Glonce

AI hyperscalers need to raise productivity 2.7x to break even by 2030

businessThe Glonce editorsMIT Technology Review🔥71Glonce rating71Impact68Novelty50Relevance95Surprise58Players88Credibility82
AI hyperscalers need to raise productivity 2.7x to break even by 2030

MIT Technology Review examines whether the AI industry's infrastructure spending can pay for itself. Rather than debate how useful AI models will turn out to be, Jessica Wachter, a finance professor at the University of Pennsylvania's Wharton School and previously the SEC's chief economist and director of its division of economic and risk analysis, started from what she calls a "remarkable fact": hyperscalers are already committed to spending huge sums on data centers. She and an unnamed coauthor calculate that hyperscalers need to raise their own productivity by a factor of 2.7 to break even by 2030, once the cost of capital, a 15% required return and asset depreciation are factored in. Wachter calls that "not impossible": the growth required would resemble the productivity surge the US saw over about ten years during the IT boom that began in the mid-1990s, only much faster this time, which she notes is "a lot of growth compressed into a few years." If it does not happen, hyperscalers, in her words, "will fall behind on their interest payments, and that risks bankruptcy." She and her coauthor write in their research paper that, absent that productivity boom, the buildout will be "the largest misallocation of capital in history."

The numbers behind that fact are stark. Hyperscalers will spend about $750 billion this year. Next year's data-center spending is projected to exceed $1 trillion, and Wachter's own estimate for hyperscaler spending through 2027 comes in at nearly $1.1 trillion. Some projections put total AI capital investment from the five hyperscalers, Alphabet, Microsoft, Amazon, Meta and Oracle (which partners with OpenAI), at more than $5 trillion over the next four years. Against that, Gary Gensler, who ran the SEC during the Biden administration and now teaches at MIT's Sloan School, puts total AI revenue this year at just $150 billion to $200 billion: "The challenge is that the spending does not have commensurate revenues yet. That's a fact," he says, adding that the open question is whether the spending is an investment that eventually pays off. The investments could soon reach about 3% of US GDP.

The strain is already visible on hyperscaler balance sheets. Alphabet, historically a cash-hoarding company, reported nearly $120 billion in revenue last quarter but saw free cash flow swing to a deficit of about $5.9 billion once AI infrastructure spending was subtracted, its first shortfall since Google's 2004 IPO. Free cash flow for the hyperscaler group as a whole is expected to turn negative soon. That is pushing companies toward debt: Morgan Stanley calculates that more than half of the $2.9 trillion hyperscalers will spend on data centers between 2025 and 2028 will be financed with external capital rather than their own cash. Stijn Van Nieuwerburgh, a finance professor at Columbia Business School, says that debt is spreading the risk well beyond the tech sector, through lenders, debt guarantors and the private credit funds that back much of the borrowing, and ultimately into instruments like pension funds and life insurance policies that ordinary people hold without realizing it.

There is also a physical clock running: GPU chips account for roughly 60% of data-center costs and their performance is doubling roughly every two years, so owners of the data centers coming online this year and next will have to keep spending billions more on next-generation chips through the end of the decade just to stay competitive. Skip that reinvestment, warns Mihir Kshirsagar at Princeton's Center for Information Technology Policy, and data centers risk becoming "hulks," stranded assets "scattered all over the place."

Separately, Van Nieuwerburgh models a scenario in which about 183 gigawatts of AI compute capacity are built between 2025 and 2032 at roughly $41 billion per gigawatt; assuming a 10% return, the minimum he says most investors would accept, hyperscalers would need "required" annual revenue of roughly $3.7 trillion by 2032, a figure the piece says other analysts land near independently. Gensler frames the whole situation as "a parlay bet by the capital markets and the economy": hyperscalers must generate that scale of revenue, AI must lift economy-wide productivity growth, and expensive frontier models must keep fending off cheaper alternatives that many businesses might find good enough, three wagers that all have to land together. The piece adds a fourth, more speculative one: that the public and local communities feel they are benefiting too.

On the productivity side, the picture is still thin. A survey of about 6,000 senior executives across the US, UK, Germany and Australia found that roughly 90% report no productivity increase from AI over the past three years, though they expect a combined 1.45% boost over the next three (2.25% among US respondents). A follow-up survey found those same executives planning to spend more on AI, which the survey's authors used to project about $280 billion in private-sector AI spending by the end of 2026, but the productivity gains they describe come mainly from cutting jobs while growing sales. Daron Acemoglu, an MIT economist and 2024 Nobel laureate, warns that if productivity gains do not show up, "people are going to sour on AI, and that will bring down investments and it would also limit revenue growth"; over the next five to ten years, he says, the investments need real productivity gains to be sustainable. The piece flags two ways the bet could fail from opposite directions: if the productivity gains that do appear come from businesses running cheaper models, such as DeepSeek, hyperscaler revenue could collapse instead of growing into the spending; if the gains come mainly from job cuts, the resulting public backlash could block planned investments and choke off the revenue growth those investments are meant to produce.

As an illustration of how tangled the financing has become, the piece points to Meta's Hyperion data center in Richland, Louisiana: two gigawatts of compute capacity at a price tag of about $10 billion when announced in late 2024, at the time Meta's largest planned data center, welcomed by state and local politicians as a boost to the rural community and reliant on Entergy Louisiana, the state's largest utility. The version of the article captured for this summary breaks off mid-sentence at that point, so whatever the piece concludes about that project, or overall, beyond here is not reflected in this retelling.

Key facts

  • Hyperscalers are set to spend more than $1 trillion on data centers next year, up from about $750 billion this year; Wachter separately estimates their spending through 2027 at nearly $1.1 trillion.
  • Wharton's Jessica Wachter calculates hyperscalers need to raise their own productivity by a factor of 2.7 to break even by 2030, once the cost of capital, a 15% return and depreciation are factored in.
  • Total AI revenue this year runs just $150 billion to $200 billion, per Gary Gensler (who ran the SEC under the Biden administration, now at MIT Sloan), against AI infrastructure investment that could soon reach about 3% of US GDP.
  • Alphabet's free cash flow swung to a $5.9 billion deficit last quarter despite nearly $120 billion in revenue, its first shortfall since Google's 2004 IPO.
  • Morgan Stanley estimates more than half of the $2.9 trillion hyperscalers will spend on data centers from 2025 to 2028 will come from external capital (debt), a risk Columbia's Stijn Van Nieuwerburgh says is spreading invisibly into pension funds and life insurance.

Why it matters

This is not one company's capex decision: it is a test of whether the entire AI industry's capital allocation adds up. Hyperscalers are set to spend more than $1 trillion on data centers next year on top of about $750 billion this year, and some projections put total AI capital investment above $5 trillion over the next four years. That scale now represents a measurable share of the US economy: the investments could soon reach about 3% of GDP. Wachter's framing turns a subjective argument about whether AI is useful into an arithmetic one: given the spending already committed, how much productivity growth is required to pay for it, regardless of anyone's opinion of the technology itself.

Who it affects

Directly: the five hyperscalers named in the piece, Alphabet, Microsoft, Amazon, Meta and Oracle (which partners with OpenAI), plus their investors and creditors. Indirectly: the wider economy, since a large share of the spending is financed with debt that runs through banks, guarantors and private credit funds and ends up, per Van Nieuwerburgh, inside ordinary pension funds and life insurance policies. Businesses buying AI subscriptions and tokens are on the hook too, eventually having to show productivity gains to justify their own spending. Workers are affected because the executives surveyed expect to lift productivity mainly by cutting headcount while growing sales, the mechanism Acemoglu warns could trigger public backlash. The communities that host the data centers themselves, such as the area around Meta's Hyperion project in Richland, Louisiana, and the utilities that supply them, are affected as well.

How to use it

There is no product here to buy, so treat this as a scorecard for judging the AI buildout going forward. The benchmarks the piece points to: hyperscaler free cash flow (already negative at Alphabet last quarter, its first shortfall since the 2004 IPO); the share of new data-center spending funded with debt rather than cash (Morgan Stanley puts more than half of the 2025-2028 total at external capital); and economy-wide productivity growth, still close to zero after three years per the cited survey. If those numbers start moving the way Wachter's math requires, a 2.7x productivity gain by 2030, the bet is on track. If spending keeps compounding while they stay flat, so does the risk the piece describes.

How solid is it

The piece leans on named, checkable sources: Wachter, a Wharton finance professor and previously the SEC's chief economist, presenting her own paper's break-even math, though the piece does not name her coauthor or say where the paper is published; Gensler, who ran the SEC under the Biden administration and now teaches at MIT's Sloan School; Van Nieuwerburgh at Columbia Business School, whose independently derived $3.7 trillion required-revenue figure for 2032 the piece says other analysts reach too; Acemoglu, an MIT economist and 2024 Nobel laureate; and Kshirsagar at Princeton. Alphabet's revenue and free-cash-flow numbers come from its reported quarter rather than a projection, and the debt-financing estimate is credited to Morgan Stanley. The productivity figures rest on a named survey of about 6,000 executives across four countries. Two gaps stand out: the piece does not say whether Gensler's $150 billion to $200 billion total AI revenue figure covers just the five named hyperscalers or the wider AI industry, and it includes no response from Alphabet, Microsoft, Amazon, Meta or Oracle to the spending and productivity critiques laid out here.

Risks and caveats

The central risk is the one Wachter states outright: if the productivity boom needed to hit 2.7x by 2030 does not materialize, hyperscalers "will fall behind on their interest payments, and that risks bankruptcy," and she and her coauthor write that the buildout would then be "the largest misallocation of capital in history." A second, physical risk sits in the hardware itself: GPUs run about 60% of data-center costs and their performance is roughly doubling every two years, so facilities that are not continually upgraded risk becoming stranded "hulks." A third is financial contagion: Morgan Stanley estimates more than half of the $2.9 trillion hyperscalers will spend from 2025 to 2028 will be debt-financed, and Van Nieuwerburgh warns that risk is already spreading, largely invisibly, through pension funds and life-insurance backing. A fourth is competitive: if the productivity gains that do appear come from businesses running cheaper models (the piece names DeepSeek as an example), hyperscaler revenue could collapse instead of growing into the spending. A fifth is social: the productivity gains executives expect are tied to cutting jobs, which Acemoglu warns could produce a backlash severe enough to choke off the investments. One caveat on this retelling: the captured source text breaks off mid-sentence in its closing section on Meta's Hyperion data center, so any conclusion the original piece reaches beyond that point is not reflected here.

“People don't even know they're holding this stuff. It's somewhere deep inside their pension fund. Ultimately, it's backing their life insurance policies. And that risk is getting distributed everywhere in places that are invisible.”

— Stijn Van Nieuwerburgh, finance professor at Columbia Business School