What's at stake in AI's trillion-dollar bet

David Rotman's article in MIT Technology Review starts from what Wharton finance professor Jessica Wachter calls a "remarkable fact": a handful of hyperscalers are pouring colossal sums into building AI data centers, and that figure will reach close to $1.1 trillion cumulatively by 2027. Instead of trying to predict how useful AI models will be, Wachter and her co-author calculate how much those companies' profits would have to grow to justify that spending: by their estimate, the productivity of the AI companies themselves would have to multiply by 2.7 between now and 2030 to cover the cost of capital, a 15% return and the depreciation of the assets. It is not impossible —it would be equivalent to the kind of growth brought by the US tech boom of the mid-1990s, but compressed into very few years. If that productivity surge does not arrive, warns Wachter (a former chief economist at the SEC), companies will start falling behind on interest payments, with a risk of bankruptcy, and the current infrastructure boom could turn into "the biggest misallocation of capital in history".

The article's figures are stark: the hyperscalers (Alphabet, Microsoft, Amazon, Meta and Oracle, the latter a partner of OpenAI) will spend some $750 billion this year on data centers, and some projections put cumulative spending at more than $5 trillion over the next four years. Against that, total revenue generated by AI will come to only between $150 billion and $200 billion this year, according to Gary Gensler, former SEC chair during the Biden administration and now a professor at MIT's Sloan School: "The spending doesn't yet have proportionate revenue. That is a fact," he says. Total investment could come to represent close to 3% of US GDP, which raises the profitability question to a matter of macroeconomic health, not just a corporate one.

The risk has worsened this year because the companies have begun financing much of that expansion with debt instead of only their own cash. The article notes that free cash flow for the sector as a whole (operating cash minus capital expenditure) is about to turn negative. Even Alphabet, famous for piling up cash, posted revenue of close to $120 billion in its latest quarter that was devoured by AI infrastructure spending, leaving a free cash deficit of some $5.9 billion, its first since going public in 2004. On top of this comes a structural problem: GPU chips account for about 60% of the cost of a data center and their performance roughly doubles every two years, forcing constant reinvestment in the next generation of hardware. Mihir Kshirsagar, of Princeton's Center for Information Technology Policy, warns that without that constant reinvestment data centers risk becoming "hulks", stranded assets scattered across the country.

Gensler describes the situation as a "parlay bet" by the capital markets: for the play to pay off, three simultaneous and interdependent conditions must be met. First, the hyperscalers must generate multibillion-dollar revenues; Columbia Business School finance professor Stijn Van Nieuwerburgh calculates, starting from a scenario of 183 gigawatts of AI computing capacity planned between 2025 and 2032 (at an approximate cost of $41 billion per gigawatt), that annual revenues of some $3.7 trillion would be needed by 2032 to achieve a 10% return, the minimum acceptable to most investors. Second, AI must translate into broad productivity growth across the whole economy, not just into subscriptions. Third, the most expensive and powerful models (the so-called "frontier models") must continue to be preferred over cheaper alternatives that many companies might consider "good enough".

On productivity, MIT economist and 2024 Nobel laureate Daron Acemoglu warns that if the productivity gains do not materialize, people will start to become disenchanted with AI, which would slow both investment and revenue growth. For now, according to the article, most economists agree that the economy's aggregate statistics show little or no productivity growth attributable to AI. A recent survey of some 6,000 senior corporate executives in the United States, the United Kingdom, Germany and Australia shows that close to 90% saw no productivity increase over the past three years, although they expect a cumulative improvement of around 1.45% over the next three (2.25% among US executives). Those same executives plan to increase AI spending, which points to some $280 billion in private AI spending by the end of 2026. But the figure comes with a warning: executives expect to improve productivity mainly by increasing sales while significantly cutting headcount, which could intensify the social backlash against the technology if productivity gains are achieved by destroying jobs rather than creating new value.

The report stresses that the risk is no longer confined to the tech companies' own balance sheets, as it was when they were spending only accumulated cash. According to Morgan Stanley calculations cited in the article, more than half of the $2.9 trillion the hyperscalers will spend between 2025 and 2028 on AI data centers will be financed with "outside capital". Van Nieuwerburgh warns that many financial institutions are exposed to these data centers, directly or indirectly, as lenders, debt guarantors or backers of private credit funds: "People don't even know they have this. It's somewhere deep in their pension fund. Ultimately, it backs their life insurance policies. And that risk is being spread everywhere, in invisible places," he says.

The article illustrates this increasingly complex financial engineering with 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 some $10 billion, it was greeted enthusiastically by local and state politicians, and the utility Entergy Louisiana proposed building three large natural gas plants to supply it. Last fall, with the projected cost already at $30 billion, Meta transferred an 80% stake to the private credit firm Blue Owl Capital, forming a joint venture called Beignet to finance the project. Meta then signed a series of four-year leases with that joint venture, backed by a residual value guarantee, and set up a structure of subsidiaries (Laidley LLC as owner and lessor, and Pelican Leap LLC as Meta's tenant) whose contract terms, according to Van Nieuwerburgh, coincide not by chance with the expected useful life of the data center's GPUs. In July, Meta announced an expansion to five gigawatts with a total cost of $50 billion, while Entergy is now planning seven additional gas plants, up to some 7.5 gigawatts of capacity —about six times New Orleans's electricity consumption.

That financial complexity is a direct concern for the affected community. Entergy says it has a 20-year guarantee that Meta will buy electricity to cover the cost of the gas plants, but analysts such as Paul Arbaje, of the Union of Concerned Scientists, question whether the company's commitment will hold up over time given how fast the AI industry's fortunes change, and have asked —so far without success— for greater transparency from the Louisiana Public Service Commission. Logan Burke, of the Alliance for Affordable Energy, warns that if Meta ends up needing less power than projected, residential consumers could end up paying for the plants' surplus, and doubts that the shifting web of financial entities involved will honor the deal for two decades. For Arbaje, the summary is simple: these companies are betting that the data centers will be worth it, but they should bet with their own money, not with that of electricity customers.

On how it all ends, the article is cautious: predicting when the AI investment bubble will burst is, in the author's own words, "a fool's errand", but there is little doubt that at some point a retrenchment will come, whether because spending stalls next year or because by 2028-2029 companies find they already have enough capacity, as Gensler suggests. Venture capital investor Vijay Pande went so far as to write in a blog post that a collapse would be "the best thing that can happen to this technology", an argument Rotman qualifies by recalling the human cost of past bubbles: the dot-com bust left hundreds of thousands of layoffs, bankruptcies and a mild recession in 2001, while the great recession that began in late 2007 —with its own special purpose vehicles (SPVs), now back in the financing of AI data centers— was even more devastating. Even so, the article recalls that the underlying technology usually survives and even thrives after these collapses: it mentions the fiber-optic infrastructure built during the telecom bubble, and notes that it is no coincidence that each of today's hyperscalers emerged from the ashes of the dot-com crash or was born shortly afterward.