A handful of technology giants building artificial intelligence infrastructure will need to increase their own productivity by a factor of 2.7 to break even by 2030, accounting for capital costs, a 15% return, and asset depreciation, according to research published September 15 by Jessica Wachter, a finance professor at the University of Pennsylvania's Wharton School. If these companies can't hit such profit targets, Wachter warns they risk falling behind on interest payments and potential bankruptcy. Should the anticipated productivity boom fail to appear, she and her collaborator conclude, the current buildout would represent the largest misallocation of capital in history.

The so-called hyperscalers—Alphabet, Microsoft, Amazon, Meta, and Oracle—will spend roughly $750 billion this year constructing massive data centers, with projections suggesting total AI capital investments could exceed $5 trillion over the next four years, making this one of the largest capital investments by any industry in history. Meanwhile, total AI revenues will reach only $150 billion to $200 billion this year, according to Gary Gensler, who led the SEC during the Biden administration and now teaches at MIT's Sloan School. If hyperscalers keep pouring money into data centers through the next decade, required annual revenues will hit approximately $3.7 trillion by 2032, estimates Stijn Van Nieuwerburgh, a finance professor at Columbia Business School, assuming a 10% return that most investors would consider minimum acceptable. More than half of the $2.9 trillion that hyperscalers will spend between 2025 and 2028 will be financed with external capital, according to Morgan Stanley calculations. Alphabet reported a free cash deficit of around $5.9 billion in its latest quarter—its first shortfall since Google went public in 2004—as nearly $120 billion in revenues were consumed by AI infrastructure spending.

Wachter, who previously served as the SEC's chief economist and director of its division of economic and risk analysis, notes that for the growth required to justify spending through 2027—when she and her collaborator estimate expenditures will reach nearly $1.1 trillion—to happen by 2030 represents "a lot of growth compressed into a few years." The amount of growth needed would mirror what the US experienced during its IT boom over roughly 10 years starting in the mid-1990s. Daron Acemoglu, an MIT economist and 2024 Nobel laureate, says that for the investments to remain sustainable over the next five to 10 years, "we definitely need to see productivity gains," warning that without them people will eventually sour on AI, bringing down investments and limiting revenue growth. Most economists monitoring the data closely agree that economy-wide statistics currently show little or no productivity growth from AI, though a recent survey of 6,000 senior business executives across the US, UK, Germany, and Australia found that while roughly 90% report no productivity increase over the past three years, they anticipate a boost of around 1.45% total over the next three years.

The risks are spreading beyond the hyperscalers' own balance sheets as borrowed money flows through complex financial arrangements into the broader economy. Van Nieuwerburgh warns that financial institutions are now exposed to these data centers either as lenders, guarantors of debt, or backers of private credit funds, adding that "people don't even know they're holding this stuff" because it sits deep inside pension funds and backs life insurance policies. Meta's Hyperion data center in Richland, Louisiana illustrates the Byzantine financing: the company transferred an 80% stake to private-credit firm Blue Owl Capital, forming a joint venture that raises financing while Meta signs four-year leases and provides a residual value guarantee covering the facility's value if leases aren't renewed or are terminated. The lease length matches the expected lifetime of the data center's GPU chips, which represent roughly 60% of costs and double in performance roughly every two years, meaning owners of AI data centers coming online now will need to spend billions more on the next generation of chips by decade's end to stay competitive. Van Nieuwerburgh notes that the length of Meta's leases means if the company terminates early and pays off its loan, investors are left "with an empty building and no cash flow," adding they'd need to "find a new tenant for a huge data center, and good luck with that." The report notes that special purpose vehicles—financial structures that contributed to the great recession starting in late 2007—are back, raising concerns about how severe the fallout could be if investments from hyperscalers become entangled throughout the economy.

Gensler predicts a retrenchment is inevitable, though timing when the AI investment bubble will burst remains impossible, observing that "history tells us that at some point you get a retrenchment, and it's just a question of when and how severe." The current $750 billion annual spending rate could flatten or decrease as soon as next year, or hyperscalers might keep building until 2028 or 2029 before suddenly pulling back because they've accumulated sufficient capacity. While a crash might refocus investors on creating sustainable value with AI technology—and some Silicon Valley insiders are already hoping for one—the consequences could be severe: the dot-com bubble burst at the start of the 2000s cost hundreds of thousands their jobs, bankrupted companies large and small, and sent the US into a mild recession in 2001. This time presents a unique danger, as the enormous financial investments by hyperscalers have entangled the future of AI itself with the fortunes of massive data centers spreading nationwide, a logic founded on an unproven and risky belief that bigger means smarter. The financial bubble around colossal hyperscaler spending will likely burst eventually, possibly soon, and while that might prove financially painful, Wall Street and AI itself will survive—but the financial fate and future utility of the massive data centers fueled by trillions of dollars remain far less certain. Companies betting on frontier models housed in billion-dollar facilities face pressure not only from financial realities but from strategic choices about which version of intelligence is worth building at scale.