The AI buildout is no longer just a chips-and-data-centers story. The Bank for International Settlements now sees it as part of the same risk map as public debt, bond liquidity and leveraged finance.
The important line in the BIS annual economic report released on June 28 is not that AI investors could be disappointed. You already knew that. The harder point is that the central bank for central banks is now treating AI spending, government borrowing and market plumbing as connected pressure points in one fragile system.
If you are building or funding an AI company, you cannot treat demand for compute as separate from credit spreads, interest rates or sovereign bond stress. The same money that finances the next data center also prices government debt, private credit funds and the equity valuations that make ambitious AI budgets look painless.
The BIS says the five largest hyperscalers are set to spend more than a trillion US dollars on AI-related capital expenditure from 2025 through 2026. Those commitments are outpacing earnings and free cash flow, the report says, pushing some firms toward debt issuance. According to the Financial Times, the BIS warned that disappointing returns could turn the capex boom into a long investment bust, with knock-on effects across financial conditions.
Frankly, that is a stronger warning than the usual bubble chatter. A bubble take is easy. The BIS is describing how trouble could travel.
AI optimism helped global growth hold up in 2025, the BIS says, even as tariffs and geopolitical uncertainty weighed on the economy. It also kept financial conditions loose by lifting stock valuations. That is the friendly version of the story: AI spending supports investment, raises confidence and gives the economy another engine when old engines are sputtering.
The unfriendly version is sitting a few pages later. The BIS says equity valuations are elevated for firms at the center of AI development, with implied long-term earnings growth for the largest companies above recent historical benchmarks. It also says US stocks account for about 64% of the MSCI Global index, so a US-led AI repricing would not stay neatly inside Silicon Valley portfolios.
Credit is where this becomes less abstract. The BIS says fixed income markets are vulnerable because hyperscalers, AI labs and engineering, procurement and construction contractors have issued large volumes of debt tied to the buildout. If hyperscalers slow capex, suppliers across the chain may struggle to replace revenue and service debt. You do not need a model to understand the risk. A data center contractor with borrowed money and a customer pause is not holding an AI thesis. It is holding a refinancing problem.
The report also flags circular financing, where hyperscalers take equity stakes in AI labs that then commit to buying chips or compute. The BIS says those announced multi-year commitments were widespread as of April 2026, and its own graph cites Bloomberg, CNBC, LSEG Datastream, S&P Global Market Intelligence and The Wall Street Journal as sources. That is exactly the sort of detail investors should dislike, because a clean revenue chart may be hiding a messier balance sheet underneath.
Public debt is the other half of the story
The AI warning landed in a report that is just as worried about government bond markets. The BIS says non-bank financial institutions have become the largest holders of sovereign debt in advanced economies. Their share of total sovereign debt holdings rose from 44% in 2021 to 53% in 2025.
That is not a footnote. Hedge funds, money market funds, open-ended bond funds and other non-banks do not behave like old bank balance sheets. The BIS says leveraged, funding-dependent non-banks can amplify liquidity shocks, especially through repo and derivative markets. The phrase it uses is a new fiscal-financial stability nexus. You can translate that plainly: when markets lose confidence in public debt, the selloff can move through intermediaries faster than governments and central banks would like.
The Times, covering the same report, pointed to the BIS warning about fire sales in bond markets and compared it with the UK gilt turmoil after Liz Truss's 2022 mini-budget. That episode forced the Bank of England to intervene after yields spiked and pension fund strategies came under stress. The BIS is not saying every bond scare becomes that. It is saying the structure that can turn a scare into a market-functioning problem now matters more.
For startups, the practical point is uncomfortable. You may think your risk is product-market fit, GPU access or whether customers move pilots into production. Those still matter. But if your business model assumes cheap capital, rising cloud budgets and public markets willing to reward losses in the name of future AI dominance, you are already exposed to the same macro trade as everyone else.
Investors should be blunt with themselves too. A founder selling AI demand in 2026 is also selling a view on rates, credit availability and the willingness of hyperscalers to keep spending through disappointment. If corporate credit tightens, the BIS says that could expose vulnerabilities in private credit, where smaller firms have become more dependent. That reaches the venture market through customer budgets, exit multiples and the cost of every infrastructure-heavy promise.
None of this proves AI is fake. The BIS explicitly says AI could raise productivity significantly over the coming decade, and its report includes scenarios where output growth improves. But belief in the technology is not the same as belief in every financing structure wrapped around it. Canals, British railways and the dotcom boom all involved real technology. They still overbuilt.
So the current state of the AI boom is not simply bullish or bearish. It is leveraged. If you are building into it, watch the customers. If you are funding it, watch the spreads. The BIS has made clear that those are now the same conversation.
Also read: How to Make Money With AI: Nine Models That Don't Need Code | China's LineShine ends nine years of US dominance in supercomputing | The $680 billion question is not how much gets spent on AI but who actually keeps the money
