surpluses and declining debt-to-GDP ratios during Clinton’s administration alleviated pressure on capital markets, keeping government interest payments lower and thus limiting crowding out.

This time, the situation has reversed. The ongoing U.S. government deficit of 6% of GDP (around $1.8 trillion) and net interest payments of $1 trillion have not only reduced financial space but are now expected to fund the same savings group for clean energy construction, rising defense budgets, and a surge in energy-hungry data centers. In practice, this demand manifests as higher borrowing costs, discouraging new housing construction and pushing long-awaited infrastructure projects to the back of the priority list.

Public finances are also affected. An increased debt stock implies that positive real interest rates will quickly translate into higher interest payment bills, likely crowding out programs that families rely upon. During the surplus period of the late 1990s, debt declined, and the Treasury even bought back bonds, meaning the government could invest alongside the private boom without crowding it out. Today, rising borrowing and a larger interest bill leave less room for maneuver when growth slows. If AI returns do indeed materialize but slowly, the calculations become more challenging. We will see more dollars go to bondholders and less to Social Security, healthcare, and essential services; if the business cycle turns downward, the trade-offs will be sharper.

Financing has also changed. The downturn in the early 2000s was mostly a stock story: stock prices plummeted, and venture capital investors targeting long-term returns suffered a severe blow; yet, as painful as it was, its intensity faded relatively quickly. As Carmen Reinhart and Kenneth Rogoff noted in their 2009 book chronicling the financial crisis, “This Time Is Different,” asset bubbles primarily threaten the macroeconomy when driven by credit and when they hit bank balances. Since the dot-com crash was largely a repricing of equities (except for telecoms), and not a banking crisis, a systemic failure did not occur despite the massive losses incurred by investors.

This time, risks are accumulating through credit. As investor Paul Kedrosky notes, financing is shifting from equities to bonds, special-purpose instruments, leases, and private credit—all forms of borrowing ultimately linked to banks and insurance companies. If AI revenues and data centers decline, problems are likely to first surface in credit markets, not in stock prices. Look for unrealized coverage targets, tightening loan conditions, and refinancing pressures impacting lenders and insurance companies through long-term leases and inventory-backed loans of semiconductor chips.

This represents systemic danger. Unlike the dot-com era, current expansion and proliferation increase exposure at the core of financial engineering, making it likely that pressure will spread through lenders and synthetic instruments. Already, we can see growing concern among market watchers, with a warning from Moody’s that a significant portion of Oracle’s data center growth depends on OpenAI, which has yet to establish a path to profitability.

Of course, if AI generates widespread and sustainable productivity gains swiftly, the outlook will improve. Faster growth would relieve financial pressure, reduce debt ratios, and support these financing structures. However, if returns come late or fall short of expectations, the hefty upfront costs may not be compensated.