A long Reuters analysis published on October 3 highlights that artificial intelligence is attracting capital on a scale never seen before. No previous technology has drawn in cash flows comparable to those now pouring into AI, exceeding even the sums invested during the railway boom or the internet bubble—even after adjusting for inflation.
Cumulative global spending on data centers alone could top $30 trillion by 2050, according to a PwC projection, nearly matching the value of outstanding U.S. Treasuries. PwC described the outlay as one that “dwarfs” what was spent in the railroad or dotcom eras.
Anthropic, one of the leading players in the AI race, plans to spend $518 billion in the coming years. The figure, drawn from an IPO prospectus seen by Reuters, exceeds 100 times the company’s 2025 revenue. Its backers argue that AI will prove more transformational than the advent of the steam engine and the industrialization it powered. Yet economists caution that these dizzying investment plans, massive outlays, and elevated valuations rest on assumptions of vast, broad-based productivity gains and future profits—assumptions for which there is so far little evidence or historical precedent.
Productivity Gains “Remain Elusive”
JP Morgan noted in August that broad-based productivity gains in the United States, the leader in the AI race, “remain elusive,” raising questions about the sustainability of current AI valuations. Historical precedent, the bank wrote, suggests that technology-driven booms often end when infrastructure buildouts cease to deliver sufficient returns.
A Bain & Company study published last month went further, arguing that productivity gains from existing markets will not be enough to justify the current level of spending. “Entirely new markets must emerge to close the funding gap,” the report stated, pointing to possibilities ranging from AI-guided robots to the development of new materials for batteries and semiconductors. Bain estimated that U.S. hyperscalers—companies such as Google, Amazon, and Microsoft that are rolling out infrastructure worldwide—along with other participants in the AI race, will need to generate more than $4.2 trillion in new revenue over the next five years simply to fund the buildout. “The question is whether the applications arrive in time to pay for it,” the study concluded.
U.S.-Dominated Investment and the Pressure for Returns
Some estimates suggest the United States accounts for roughly three-quarters of global AI investment. Columbia Business School economist Stijn Van Nieuwerburgh projects that U.S. AI investment alone could reach about $9 trillion between 2025 and 2032—equivalent to spending 3.2 percent of U.S. GDP each year. To earn a 10 percent return on that investment, he calculates, the U.S. AI sector would need to generate approximately $3.55 trillion in annual revenue by 2032. It currently earns only a fraction of that amount.
In a conference paper revised in October, Van Nieuwerburgh also warned that the leveraged structure of much of the debt financing AI infrastructure means “a relatively modest deterioration in demand, delays, or asset values can therefore produce much larger losses.”
Using Nvidia—whose chips form the backbone of the AI revolution—as an example, JP Morgan estimated that U.S. productivity gains would need to average 3 to 5 percent annually over the next decade to justify the company’s valuation. That would represent a substantial increase over the U.S. Congressional Budget Office’s baseline expectation of 1.75 percent annual productivity growth for the same period.
Optimistic Visions and “Recursive Self-Improvement”
The scale of the numbers has not dampened the enthusiasm of U.S. AI leaders. Anthropic’s Dario Amodei has described an AI future as potentially “a thing of transcendent beauty.” OpenAI’s Sam Altman has spoken of “the rate of new wonders being achieved” becoming immense as models learn to improve themselves and accelerate breakthroughs.
Jasjeet Sekhon, chief strategy officer at Google DeepMind, told a summit at UC Berkeley in August that this capacity for self-teaching—known as recursive self-improvement—is “a key part of the investment thesis.” If achieved, he said, it could deliver unprecedented productivity gains. The same capability, however, has also raised concerns about existential risks to humanity.
Anthropic’s economics team modeled several scenarios for the additional growth AI could deliver by 2030. Assuming a non-AI baseline of 2 percent annual growth, the team projected 2.4 percent under a modest-impact scenario, 5.4 percent under a substantial scenario, and 15.4 percent under an extreme scenario. Higher growth, the team noted, would also mean more jobs lost—though it assigned no probabilities to the outcomes.
Early Signals in the Labor Market
Amodei forecast last year that AI could eliminate half of all entry-level white-collar jobs within five years. So far, researchers say the most visible effect appears limited to making it harder for those seeking office work to find employment.
Studies in both the United States and Britain have pointed to a slowdown in early-career hiring for white-collar positions that perform tasks at which AI is already adept, even as overall employment remains strong.
Researchers at Stanford University reported in August that employment of workers aged 22 to 25 in AI-exposed occupations—such as accountants and paralegals—was 19 percent lower than in jobs that AI finds hard to replicate, such as janitors and construction workers.
Historical Perspective: The Value of Remaining Infrastructure
Few doubt AI’s potential to transform everything from office work to research laboratories, just as earlier revolutions compressed journey times from days to hours or connected the world at the touch of a keyboard. What remains more immutable are the mathematics of return on investment and the deadlines for repaying loans. Economists must therefore assess not only the ups and downs of the investment cycle but also the longer-term implications for the global economy.
Diane Coyle, an economist at the University of Cambridge, observed that the productivity impact of past revolutionary technologies typically took between 10 and 50 years to fully feed through. Even if the transformation AI ultimately delivers arrives more slowly than current valuations and investment figures imply, real economic benefits should still remain—just as trains continued to run after the Panic of 1873 bankrupted many railroad barons, and the internet did not shut down after the late-1990s dotcom bubble burst.
“History is our friend in trying to understand this,” Coyle said. “As long as one is left with the infrastructure that’s needed to support all the productivity effects down the road, that’s okay.”
The AI race has entered a high-stakes, high-speed phase.
Capital is flowing into data centers, chips, and model training at an unprecedented pace, yet a clear gap remains between the timelines for generating returns and the timelines for realizing broad productivity gains. The fundamental question is not whether AI has the potential to transform the world, but whether that transformation can arrive in time—before the money runs out—to justify the extraordinary investments and valuations now in place. The outcome of this race will shape not only the fortunes of technology giants and AI startups, but also the longer-term structure of the global economy and the future of work.