AI Infrastructure's Trillion-Dollar Bet and the Bubble Question

MIT Technology Review published a long read by David Rotman asking the "trillion-dollar question": with a handful of hyperscalers pouring roughly $750 billion into AI data centers this year and possibly more than $1 trillion next, can they earn enough to sustain the boom? The estimates and judgments below are from that article.

How steep the payoff curve must be#

MIT TR reports that Jessica Wachter, a Wharton finance professor and former SEC chief economist, calculated that—accounting for cost of capital, a 15% return, and depreciation—these companies must raise their own productivity by a factor of 2.7 to break even by 2030, compressing into a few years the growth the US IT boom delivered over roughly a decade from the mid-1990s. She and her coauthor write that if the productivity boom "fails to materialize," "the current buildout will be the largest misallocation of capital in history," with firms falling "behind on their interest payments" and risking bankruptcy.

The spend-versus-revenue gap#

MIT TR quotes former SEC chair and MIT Sloan professor Gary Gensler: the hyperscalers (Alphabet, Microsoft, Amazon, Meta, and Oracle, which partners with OpenAI) plan to spend trillions, while total AI revenue this year is roughly $150-200 billion, so "the spending does not have commensurate revenues yet." Alphabet is cited as an example: nearly $120 billion of revenue in its latest quarter was devoured by AI infrastructure, leaving a free-cash-flow deficit of about $5.9 billion—its first shortfall since going public in 2004. Morgan Stanley estimates that more than half of the $2.9 trillion hyperscalers will spend on data centers from 2025 to 2028 will be financed with external capital.

After the bubble#

MIT TR frames the bet as a "parlay": the firms must generate massive revenues, AI must lift economy-wide productivity, and both must hold up while frontier models fend off cheaper rivals. Columbia's Stijn Van Nieuwerburgh calculates ~$3.7 trillion in required annual revenue by 2032; 2024 Nobel laureate Daron Acemoglu stresses that productivity gains are essential; a survey of some 6,000 executives found about 90% reporting no productivity gain over the past three years. Separately, a sponsor-labeled piece published by MIT TR and authored by ON.energy—which the site marks as "not written by MIT Technology Review's editorial staff"—cites a July 22, 2026 fault in Ashburn, Virginia that knocked more than 3 gigawatts off the grid in a second, arguing the grid strain is an "architecture problem," not a shortage of power.

Our take: Read together, the pieces suggest the boom's real weakness is not compute or chips but whether returns arrive on schedule and where the risk ends up—binding hyperscaler balance sheets to public electricity prices and pension funds is the alarm MIT TR keeps ringing.