The Central Bankers Are Scared: BIS Warns AI's $100B Debt Stack Could Seed the Next Financial Crisis
The Bank for International Settlements — the institution that coordinates the world’s central banks and whose reports carry the weight of collective official concern — published a late-June warning about the financial architecture underpinning the AI boom. Yann LeCun amplified it on LinkedIn. The BIS is not in the business of overreaction.
The core argument is structural, not speculative: the AI investment cycle has generated a debt stack with embedded feedback loops that make real demand difficult to distinguish from circular financing, and the exposure is concentrated in asset classes that historically amplify downturns rather than absorb them.
The Debt Numbers
Hyperscaler bond issuance topped $100 billion in 2025. That number alone is not the problem — large companies issue bonds regularly. The concern is where the risk landed after issuance. Off-balance sheet vehicles shifted substantial data-centre obligations toward private credit funds, insurers, and other non-bank lenders. Those entities do not have the same liquidity backstops as regulated banks, and they are not subject to the same oversight.
Private credit funds have quadrupled their AI and IT exposure over five years, reaching approximately 15% of total portfolios. Some retail-facing funds within that category are already experiencing redemption pressure — meaning the exit liquidity assumption behind the position is being tested before the underlying assets have proved their cash flow.
Circular Financing
The more novel concern the BIS raises is circular financing. Chipmakers, hyperscalers, AI labs, and compute providers are funding each other while simultaneously booking future sales from each other. Nvidia’s equity participation in CoreWeave and other AI clouds. Hyperscaler revenue-sharing arrangements with inference providers. AI labs whose compute commitments are financed by the hyperscalers whose cloud contracts they simultaneously anchor. Oracle and Microsoft financing SpaceX compute that trains models that generate cloud revenue that finances more Oracle and Microsoft infrastructure.
The circularity is not fraud. But it means that the network of transactions looks like demand when it may partly be balance-sheet recycling. Outside observers — and the BIS explicitly — cannot cleanly distinguish real end-user AI demand from AI companies paying each other in an increasingly closed loop.
Why the Household Exposure Matters
US equities represent 64% of the MSCI Global index. Household equity exposure is at levels above previous cycles, including the periods before the 2000 and 2008 corrections. Goldman Sachs separately projects that data centres will consume 8.5% of US peak power demand in 2027, up from 4.1% in 2025. The infrastructure buildout is now visible in grid planning documents, utility rates, and congressional hearings — it has a physical presence that makes it feel permanent.
But the financing of that infrastructure assumes AI revenue arrives fast enough and at sufficient scale to service debt, justify private credit positions, and support the equity multiples priced into the hyperscalers. If any of those assumptions slip — demand slower than modelled, model commoditisation faster than priced, enterprise deployment cycles longer than projected — the correction propagates through channels that reach retail portfolios and pension funds, not just institutional tech investors.
The BIS is not predicting collapse. Its report explicitly notes that AI can still deliver real productivity gains. The concern is that the financing stack now assumes delivery arrives faster than historical technology diffusion patterns would suggest, and that the non-bank exposure amplifies the correction if delivery is late.
What to Watch
The early indicators would be: hyperscaler capex guidance revisions, CoreWeave and similar infrastructure firms adjusting their contract backlog estimates, or any sign that private credit redemption pressure is becoming systemic rather than isolated. None of those signals are flashing yet. The BIS report is a warning about structure, not a call on timing.
The more immediate question is whether the warning lands. In the 2000 cycle, similar structural concerns were raised about telecom infrastructure investment. The response was to dismiss them as failing to account for the transformational nature of the internet. The internet was transformational. The debt was still a problem.