Product1 publisher3 min readPublished
A Wharton economist prices the AI buildout's break-even at 2.7x productivity by 2030
Jessica Wachter asked how fast hyperscaler earnings must grow to justify nearly $1.1 trillion of spending through 2027. Her answer is a 2.7x productivity gain by 2030. Anyone forecasting an AI budget now has a number to test.
The Product Desk · Product desk
What happened
- Wharton finance professor Jessica Wachter and a coauthor put hyperscaler spending on AI data centers at nearly $1.1 trillion through 2027, and used that figure as the starting point of their analysis.
- Their calculation finds the AI companies must raise their own productivity by a factor of 2.7 to break even by 2030, after the cost of capital, a 15% return and depreciation of the assets.
- Alphabet's latest quarter produced nearly $120 billion of revenue and a free cash deficit of some $5.9 billion, the company's first shortfall since Google went public in 2004.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- decision A 2027 AI budget can now be argued against a stated threshold with stated assumptions, so the disagreement is about whether 28% a year is reachable, not about the word bubble.
- constraint Four years is the constraint: the same 2.7x over the ten years of the IT boom Wachter cites asks for roughly 10% a year, a rate the 1990s actually delivered.
- exposure Buyers sitting on debt-funded capacity meet the risk at renewal, because a lender is owed the same money whether or not compute demand holds, and the borrower can raise price.
- cost The bill for a miss lands well outside the hyperscalers, with the investments heading toward around 3% of GDP and the paper's own label for that outcome being the largest capital misallocation in history.
Buyers have to sign before the evidence arrives. Someone signs a 2027 AI contract in the next quarter, and the question of whether the capacity behind it pays for itself stays open until 2030.
Wachter and her coauthor worked backwards from the spending. She skipped forecasting how good the models get or how widely they get deployed, and asked what earnings growth the spending through 2027 already implies, priced with the cost of capital, a 15% return and depreciation of the assets [3][4]. Before Wharton, she was the SEC's chief economist and ran its division of economic and risk analysis [1][5].
Four years is the hard part. From a 2026 base, 2.7x by 2030 works out to about 28% a year, since 2.7 to the power of one quarter is 1.28 [1]. Wachter compares the required growth to the US IT boom, which ran about ten years from the mid-1990s [6]. Spread over ten years, the same multiple is about 10% a year [2]. The article does not give the paper's base year for the multiple. "That's a lot of growth compressed into a few years," she said [7].
Hyperscalers will spend about $750 billion this year [10], against total AI revenues Gensler puts at $150 billion to $200 billion [11]. That is between $3.75 and $5 of capital spending for every dollar of AI revenue [3]. "The challenge is that the spending does not have commensurate revenues yet. That's a fact," Gensler said [12]. Some projections put total AI capital investment from Alphabet, Microsoft, Amazon, Meta and Oracle, which partners with OpenAI, above $5 trillion over the next four years [13].
For a buyer, 2.7 is a claim about output per dollar at the vendor's end, and it gets funded at yours. The thing to measure is change in output per person, landing on a cost line with a named owner. Seat counts and weekly actives on an assistant will not tell you that.
The financing pulls this into procurement. Free cash flow for the group is expected to dip into negative territory [15]. Alphabet's latest quarter turned nearly $120 billion of revenue into a free cash deficit of some $5.9 billion, its first since Google went public in 2004 [16], equal to about 5% of the quarter's revenue [4]. If demand for the data centers' compute drops, the companies still have to repay what they borrowed [17].
Two tests apply to a 2027 AI line item. One is whether the saving lands in a cost line somebody owns or only in a review deck. The other is whether the line item survives if the vendor raises price to service its debt. Fail the first and the item never produced the productivity Wachter's model needs. Fail the second and the item is priced on the vendor's balance sheet.
Wachter says the growth is not impossible [19]. On what happens if the hyperscalers miss those profit goals, she said: "Then they will fall behind on their interest payments, and that risks bankruptcy" [8]. If a productivity boom "fails to materialize," she and her coauthor conclude, "the current buildout will be the largest misallocation of capital in history" [9].
What to watch
- Whether group free cash flow actually turns negative next quarter, or capex guidance gets trimmed first.
- Whether AI revenue moves out of the $150bn to $200bn band Gensler cites for this year.
- Whether Wachter and her coauthor publish the base year and depreciation schedule behind the 2.7x figure.