The economic risks of the AI boom

Artificial intelligence has many — mostly far from fully understood — implications for humanity. Among these is something concrete and current: it has launched a huge surge in spending on physical capital, to meet the demand for huge data centres. Indeed, according to a fascinating paper, “Financing the AI Buildout” by Stijn van Nieuwerburgh of the Columbia Business School, spending in the US will average 3.6 per cent of GDP from 2025 to 2032 (inclusive). This would be larger than the average of 2.2 per cent of GDP spent on railways between 1870 and 1890, let alone the 1.1 per cent spent on highways from 1956 to 1973. In fact, it would be the biggest sector-specific investment boom of all.

Some of those earlier transformations were sustained over far longer periods. But this fact makes the AI build-out even more concentrated and so, quite possibly, even more disruptive. Moreover, as van Nieuwerburgh explains, financing the boom is now creating some classic risks, despite the strength of the huge companies (mainly Alphabet, Amazon, Meta, Microsoft and Oracle) driving the huge surge in spending.

Moreover, these risks are also macroeconomic. Booms can easily turn into busts, as stock market valuations collapse, companies go bankrupt, lenders lose money and planned investment vanishes. Will this time be different? Ultimately, that depends on the profitability of the new businesses and, in this case, hopes are very high indeed.

As the paper points out, aggregate capital expenditures by the companies listed above rose from $97bn in 2020, to $416bn in 2025 and a forecast of $801bn in 2026. In 2020, the surplus of operating cash flow of these companies over capital spending was $155bn. In 2025, they still had a cash flow surplus of more than $187bn. But over these years, the ratio of capital spending to operating cash flows rose from 38 per cent in 2020 to nearly 69 per cent in 2025. Then, in 2026, according to the companies’ forecasts, capital spending will exceed cash flow by $93bn and the ratio of spending to cash flow will hit 113 per cent. Alphabet and Microsoft will still spend a little less than their cash flow. But Oracle’s spending will be 138 per cent above it, Amazon’s 36 per cent above it and Meta’s close to it.

As a result, ever more complex financial arrangements are being made. Thus, not for the first time, an investment boom is causing an explosion in leverage. While the hyperscalers’ direct investment is still important, “the marginal financing model is shifting. Specialized data center developers, REITs, infrastructure funds, and private equity investors increasingly build or acquire facilities and lease capacity to hyperscalers and other large users.” The result, according to the paper, is an increasingly complex and non-transparent financial structure. In particular, it notes, “publicly rated debt represents only a relatively small share of the datacenter financing; much larger volumes are provided through syndicated bank loans and private-credit facilities, for which little information is publicly available.”

In sum, “the central issue is not simply that AI infrastructure is exposed to technological and operating risks, but that the sector’s financing structure can transmit and amplify those risks.” This is historically familiar behaviour. Risks for data centre owners include: concentration in the tenants and also in sources of final demand; technological obsolescence and so the need to scrap and replace relatively quickly; and political and operational difficulties involved in making the investments work. Moreover, van Nieuwerburgh adds, “Chip designers, cloud providers, data center operators, and infrastructure investors increasingly depend on one another through long-term contracts, strategic investments, and financing arrangements.” This creates fragility.

These are not just issues for the AI sector itself. The investment boom is big enough to have large macroeconomic implications. A paper from the Federal Reserve suggested that spending on computer hardware, software and data centres contributed more than a percentage point to annualised quarter-over-quarter US GDP growth in the first quarter of this year, partially offset by a half a percentage point of GDP increase in net imports in those categories. Both the IMF’s World Economic Outlook and a recent Substack from Matt Klein shed light on the international spillovers of the global AI boom, especially on to Taiwanese and South Korean products.

A crucial question, however, is how far the AI mania has infected the stock market, above all in the US. Here some of the vulnerabilities are evident. First, the Magnificent Seven (Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia and Tesla) have, until recently, indeed been pulling the US stock market ever higher. Second, the market is at valuations rarely seen before, with a cyclically adjusted real earnings yield of 2.5 per cent, which is only fractionally above the current real interest rate. This only makes sense if earnings from stocks are about to grow far faster than ever before. In the past, however, whenever earnings have been this low relative to stock prices, the latter have crashed. Is this AI-driven boom, with all its unavoidable uncertainties about future profitability, returns on the huge and rapidly growing investment, and now the elaborate financial engineering, going to be any different?

The longer the investment boom and the associated financial engineering continue, the bigger the negative economic impact is likely to be when it ends. This does not mean that AI itself will fail to be the transformative technology it seems to be, though only a species that adores playing Russian roulette would tolerate a business (Anthropic) whose prospectus says it could create “catastrophic or existential risks to humanity”. New technologies have often triggered stock market manias and investment booms that end in busts. Recent market tremors suggest this could even be happening to AI now. A return to reality will come when it will come. The readiness is all.

martin.wolf@ft.com

Follow Martin Wolf with myFT and on Twitter