AI boom poses new financial stability risks, BIS head says

BIS chief Pablo Hernandez de Cos warned that the AI investment boom is creating financial-stability risks as the five largest technology firms are expected to spend more than $1 trillion on AI in 2025-26 and global AI investment could reach $4 trillion by 2030. He said increasingly debt- and private-credit-funded AI infrastructure investment is opaque and interconnected, leaving markets vulnerable if commercial returns fail to meet lofty expectations. AI could lift total factor productivity growth by roughly 0.5 percentage point annually, but job displacement risks are emerging in customer service, programming and administrative work, while AI supply-chain economies including South Korea, Singapore, Malaysia and Taiwan have benefited from stronger chip and equipment export prices.

Analysis

The investable implication is not a broad AI de-rating, but a widening split between cash-funded platform owners and leveraged infrastructure builders. MSFT, GOOGL, META and AMZN can absorb elevated capex from operating cash flow; suppliers and developers reliant on project debt or private credit face a materially higher hurdle if utilization, GPU pricing or enterprise AI monetization disappoints. That argues for preferring “toll collectors” with contracted demand—VRT, ETN, PWR, DLR and EQIX—over lower-visibility compute leasing and unlisted AI infrastructure credits.

Over the next 1-3 months, financing disclosures are the key catalyst rather than model-performance announcements. Any earnings call showing capex funded by incremental debt, weaker free-cash-flow conversion, lengthening customer-payment terms, or a rise in data-center project cancellations should trigger multiple compression across the AI infrastructure complex. Private-credit stress would initially be most visible in BDC marks and direct-lending spreads—ARCC, ARES and BX are useful liquid read-throughs—but public lenders are not necessarily shorts absent evidence of concentrated exposure.

The consensus risk is focused on an AI-equity bubble, while the underappreciated transmission channel is higher real rates: sustained infrastructure spending can keep power, equipment and construction demand firm even as software-driven productivity lowers unit labor costs. That mix complicates a clean disinflation narrative and favors companies with pricing power in electrical equipment and grid buildout. Over 6-18 months, the larger structural winner may be power availability rather than semiconductor supply: utilities and generation providers serving constrained data-center regions can capture scarcity rents, while hyperscalers face margin pressure if power contracts reset higher.

A bearish AI thesis is falsified if enterprise software vendors demonstrate recurring AI revenue that offsets inference costs and hyperscalers maintain capex growth without deterioration in aggregate free cash flow. Conversely, a meaningful widening in private-credit defaults, an upward revision to data-center power costs, or reduced 2027 capex guidance would validate the relative-value short leg.

Key Decisions for Investors

• Initiate a 3-6 month pair: long VRT and ETN / short an equal-dollar basket of high-duration AI software via IGV. The pair isolates physical-infrastructure bottlenecks from broad AI enthusiasm; reassess if VRT/ETN backlog growth decelerates materially or IGV earnings revisions accelerate.
• Maintain long TSM and MU versus a short SMH hedge only after checking current valuation dispersion. TSM’s leading-edge manufacturing and MU’s memory-cycle exposure are more directly tied to AI unit demand, but use the hedge if semiconductor-capex expectations have already rerated sharply; exit on evidence of hyperscaler order deferrals.
• Add DLR or EQIX on financing-driven weakness rather than chase momentum. Their contracted interconnection and colocation economics offer cleaner exposure than speculative GPU capacity, but require confirmation that incremental development yields remain above funding costs; a sustained rise in debt costs without rent repricing invalidates the trade.
• Create an alert for ARCC, ARES and BX quarterly disclosures: rising non-accruals, NAV markdowns in technology/infrastructure loans, or weaker fee-related earnings would be a signal to reduce AI-infrastructure beta and consider a tactical short in the most levered public data-center or power-development names.
• Avoid treating broad AI exposure as a single macro position for the next quarter. Size longs around cash-flow durability and contracted backlog, while retaining index hedges ahead of hyperscaler earnings, where capex guidance—not headline AI adoption claims—will determine whether valuation support persists.