The wrong question is whether AI is a bubble. The right question is whether the financing of AI is brittle. History’s worst market breaks did not require technology to fail; they required balance sheets to be built like scaffolding in a windstorm. If the Bank for International Settlements is right, today’s pressure points are familiar: debt, leverage, opacity, and a herd convinced that this time cash flow will show up before the bill is due.
Consider the paradox. If AI delivers real productivity gains, the first-order winner is the sector that can deploy capital fastest. But rapid deployment is not the same as resilient financing. Hyperscalers are committing to over a trillion dollars of AI-related capex in the next two years, with obligations that run ahead of current earnings. That is not sin; it is a duration bet. The cash coming back must arrive on schedule and at scale. Meanwhile, energy, chips, real estate, and labor need to keep pace. If they do not, success becomes inflationary. Higher input costs and higher rates would compress multiples at the exact moment debt service climbs. The tech cycle can work while the financing cycle fails. Markets are machines that punish timing mismatches.
The BIS flag on circular deals is overdue. Vendor financing helped inflate the late-1990s telecom buildout. Equipment makers funded customers to buy more equipment, reporting growth that sat on a pile of IOUs. Today’s version is cleaner but rhymes: chipmakers and hyperscalers take stakes in AI labs or neocloud providers who in turn pledge multi-year purchases of chips and compute. Data centers are outsourced to third parties and leased back on long contracts with exit clauses. Economically, that stack resembles debt with options sprinkled in. When terms are thinly disclosed, the same revenue is counted as both customer demand and collateral support. When the pressure arrives, obligations telescope. Contracts that were treated as growth become leverage. As with the 2001 unwind, vendor IOUs do not default one at a time; they reprice together.
Markets do not collapse because an asset goes from 100 to 80; they collapse because that 20-point drop triggers margin, then funding, then forced sales elsewhere. Poorly disclosed AI-related arrangements add a new collateral chain to an old problem. If the same future cash flow supports multiple promises, the system is short convexity. One counterparty’s risk reduction is another’s loss of funding. Post-2008 reforms ring-fenced banks; the fragility moved. Hedge funds and non-banks now intermediate duration with short-term repo and basis trades. In normal times, collateral velocity is efficient. Under stress, it becomes a sandpile: a few extra grains and the slope gives way. UK gilts showed the feedback loop in 2022. A disorderly repricing in any crowded trade now propagates across borders in hours, not weeks. Chains fail at their weakest link, not their average strength.
Inflation risk is no longer a tail. A persistent overshoot of targets has rewired expectations. AI is electricity plus silicon plus real estate. Power grids and transformers have multi-year lead times; diesel backup is not a growth plan. If the AI buildout accelerates into tight energy markets, input costs rise. Higher electricity prices bleed into everything from food processing to logistics. The positive productivity story competes with the negative cash flow math of higher rates. In game theory terms, the equilibrium is fragile. If central banks lean against inflation with tighter policy, they hit a capital-intensive sector mid-sprint. If they do not, they risk unanchoring expectations. Either path exposes duration mismatches embedded in circular deals and leasebacks. The claim that nominal growth bails out capex assumes funding costs do not move first.
Fiscal fragility is the oldest risk in finance. High debt loads reduce the system’s shock absorbers. The marginal buyer of sovereign bonds has changed. Levered funds have taken a larger role, harvesting tiny spreads through basis trades funded in repo. It works until haircuts rise or volatility jumps. Then it becomes a deleveraging machine. A term premium that drifts up by 50 basis points sounds manageable until you stack it on top of mark-to-market losses, collateral calls, and redemptions. The BIS warning is simple: when the funding base is short and mobile, liquidity transforms into exit pressure. A gilt-like episode in a larger market would not stay local. The feedback loop between sovereign yields, bank balance sheets, and non-bank funding is a roundabout with too many entries and no lights.
It is not just macro. The plumbing is changing beneath regulators’ feet. Stablecoins present as cash, trade as money, and behave as money until redemption frictions arrive. Their rapid growth risks fragmenting monetary control and adds a new layer of runnable liabilities outside the banking perimeter. Permissionless blockchains layer on operational and legal risk for banks that touch them. Unknown validators and untested fallback procedures are not comforting in stress. The BIS is steering the system toward tokenized versions of bank money, where claims are clear and governance has a phone number. That migration will take time. Meanwhile, an on-chain off-ramp panic could propagate across exchanges, market makers, and prime brokers before supervisors finish their first call. Liquidity illusions end the same way no matter the substrate.
Call it an inversion test. Assume AI works. Assume it earns its cost of capital in the long run. Does today’s funding structure survive a two-year earnings air pocket, a power bottleneck, or a modest equity correction? The answer depends on leverage, disclosure, and the replaceability of funding. If a growth story needs long-dated, patient capital but is fed by short-dated repo, circular commitments, and off-balance-sheet leases, the system is poised for a classic liquidity squeeze. If the same assets and contracts are promised to multiple parties, the unwind will be nonlinear. Investors chase antifragility and often buy fragility packaged as scale. Big is not the same as robust. The Roman aqueduct survived millennia because it was overbuilt and underlevered. Our data aqueduct is optimized for quarterly return on capital.
The remedy is boring and hard. Treat long-term purchase commitments as debt. Map collateral chains and cap rehypothecation. Align funding tenors with asset lives instead of praying for rollover. Demand plain-language disclosure of circular agreements, including embedded options and exit clauses. Stress test data center landlords and hyperscaler counterparties for a 200 basis point move in term premia plus a 20 percent equity drawdown. Stop counting vendor-financed demand as organic growth. Recognize that energy infrastructure is the binding constraint and price projects accordingly. On the policy side, keep monetary and fiscal lines clean; credibility is cheaper than rescue. On the market structure side, assume that a runnable, global, always-on payments layer will be tested at peak stress. In probabilistic terms, the tails are fat because the networks are tight. Robust beats elegant. Cash flow beats narratives. The system will not fail because machines think; it will fail, if at all, because we still finance like optimists.
