When Jessica Wachter, a finance professor at the University of Pennsylvania’s Wharton School, set out to measure AI’s near-term economic impact, she confronted many business and technical uncertainties. She began instead from one indisputable fact: a few so-called hyperscalers are investing enormous sums to build AI data centers.
Avoiding speculation about how useful or widely deployed future AI models will be, Wachter and her coauthor asked how fast hyperscalers’ earnings must grow to justify their spending through 2027 — a point by which they estimate expenditures will reach nearly $1.1 trillion. This is a straightforward accounting method to make sense of today’s historic buildout.
How much growth is required?
Wachter’s calculations indicate hyperscalers would need to raise their productivity by about 2.7 times to break even by 2030 after accounting for the cost of capital, a 15% return requirement, and asset depreciation. That level of increase is not impossible, she says, but it would compress into a few years the kind of economic expansion seen during the U.S. IT boom that began in the mid-1990s. If that productivity surge does not happen, Wachter and her coauthor conclude the current buildout could become “the largest misallocation of capital in history.”
The gap between investments and revenues
Hyperscalers are slated to spend roughly $750 billion this year on data centers, and the spending spree appears to continue. Some projections put total AI capital investment from Alphabet, Microsoft, Amazon, Meta and Oracle (which partners with OpenAI) at more than $5 trillion over the next four years.
Yet total AI revenues this year are only about $150–$200 billion, according to Gary Gensler (MIT Sloan School), meaning the scale of spending is not commensurate with current revenues.
Debt, cash flow and financing risks
Much of the buildout is debt-financed and free cash flow (operating cash flow minus capital expenditures) for the group is expected to turn negative. Even Alphabet reported that nearly $120 billion in revenue was offset by AI infrastructure spending, leaving a free cash shortfall of about $5.9 billion — its first deficit since Google’s IPO in 2004.
Morgan Stanley estimates that more than half of the $2.9 trillion hyperscalers plan to spend between 2025 and 2028 will be financed with “external capital.” That borrowing has produced complex financing webs that spread exposure across banks, private-credit funds, pension funds and life insurers.
Chip depreciation and technological obsolescence
High-performance GPU chips make up roughly 60% of data-center costs, and their performance is roughly doubling every two years. That pace explains rapid model improvements but forces operators who bring centers online now to plan for billions more in expenditures on next-generation chips by the end of the decade if they wish to stay competitive. Without such reinvestment, Princeton’s Mihir Kshirsagar warns, these facilities risk becoming stranded “hulks.”
What revenues would be necessary long term?
Stijn Van Nieuwerburgh of Columbia Business School models a scenario in which about 183 gigawatts of compute capacity are built between 2025 and 2032; at roughly $41 billion per gigawatt, and assuming a 10% investor return, required annual revenues would be about $3.7 trillion by 2032. Other estimates arrive at similar magnitudes.
The third leg: economy-wide productivity growth
Even if hyperscalers can sell subscriptions and capacity to companies eager to adopt AI in the short term, long-term sustainability requires that AI deliver broad productivity gains. Customers will need to see bottom-line improvements — higher output, greater efficiency, or new profitable products — or they will stop paying premium prices.
Most macroeconomic statistics to date show little or no productivity growth attributable to AI, though there are tentative signals. A survey of roughly 6,000 senior executives in the U.S., U.K., Germany and Australia found about 90% reported no productivity increase over the past three years, yet they expect roughly 1.45% aggregate productivity growth over the next three years (U.S. executives expect 2.25%). The same and related surveys anticipate roughly $280 billion in private-sector AI spending by the end of 2026.
If AI boosts productivity by reducing employment, public backlash could intensify over the technology and infrastructure such as data centers. Public acceptance and local benefits are therefore an additional, implicit requirement for the hyperscalers’ bet to pay off.
Complex financing and local tensions: the Meta Hyperion example
Meta’s Hyperion data center in Richland Parish, Louisiana, illustrates financing and local-risk complexity. Announced in late 2024 as a two-gigawatt, roughly $10 billion project, costs later rose toward $30 billion and Meta transferred an 80% stake to private-credit firm Blue Owl Capital, forming a joint venture called Beignet. Meta signed a series of four-year leases with the joint venture and provided a residual value guarantee to backstop the arrangement. In July, Meta announced an expansion to five gigawatts, with a total price tag of about $50 billion so far; Entergy has proposed building seven additional gas-fired power plants to reach around 7.5 gigawatts of generation capacity.
Local officials initially welcomed the project, but residents and consumer advocates worry who will pay if Meta reduces its power purchases or terminates leases. Entergy says it has a 20-year guaranteed power purchase from Meta, yet observers — including the Union of Concerned Scientists and local consumer groups — remain skeptical that long-term obligations will hold or that ratepayers won’t shoulder unexpected costs.
Potential outcomes and historical parallels
Predicting when a possible investment bubble might burst is uncertain, but many experts expect some retrenchment. Gary Gensler suggests the current $750 billion annual spending could flatten or decline in coming years, or that companies may suddenly stop expanding once they perceive sufficient capacity.
Past busts — the dot-com crash and the 2007–2008 financial crisis — caused deep pain but also refocused capital on more sustainable technologies: the telecom and fiber-optic investments of the earlier era now underpin today’s internet, and many hyperscalers emerged or strengthened after those corrections.
This episode differs in one key respect: hyperscalers’ financial engineering has tied the fortunes of AI itself to dispersed and expensive physical infrastructure and to complex debt obligations. If demand or productivity gains lag, the fallout could be not just corporate but systemic.
Conclusion: three bets that must all win
The hyperscalers’ strategy amounts to a parlay bet that requires three simultaneous wins: massive revenues for hyperscalers, broad productivity gains across the economy, and continued dominance of expensive frontier models over cheaper, “good-enough” alternatives. Any failure among these conditions would increase the likelihood of large financial losses, local community harm and stranded assets — even if AI as a technology continues to advance in other forms.
