Wachter: AI data centers need 2.7× productivity by 2030

The debate over AI's economic payoff has largely been about how useful and widely deployed the models will become. Wachter's approach sidesteps that question: she starts from the one 'remarkable fact' not in dispute — a handful of hyperscalers are pouring vast sums into data centers — and works backward to ask how fast earnings must grow to justify the spending through 2027.

Her answer lands in the same territory as other estimates. Columbia's Stijn Van Nieuwerburgh calculates that roughly 183 gigawatts of planned AI compute capacity, at about $41 billion per gigawatt, would require roughly $3.7 trillion in annual revenues by 2032 assuming a 10% return. MIT's Gary Gensler frames the whole thing as 'a parlay bet by the capital markets and the economy' — one that requires hyperscalers to generate massive revenues, AI to boost broad economic growth, and expensive frontier models to fend off cheaper alternatives, all at once.

What sharpens the stakes this year is how the money is being raised. Morgan Stanley calculates that more than half of the $2.9 trillion hyperscalers will spend between 2025 and 2028 will be financed with 'external capital,' and complex arrangements like Meta's Beignet joint venture with Blue Owl Capital spread that risk through pension funds and insurance policies in ways the public cannot easily see. The outcome hinges on whether measurable productivity gains arrive before lenders' patience runs out — and on whether communities asked to host the data centers and their power plants feel they share in the upside.