Bryan Judge is the director of research for the finance and democracy program at the University of California, Berkeley. Recently, Kevin Warsh, president of the Federal Reserve System (Fed), announced the creation of a new working group to "examine the pace of dissemination of new general-purpose technologies, including artificial intelligence (AI), their adoption and economic impact, as well as their implications for the Federal Reserve" in carrying out its mandates regarding employment and inflation. Notably, Warsh's statement made no mention of the potential impact of AI on financial stability, even though financial stability is effectively the Fed's third mandate.
The Fed has somewhat recognized the risks that AI poses to financial stability. In April, former Fed Chair Jerome Powell, along with U.S. Treasury Secretary Scott Pessen, held a meeting to assess how advanced AI models might affect the cybersecurity of the banking system. However, even this approach was too limited.
In the United States, both the financial system and asset markets have become one-sided bets on the success of AI. If current trends continue, by the end of this decade, the unpaid debt for financing AI data centers will exceed mortgage debt. Now the Fed must determine whether the revenues from these data centers will generate sufficient cash for creditors to make timely repayments.
This question is often associated with two other questions: whether artificial intelligence is a bubble and whether it is a transformative technology. There is a temptation to present this as a dichotomy: AI is either a bubble or a transformative technology, but that would be a mistake. In the 19th century, the railroad boom in the United States ended with the panic of 1873, and the telecommunications and internet boom of the 1990s ended with the stock market crash. In both cases, the technology was indeed transformative, but creditors and shareholders still suffered heavy losses as the volume of investment exceeded the realistically expected returns in the short term.
Similarly, when it comes to AI, technological success does not guarantee financial success. It is still unclear how much revenue AI will generate, but the repayment terms for borrowers' obligations are predetermined. AI tools may be widely applied, and data centers may be widely used, but they may not generate sufficient cash to service their owners' debts. Concerns about financial stability arise from the mismatch between projected future revenues and current contractual obligations.
These calculations are staggering. According to David Kani, a partner at Sequoia Venture Capital, this year, capital expenditures on AI by hyperscale companies are expected to be around $750 billion, which must generate about $1.5 trillion in revenue from end customers to justify those expenditures. According to his calculations, since the launch of ChatGPT in 2022, at least about $3 trillion in cumulative revenue will be needed to offset the investments made to create the entire AI infrastructure. Reports suggest that Anthropic's annual revenue is around $60 billion.
According to Bain & Company, to meet the expected demand for AI by 2030, funding for computational power will require about $2 trillion in new annual revenue in the AI sector. Given that a financial "bubble" is considered a situation where the price of an asset significantly exceeds the cash flows it generates, such forecasts seem to support warnings that AI is indeed a bubble.
Funding for the development of AI infrastructure has decisively shifted from the cash flows of tech giants to capital markets. Circular financing schemes have become widespread: chip manufacturers invest in AI laboratories, which use those funds to purchase chips, while cloud service providers fund startups that rent their servers. As a result, a positive feedback loop is formed between rising valuations and capital expenditures.
The chip manufacturing giant Nvidia has become a financial backbone for "neocloud" companies, allowing cloud providers with limited capital to attract private funding on favorable terms. Tech giants are accumulating massive off-balance-sheet obligations through joint ventures and leasing structures. The growing share of capital is provided by private credit funds that often finance projects linked to their sponsors.
Unlike previous investment booms that resulted in long-lasting assets such as railroads and fiber-optic cables, these investments do not leave behind such assets. Chips account for about half of the value of an AI data center, and in 3-5 years, they will practically be unusable. As a result, the value of collateral may decline faster than the debt is repaid.
Moreover, the expected productivity gains from AI and its impact on the labor market are still not visible in the data. According to the latest analytical note from Fed staff, this is because AI is still in its "development" phase. However, a sharp increase in productivity will also require companies to make significant internal investments to reorganize their processes. Nevertheless, markets are already reflecting significant profit growth expectations, partially driven by productivity gains resulting from AI, raising concerns about a "profit bubble".
Much of the discussion about the risks of a downturn in the ongoing AI boom has focused on the stock market, but a much greater risk threatens credit markets. We currently have a "market-based" financial system where lending is less done through banks and more through bond markets, securitization structures, and non-bank lenders. The danger is not a mass withdrawal of bank deposits in the style of the 1930s, but a massive outflow of funds from the shadow banking system in the style of 2007: doubts about the quality of loans lead to a reduction in short-term financing, and borrowers are forced to sell assets in an already declining market, leading to further price declines.
Due to the paralysis of short-term funding markets, the Fed will be under enormous pressure to support non-bank lenders and the debts of data centers, as it did in early 2020 during the COVID-19 pandemic by supporting money market funds. However, AI is currently less popular than Wall Street was in 2007. The salvation of both will likely destroy the last remnants of the Federal Reserve's independence.
It is possible that the massive capital expenditures in the AI sector will be justified by generating the revenues necessary to service trillions of dollars in debt. However, in that case, the anticipated shift in the labor market will have no historical precedent. This is a nightmarish scenario of "pen and ink": if the "pen" falls, the result will be financial instability, and if the "ink" falls, there will be a shock of biblical proportions in employment. In any case, financial stability must occupy a central place in the Fed's AI agenda. Even if this time everything develops differently from a technological standpoint, financially, nothing may change.
