AI's trillion-dollar bet: what has to go right for the data center boom not to end in disaster

The article, by David Rotman in MIT Technology Review, opens with an accounting exercise by Jessica Wachter, a finance professor at Wharton (University of Pennsylvania) and former chief economist at the SEC. Rather than trying to predict how useful AI will be, Wachter starts from an undeniable fact: a handful of hyperscalers are pouring colossal sums into building AI data centers, with cumulative spending that, by her estimates with a co-author, will approach $1.1 trillion by 2027. Her conclusion is blunt: to cover the cost of capital, earn a 15% return and offset asset depreciation, those companies would have to multiply their productivity by 2.7 before 2030. That is not impossible —it would be equivalent to the growth generated by the US tech boom from the mid-1990s— but compressed into very few years. If they fail, Wachter warns, they will start missing interest payments, with a risk of bankruptcy; and if the productivity surge 'fails to materialize,' the current infrastructure boom could become 'the greatest misallocation of capital in history.'

The scale of the spending is hard to overstate. This year hyperscalers will spend around $750 billion on data centers, and some projections put the combined investment of Alphabet, Microsoft, Amazon, Meta and Oracle (an OpenAI partner) above $5 trillion over the next four years, one of the largest capital investments by any industry in history. The problem, notes Gary Gensler —who ran the SEC during the Biden administration and is now a professor at MIT's Sloan School— is that total revenue generated by AI is barely $150 billion to $200 billion this year. 'The spending doesn't have equivalent revenue yet. That's a fact,' Gensler sums up. The question is whether that investment will be recouped in the future. What is at stake, according to the article, is the financial health of Big Tech and of the US economy as a whole, since this spending could approach 3% of GDP.

On top of the usual risks of any investment boom, this year adds a new ingredient: debt. Free cash flow across the hyperscaler group (operating cash minus capital spending) is about to turn negative. Even Alphabet, famous for piling up cash, posted revenue of nearly $120 billion last quarter that was devoured by AI infrastructure investment, leaving a free cash flow deficit of some $5.9 billion, its first since Google went public in 2004. Added to this is the problem of chip depreciation: GPUs account for close to 60% of a data center's cost and their performance roughly doubles every two years, forcing billions to be reinvested in new chip generations to stay competitive. Without that constant renewal, warns Mihir Kshirsagar of Princeton's Center for Information Technology Policy, data centers risk becoming 'shells' stranded across the country.

Gensler sums up the situation as a 'parlay bet' by capital markets and the economy, one that requires winning three related but independent bets at once: that hyperscalers generate massive revenue, that AI drives broad economic growth, and that expensive frontier models remain preferable to cheaper alternatives that many companies might consider good enough. Each bet depends on the other two. If the pace of spending holds into the next decade, the revenue required soars to astronomical figures: Stijn Van Nieuwerburgh, a finance professor at Columbia Business School, calculates that for the roughly 183 gigawatts of AI computing capacity projected between 2025 and 2032 (at a cost of about $41 billion per gigawatt), and assuming a minimum 10% return acceptable to investors, the annual revenue 'required' would be around $3.7 trillion in 2032. Other analysts arrive at similar figures.

Winning the second bet —productivity growth across the whole economy— is crucial. For a few years hyperscalers could grow revenue simply by selling subscriptions and tokens to companies eager to ride the AI wave, but sooner or later those customers will need to justify the spending with tangible benefits. Daron Acemoglu, an MIT economist and 2024 Nobel laureate, warns that if those productivity gains do not arrive, people will eventually grow tired of AI, which would slow investment and cap revenue growth; for the investments to be sustainable over the next five or ten years, he argues, 'we need to see productivity gains.' For now, most economists agree that aggregate economic statistics show little or no productivity improvement attributable to AI. There are, however, signs of optimism: a recent survey of some 6,000 senior corporate executives in the United States, the United Kingdom, Germany and Australia found that while around 90% saw no productivity increase over the past three years, they expect a 1.45% pickup over the next three (US executives anticipate 2.25%). In a follow-up survey, the same executives said they planned to spend more on AI, leading the study's authors to project some $280 billion in private AI spending by the end of 2026.

That good news for the hyperscalers comes with bad news for jobs: the executives surveyed expect to raise their companies' productivity by increasing sales while significantly cutting headcount. If AI's productivity is achieved by destroying jobs, the public backlash against the technology —already visible around data center construction— will probably intensify. The article suggests adding an implicit fourth bet to Gensler's list: that the public and local communities feel they are benefiting from these massive investments too. And it underscores how interdependent these bets are: if productivity ended up coming from cheaper models like DeepSeek, hyperscaler revenue could collapse; and if productivity comes at the cost of jobs, the social backlash could block many of the planned investments and choke off expected revenue.

The report puts particular emphasis on how the nature of the risk has changed: while AI companies were spending accumulated cash, the risk was largely confined to their own balance sheets and shareholders; now that much of the money is borrowed, the game has far broader consequences. By Morgan Stanley's calculations, more than half of the $2.9 trillion hyperscalers will spend between 2025 and 2028 building AI data centers will be financed with 'outside capital.' That borrowing is creating ever more complex financial structures that entangle these investments with the rest of the economy. Van Nieuwerburgh warns that many financial institutions are exposed, directly or indirectly, to these data centers as lenders, guarantors of part of the debt or backers of the private credit funds that finance them: 'people don't even know they have this in their portfolio. It's hidden somewhere inside their pension fund. Ultimately, it backs their life insurance policies. And that risk is being distributed everywhere, in invisible places.'

As a concrete example of this increasingly byzantine financial engineering, the article details the case of Meta's 'Hyperion' data center in Richland, Louisiana. Announced in late 2024 with a capacity of two gigawatts and a cost of about $10 billion —then the largest data center Meta had planned— the project was welcomed enthusiastically by state and local politicians, and the utility Entergy Louisiana proposed building three large natural gas plants to supply it. Last fall, with the cost already up to $30 billion, the financing became far more complex: Meta transferred an 80% stake to the private credit firm Blue Owl Capital, forming a joint venture called Beignet. Meta then signed a series of four-year leases with that joint venture, an arrangement the company says gives it 'long-term strategic flexibility,' backed by a residual value guarantee under which Meta would pay cash if the contract is not renewed or is terminated. The structure gets more complicated still: Beignet created an entity, Laidley LLC, which owns and operates the site as lessor and in turn leases the facilities to Meta's subsidiary Pelican Leap LLC, the lessee. It is no coincidence, Van Nieuwerburgh points out, that the length of the leases (four years) matches the expected useful life of the data center's GPUs; if Meta cancels early it must repay the loan, but investors would be left with an empty building generating no cash flow and the difficult task of finding a new tenant for a facility of that size.

Far from pulling back, Meta has doubled down: in July it announced an expansion of the center to five gigawatts, at a total cost that now stands at $50 billion (it has not said whether Blue Owl will help finance the expansion). Entergy, for its part, now plans to build seven additional gas plants, reaching some 7.5 gigawatts of total capacity —about six times New Orleans's electricity consumption. The complexity of this financing is baffling even to investors and financial experts, and all the more so to affected residents. The main concern is how Entergy's spending on gas plants will affect electricity prices and who will bear the cost if Meta walks away. Entergy maintains that it has a 20-year guarantee from Meta to buy electricity and thereby cover the cost of the plants and associated infrastructure, but there are skeptics: Paul Arbaje, a senior analyst at the Union of Concerned Scientists —an organization that has pushed, largely without success, for the Louisiana Public Service Commission to demand more transparency about the project's financing— wonders whether Mark Zuckerberg will still be interested in the project four years from now or will 'throw in the towel.' Logan Burke, executive director of the Alliance for Affordable Energy, fears that if Meta ends up needing less electricity than expected (the projections are not public), residential consumers will end up paying for the surplus produced by the plants, and she doubts that the shifting web of financial entities will honor agreements twenty years out. Arbaje's conclusion is blunt: 'They're making enormous bets that these data centers will be worth it. Bet with your own money, not with the money of electricity ratepayers.'

The article closes by reflecting on when the bubble will burst, calling it a 'fool's errand' to predict the timing, though it takes for granted that a retrenchment will come. Gensler recalls that 'history tells us that at some point there is a contraction; it's just a question of when and how severe,' and suggests that the current pace of spending ($750 billion) could stall or shrink next year, or that the contraction could arrive around 2028-2029, when companies decide they already have enough installed capacity. The fate of the financial bubble, however, may not match that of the underlying technological revolution: some Silicon Valley investors, such as veteran venture capitalist Vijay Pande, even want a correction to come, arguing in a blog post cited in the article that 'the next crash would be the best thing that could happen to this technology,' since it could rationalize investment and refocus investors on uses that create sustainable value. The piece warns, however, that such a wish calls for caution: after the dot-com bubble burst in the early 2000s, hundreds of thousands of people lost their jobs, numerous companies large and small went under, the economies of Silicon Valley and San Francisco were devastated for a time, and the United States entered a mild recession in 2001. Worse still was the Great Recession that began in late 2007, and the article notes, unsettlingly, that the so-called special purpose vehicles (SPVs) that were central to the financial engineering preceding that crisis 'are back' in the current data center boom. If Van Nieuwerburgh is right about the dangers of letting these investments become entangled with the rest of the economy, the consequences of a retrenchment could be severe. As a counterweight, the text notes that the technologies survived and even thrived after both crises —it is no accident, it points out, that each of today's hyperscalers emerged from the ashes of the dot-com bubble or shortly after it, and that some of the fiber optic infrastructure built during the parallel telecom bubble is still in use today— though the available excerpt of the article breaks off at this point before fully developing that final comparison.