How big are the risks behind the global AI investment boom? Latest assessments from financiers including Rubin and Zhou Xiaochuan
Goldman Sachs Research estimates that global AI investment will exceed 1 trillion USD in 2026. At the 2026 Bund Summit, Robert Rubin argued that the probability of AI lending triggering systemic risk is very low, while Bank of America chief economist Irigoyen warned that the opacity of private credit is the biggest hidden danger. Nout Wellink, former president of the Dutch central bank, noted that about one-third of AI investment relies on debt financing, and about 30% of private credit funds come from banks, so risks may be transmitted back to the banking system in a roundabout way. Zhou Xiaochuan suggested studying the issue and preparing contingency plans, and stocking tools in the policy toolbox to respond to financial turmoil triggered by failed AI investments.
Body · AI translation
Main body of the article
Goldman Sachs Research estimates that global artificial intelligence (AI) investment will exceed 1 trillion US dollars in 2026. So where will the massive AI investment boom lead the financial system? At the 2026 Bund Summit, several heavyweight figures in international finance gave answers with different emphases.
Beware of hidden risks in private credit
Estimates released by Goldman Sachs Research in August show that global AI investment will exceed 1 trillion US dollars in 2026; some views currently also point out that the scale of AI investment is already far larger than during the internet bubble period. Asked whether this situation could lead to systemic risk, Robert Rubin, the 70th US Secretary of the Treasury, took a negative view.
He believes that, as far as lending to AI-related activities alone is concerned, the probability of a "crash" is quite low. The reason is that the US financial system is currently in genuinely very good shape: large banks have abundant capital and are the main providers of loans to AI-related companies; regional banks with insufficient capital have not participated in AI lending business.
"Of course there are also non-bank lenders, so there are some problems, and these problems may need to be addressed. I think individual companies may have financial difficulties, but the probability of systemic risk is very likely very low," Rubin said.
Claudio Irigoyen, Global Chief Economist at Bank of America, also mentioned the financial risks that non-bank lenders may bring.
"If you ask what else the financial system should be vigilant about, my answer is: all the opacity related to private credit (that is, debt assets in which non-bank institutions lend directly to companies without public market issuance)."
Irigoyen admitted that the reason "Lehman" happened was that banks did not understand the cross-exposures among those important participants; by the time banks realized it, it was too late. So he believes the next crisis will not come from something completely unknown—at least not from the risks banks usually worry about; what is truly worth worrying about is the part with the lowest visibility.
"The market seems to think the scale of private credit is still small, but in the end the problem may arise precisely here," Irigoyen judged.
"I will pay close attention to how much of the AI investment boom is financed by debt." Klaas Knot, one of the candidates for European Central Bank President and former President of the Dutch central bank, observed possible financial risks in the AI investment boom from the perspective of financing forms. He pointed out that AI investment appears to be a field that requires both equity financing and risk-absorption capacity; currently about one-third of AI investment is supported by debt financing. "History tells me that such figures are usually only a lower bound, because in addition there are often off-balance-sheet commitments, and it is difficult for us to adequately monitor these commitments."
Knot believes that the risks of non-bank financial intermediation stem from the fundamental difference between their liability structure and that of banks. These specific risks include liquidity mismatch, herd behavior caused by sudden surges in liquidity demand, and hidden leverage that exists in some local areas.
Essentially, Knot pointed out that non-bank institutions (when they need liquidity) often turn to banks for financing; currently about 30% of private credit funds come from banks. Therefore, the relationship between banks and non-bank institutions is closer to "symbiosis," and risks may also be transmitted back to the banking system through circuitous paths.
So how should monetary policymakers respond to this potential risk? Knot pointed out that first, data gaps must be filled and crisis management tools prepared in advance.
"In a sense we are very lucky." Knot believes the reason is that in the past, risk events in non-bank financial intermediation occurred in an environment of highly accommodative monetary policy, and asset purchases could serve as an effective response tool; now central banks more often approach the inflation target from above rather than from below, that is, when inflation needs to be controlled, central banks must be prudent in providing liquidity. This also requires policymakers to think more ex ante.
"Valuation adjustments will definitely happen at some point"
Overall, Irigoyen believes that AI's impact on the economy is mainly demand-driven. Because this is essentially an investment boom, the countries most obviously benefiting are those deeply embedded in the AI ecosystem supply chain; this boom has also pushed up financial asset valuations and affected consumption through the wealth effect.
From another important dimension of investment, Irigoyen pointed out that every link in the AI ecosystem, from energy producers and chip manufacturers to cloud computing and models, essentially has "winner-take-all" market characteristics. Therefore, every company is investing, hoping to win the market and expand share.
"This means that from the perspective of individual rationality, they will inevitably overinvest, because it is impossible for all companies to become winners." Irigoyen said that at some point, the market must answer a question: is the investment slowdown because a group of companies have exited the competition, or because final demand is not as large as initially imagined? And once the market adjusts, it will immediately transmit to US consumption through the wealth effect.
"This may be one of the biggest financial risks I will talk about today." Irigoyen emphasized that no one currently knows whether the valuations of these companies are reasonable. Under multiple equilibria, if the stock market adjusts—regardless of whether the adjustment has a reasonable basis, the adjustment itself will occur—it will directly affect consumption through the wealth effect.
Zhou Xiaochuan, former Governor of the People's Bank of China, also mentioned that in the evolution of global AI technology, some investments will succeed and some will fail. Considering that huge amounts of funds are currently flowing into AI and related infrastructure, once some (investments) fail, it may trigger financial turmoil. (In such a situation) large-scale action would be needed to support the economy and rescue some financial institutions.
"Therefore, we should conduct research and prepare plans, and reserve corresponding policy tools in the policy toolbox to respond to such turmoil and impact," Zhou Xiaochuan emphasized.
Liu Jun, President of Industrial and Commercial Bank of China, pointed out that commercial banks' business directly involves the daily lives of the general public, involves industrial development, and involves various activities in the market. Therefore, they can only adopt relatively mature and proven technologies, and cannot bet entirely on frontier technologies that are not yet mature. At the same time, in the process of applying AI, it is necessary to resolve the issues of AI's own subject-object positioning and the division of responsibilities.
"In the past, if bank practitioners used a certain computer system to conduct business, they had to bear responsibility—bank practitioners could not shirk it, nor could they tell the public, 'I did not develop this system, so I am not responsible.'" Liu Jun said that in the process of AI development and application, there are many links: algorithm developers, programmers, deployers, users, all the way to end users. All parties are in different links, and the algorithm itself is to some extent a "black box." Therefore, how should ultimate responsibility be determined and verified? Who should bear it? This is a question that is difficult to define at the current stage.
He suggested that if AI is regarded as a human-like "quasi-legal subject," and AI financial governance is combined with governance of human subjects and the existing governance system of market participants, it would be possible to establish a unified governance mechanism. Regarding the regulation of AI applications, he pointed out that the data trails generated during AI operation could be integrated with relevant data of existing market participants and incorporated into the same system, achieving unified management and full-process connectivity, covering from the initial link to the final link.
"Because we do indeed have data labeling technology and tracking capabilities; after incorporating AI into consideration, using existing mechanisms it is possible to solve the problem of verifying and identifying systemic risk," Liu Jun said. Looking further, by combining the four factors of proof of work, proof of history, proof of stake, and proof of revenue, it would be possible to identify and clearly define what ultimate responsibility AI bears in a certain business activity.
"If Industrial and Commercial Bank of China deploys a large number of AI tools and agents across different business lines, and all returns enter Industrial and Commercial Bank of China's balance sheet and income statement, that means I bear the responsibility for AI deployment; once a serious problem occurs, I should compensate the relevant victims and give the market a final answer." Liu Jun gave an example of the idea that once the attribution of returns is clear, the responsible entity is also clear.
Liu Jun pointed out that in specific practice, Industrial and Commercial Bank of China implements this rule in business tiers: for businesses that are highly data-dependent and do not directly face customers, AI is given greater autonomy, and the institution bears the consequences; for businesses that directly face customers, a relatively large human team is retained, and major decisions are made by humans "in the loop"; for businesses involving price management and large financial and social exposure, humans must be present from the very beginning of the process.
Compiled from the Bund Summit and Xinhua Finance
(Source: Securities Times) Middle of the main text Embedded advertisement