Circular financing, on-balance sheet & off-balance sheet debt, GPU-backed debt, rapid deleveraging potential, asset repricing, etc. It’s all there.
By Wolf Richter for WOLF STREET.
The risks associated with the AI investment boom – financed by a borrowing boom, circular financing, leverage and sky-high stock prices – have been talked about for some time. The Federal Reserve also warned about the financial stability aspects of this boom if and when it turns bad, particularly in its latest semi-annual Financial Stability Report (FSR), released Friday.
Fed Governor Michael Barr laid out in a lengthy speech today the risks of AI to financial stability – and they are significant, complex and opaque. Here are 10 of those risks, in Barr’s own words:
1. AI investment losses could hurt the economy.
“In one of several possible scenarios, even if AI proves to be a transformative innovation that significantly improves productivity, a key question is whether investors will see returns on AI development in a time frame and to an extent consistent with their expectations or whether a different outcome could lead to a significant revaluation.
“The potential investment losses that could result from a large-scale price revision could dampen output growth in two ways: first, through the direct effect of a decline in investment, which is a component of gross domestic product; and, second, by the knock-on effects of this loss of wealth in reducing consumer and potentially business spending.
2. Debt, circular financing and other complex financial obligations.
“The Federal Reserve’s Financial Stability Report (FSR) examines three ways in which financing AI data centers and related investments, which it calls “advanced IT debt,” could affect financial stability:
- the rapid growth of this debt,
- circular financing networks,
- other complex financial obligations.
“In the event of a significant reassessment of AI investments, tensions related to these three aspects of AI financing could amplify the reassessment of AI investments. »
3. $800 billion balance sheet debt of the largest AI companies:
“Debt on the balance sheets of the seven largest AI companies increased significantly in 2026 and is expected to exceed $800 billion before the end of the year.
“The growth in debt of these companies is reminiscent of the growth of debt in the telecom investment cycle of the late 1990s and early 2000s… Although hyperscalers have not yet surpassed their comparators from the previous boom, the recent acceleration of their capital spending suggests that they could do so soon.”
4. Opaque off-balance sheet debt of $3 trillion at the largest AI companies:
“While the debt levels on these companies’ balance sheets have increased significantly, their other obligations, such as long-term contracts, lease commitments and residual value guarantees, have increased even more rapidly.
“The FSR estimates that the maximum potential face value of these types of bonds, or off-balance sheet liabilities, now exceeds $3 trillion.
“A significant portion of these obligations arise from complex arrangements in which hyperscalers and other large companies with strong credit ratings indirectly support the financing of projects involving other companies with weaker credit, or support public benefit investments.
“These off-balance sheet liabilities include obligations related to energy supply contracts, long-term leases of data and computing centers, as well as a wide variety of off-balance sheet financing structures….
“This financial complexity can be opaque to unsophisticated investors, and this opacity can exacerbate sales when investor sentiment changes, because it is difficult to know where the risk lies….
“Perhaps just as importantly, in addition to hyperscalers, lower credit quality companies that rely most on external financing have seen their cost of capital increase significantly in recent months and, as a result, are turning to more complex asset-backed financing.

5. $100 billion in circular financing in small businesses.
“Hyperscaler companies and large producers of graphics processing units (GPUs) also have more than $100 billion in cumulative equity investments in large language model labs (LLMs) and cloud computing start-ups.
“These equity investments represent a form of ‘circular financing’ in which hyperscaler companies and GPU manufacturers invest capital in smaller companies, and the smaller companies turn around and use those funds to lease or purchase GPUs and perform computing from the company that made the investment.
“These investments represent a relatively small portion of the balance sheets of hyperscalers and GPU producers, but they are essential for the small companies that receive these investments.
“For investments in AI development to ultimately become economically viable, small businesses will eventually need to replace this funding from their suppliers, hyperscalers and GPU manufacturers, with revenue from paying customers.”
6. GPU-based asset-backed circular financing.
“Another indirect form of circular financing in which hyperscalers and GPU producers support purchases of their products and services is the use of collateral guarantees for asset-based financing.
“As companies more dependent on external financing face financing constraints, they are moving away from issuing corporate bonds to issuing asset-based financings. To support the credit of these financings, larger companies with higher credit often guarantee the resale value of the data centers or GPUs that are used as collateral to support the financings.”
“These guarantees, or collateral guarantees, are recorded as off-balance sheet liabilities because they represent the extent to which hyperscalers and GPU producers could absorb losses in the event of default and liquidation of the guarantees.”
7. Potential for “rapid deleveraging”.
“Given the still limited amount of debt supporting the current boom, I am less concerned about direct losses.
“However, with hedge funds’ leverage and risk appetite apparently quite high, a sharp reversal could lead to rapid deleveraging and, potentially, self-reinforcing flash sales.
“In addition, the leverage used in equity strategies has also increased in recent years. This leverage is provided, in part, by the prime brokerage services of large dealers. So, while high levels of capital at the largest banks provide a buffer against spillovers, counterparty exposures could increase in times of equity market volatility, so it is important to continue to monitor exposures, including exposures to counterparties that have taken large positions in AI.
8. Rapid depreciation of GPUs securing debt.
“Additionally, most of the cost of a data center comes from GPUs and other specialized chips used in AI.
“One concern is that these specialized chips could depreciate quickly as new technologies overtake them, so the capacity they represent won’t be of much use in the future.”
9. Potentially unnecessary electrical infrastructure.
“It is difficult to say what other elements of the AI-related capital stock – substations, high-voltage connections and grid interconnection – could be used for if demand for their services disappoints.
“In some parts of the country, there is no doubt that additional power and network capacity will be useful, whatever the future of AI.
“In other areas, it may be years before excess capacity becomes useful, as was the case with ‘dark fiber’ in some areas after the boom and bust of the 1990s.”
10. AI Stock Trading: “Market Rally and Possible Rush.”
“If traders rely on a small set of similar AI models trained on similar data, they run the risk of herding behavior: trading strategies could become similar and synchronize reactions to news and other information.
“Flaws in algorithms can cause failures, but a market with a large number of similar models likely to have the same flaw risks failures with systemic consequences.
“Additionally, bad actors could intentionally manipulate trades, quotes or news and potentially mislead similar AI trading systems at the same time, thereby amplifying market disruptions.
“AI models can echo unfortunate human tendencies. For example, some theoretical research shows that certain reinforcement learning algorithms, called Q-learning trading algorithms, can autonomously develop collusive behavior without explicit intent.”
“Such self-learning algorithms can also autonomously learn to manipulate markets – for example, through spoofing, which involves placing orders designed to create a false impression of market demand – or adopting trading strategies that create and exploit bubbles, thereby destabilizing markets. »
Do you like reading WOLF STREET and want to support it? You can make a donation. I appreciate it immensely. Click on the mug to find out how:

![]()
Gn bussni
