Can AI Productivity Grow Fast Enough to Justify Big Tech’s Spending?

The trillion dollars that big technology firms have committed to AI infrastructure amounts to a bet that AI productivity will roughly triple within a few years. If that bet fails, the firms making it risk bankruptcy, according to new Wharton research.

The findings come from a new paper titled “What Investment Data Implies About the AI Transition.” Wharton finance professor Jessica A. Wachter co-authored the paper with Jonathan Wachter, head of operations for macro, treasury, and risk technology at Point72, a Stamford, Conn.-based alternative investments firm.

The current AI revolution has grown faster than its supporting infrastructure in terms of data centers and power capacity, triggering a rash of new investments. “There’s been a productivity boom; people didn’t foresee it, and now everybody’s playing catch-up,” said Jessica Wachter, setting the backdrop for the massive size of the AI investments. “The initial [AI sector] boom entails an unanticipated jump in productivity, to which optimizing firms respond with a surge in investment,” the paper noted.

Five publicly held big tech firms — Amazon, Alphabet, Microsoft, Meta, and Oracle — also known as hyperscalers, make up the majority of AI infrastructure investments, which have grown from $155 billion in 2022 to a forecast $755 billion in 2026, and estimated to cross $1 trillion in 2027. Five other firms — SpaceX subsidiary xAI, CoreWeave, Crusoe, IREN, and Lambda — have forecast AI infrastructure capital expenditure totaling $95 billion in 2026.

Evaluating AI Infrastructure Investments

As one would expect, concerns are mounting over the viability of those massive AI investments. The paper advances the debate by using investment data with a theoretical model for rare productivity booms to estimate future growth. “People have looked at the abstract question of how AI would qualitatively affect the economy, but nobody is really doing it with numbers the way we are,” Wachter said. “The idea of a rare boom is also a useful device for thinking about this.”

The model has a two-year window in which each year carries a 50% chance of a further boom, with any additional booms realized in 2029 and 2030. This generates three scenarios: moderate (the initial boom only), transformative (one further boom), and singularity (two further booms).

The model in the paper calibrated the AI investment commitments through end-2027 to rare booms. It estimated that each of those booms increases the AI sector’s productivity by a multiple of 2.7 times current productivity levels, while the non-AI sector continues to grow at its historical rate. The AI sector’s initial boom implies about 5 percentage points of additional cumulative GDP growth by 2030; with further booms, the scenarios range up to 58 percentage points.

“There’s been a productivity boom; people didn’t foresee it, and now everybody’s playing catch up.”— Jessica A. Wachter

An Eye-Popping Multiple

Wachter agreed that the 2.7 multiple is “an eye-popping number,” simply because it would eclipse the multipliers in all prior boom periods across history.

For instance, the U.S. IT boom from 1995 to 2005 delivered only 1.5 times per capita GDP growth over 10 years. But that comparison would be inappropriate, Wachter said, noting that the internet boom was a stock price boom, and not a capital expenditure boom like what is now occurring in the AI sector.

The three industrial revolutions from 1760 to 1920 logged per capita GDP growth of between 1.7 and 2.8 times; that period includes the U.S. railroad era (1850-1910), which delivered a growth multiple of 2.8 times over 60 years. The East Asian growth miracles of Japan, South Korea, Taiwan, Singapore, and China produced multiples of 8 to 13 times, each over 25 to 30 years. A close comparison to the AI infrastructure spending boom is the fiber optic cable buildout of the late 1990s, which Wachter estimates implied a productivity gain of roughly 1.3 to 1.5 times.

The AI sector’s share of the economy rises from roughly 3% today to between 8% and 39% depending on the scenario, the paper stated. As this share grows, the AI sector’s rapid productivity gains increasingly dominate aggregate GDP growth. The model also estimated productivity multipliers in investment scenarios up to 2110: The expected AI-sector productivity multiplier under the singularity scenario is 7.1 over 30 years, 26.4 over 50 years, and 188 over 80 years, through 2110.

Is the Growth Multiple Achievable?

The big question, of course, is if that productivity multiple of 2.7 is a reasonable expectation. Wachter is optimistic. “We looked at today’s stock market valuations and asked what needs to be true for those valuations to not be too high. It turns out to be actually quite a reasonable number. We observe productivity from these companies’ earnings.”

“Many people see these investments unconnected from reality. But I actually don’t think that it’s unusual for a large sector in the economy to experience that kind of productivity boom,” Wachter said.

“The nature of the American economy is to jump on an opportunity and risk bankruptcy.”— Jessica A. Wachter

The paper anchors its case on the revealed preferences of firms in AI infrastructure investments. “Those are spending commitments where companies say, ‘We’re going to put dollars in the ground, and we’re going to put them as fast as we can,’” Wachter pointed out. Those commitments are stronger than the broad corporate utterances of intended investments, which may or may not materialize.

Here, the paper noted that while “the revealed-preference argument identifies the productivity boom that managers believe has occurred, it does not establish that the boom has in fact occurred.” Drawing from that, the investments “may simply reflect a bubble,” the paper noted. “Managers are not immune to collective overoptimism, and the history of technology investment is replete with episodes,” it added, pointing to the overcapacity built during the fiber-optic build-out in the late 1990s as an example. “It is possible, for example, that the productivity-enhancing power of AI may be a mirage. In this case, customers would be unwilling to pay for more AI usage.”

The macroeconomic effects of AI investments are also debatable. Some studies have suggested that as the AI economy triggers higher growth, it should lead to higher interest rates, Wachter noted. “So far, we haven’t really seen that. I think that the explanation for that is that it is very risky growth, and that’s what is keeping the interest rate low.” Another point the paper noted is that the AI boom substantially increases the equity premium, or the extra return investors expect for holding stocks instead of safe, risk-free assets such as government bonds. “Anything that’s risky increases the equity premium,” Wachter said.

Is there some irrational exuberance here? “The nature of the American economy is to jump on an opportunity and risk bankruptcy,” Wachter said. “By making these massive expenditures, these firms increase the risk of bankruptcy. But it’s because they don’t want to leave money on the table.”

What Could Upset the Math?

According to Wachter, the case for a bubble here is not compelling the way it was in the late 1990s, with the dot-com boom. “The real danger is that something outside of this could go wrong, that could disrupt this,” she said. “There could be some geopolitical event that would freak investors out and make it hard to continue to raise money, as these companies depend on very complex global supply chains.”

Beyond all the math is the substantial long-run uncertainty that bedevils a range of scenarios of what productivity the investments could bring. The paper pointed to a crucial aspect of the AI infrastructure investment ramp-up. “At its base is a productivity boom that is in investment data, but not yet in productivity data,” it noted. “If the boom fails to materialize, the current build-out will be the largest misallocation of capital in history.” On the other hand, those who are skeptical of the investments may eventually realize in the end that they missed out on a big opportunity. “Rare events are hard to imagine until they occur,” the paper stated.

Source: View the original article at Knowledge@Wharton.