The surge in investment needed to build artificial intelligence infrastructure could expose companies and the wider economy to significant risks, Martin Wolf argues in an opinion article. The concern is not that AI will cease to be transformative, but that the investment boom may be difficult to sustain if the expected returns do not materialize.
A study by Stijn van Nieuwerburgh of Columbia Business School estimates that US spending on AI infrastructure will average 3.6% of gross domestic product from 2025 through 2032. That would exceed the average shares devoted to railways from 1870 to 1890 and highways from 1956 to 1973, making it the largest recorded investment cycle in a single sector, according to the study.
Capital spending by Alphabet, Amazon, Meta, Microsoft and Oracle rose from $97 billion in 2020 to $416 billion in 2025. Company projections put the figure at $801 billion in 2026. While the five firms still generated more operating cash than they spent on capital investment in 2025, the gap is expected to reverse in 2026: projected investment would exceed operating cash flow by $93 billion, with capital spending reaching 113% of cash generation.
The financing of data-center construction is also becoming more complex. Alongside direct investment by large cloud companies, specialist developers, real-estate and infrastructure funds, and private-equity investors are building or buying facilities and leasing capacity to major technology firms and other users. Van Nieuwerburgh’s study says public-rated debt accounts for only a relatively small share of data-center financing; larger amounts come through syndicated bank loans and private credit, where public information is limited.
That structure could transmit and amplify risks, the study warns. Data-center owners face potential tenant and demand concentration, rapid technological obsolescence and political or operational obstacles to projects. The sector is also increasingly interconnected: chip designers, cloud providers, data-center operators and infrastructure investors rely on one another through long-term contracts, strategic investments and financing agreements.
The investment cycle is large enough to affect the broader economy. A Federal Reserve study found that spending on computer equipment, software and data centers contributed more than one percentage point to US annualized GDP growth in the first quarter, compared with the previous quarter. Net imports in those categories offset part of that contribution.
Wolf also points to vulnerabilities in US equity markets, where the so-called Magnificent Seven have helped drive share prices higher and valuations have reached levels rarely seen. He notes that a cyclically adjusted real earnings yield of 2.5% is only slightly above the current real interest rate, a level that would require unusually strong future profit growth to be justified.
Past investment booms driven by new technologies have sometimes ended in financial crises. Wolf argues that the longer the AI investment cycle and its associated financial engineering continue, the greater the potential economic damage when the boom ends. Recent market tremors, he writes, suggest the possibility of a reckoning, though he does not claim that AI itself will lose its transformative potential.
Sources: Folha Mercado.