The rising threat AI poses to financial stability
Debating whether AI stocks are overheated misses the bigger risk: system stability. The first wave of AI buildout has been funded mainly by equity and the cash flows of the mega platforms. The next wave won't be. Debt is piling up across chips, data centres, power, and the broader supply chain.
For finance teams, this isn't abstract. It's concentration risk, collateral risk, duration risk, and liquidity risk-arriving at the same time, tied to the same theme.
Where the leverage is building
- Data centre expansion: project finance, REIT balance sheets, private credit, and construction loans secured by highly specialized assets.
- Power infrastructure: long-dated PPAs, grid interconnection bets, merchant exposure to volatile power prices, and counterparty risk.
- Chips and hardware: vendor financing, inventory financing, and equipment loans backed by fast-obsolescing GPUs.
- Startups and model builders: venture debt, convertibles, revenue-share structures, and dependency on hyperscaler credits.
- Public markets: margin loans against AI-heavy portfolios; structured notes tied to a narrow set of tech names.
Why this cycle is different
- Concentration: a few platforms control demand, supply, and distribution. Single-counterparty exposure shows up everywhere.
- Specialized collateral: stranded-asset risk if demand softens or architectures shift.
- Long duration, short tech cycles: capex lives for decades; product cycles turn in quarters.
- Power dependency: grid delays and price spikes can break pro formas, even with PPAs.
- Shadow credit: private markets are carrying large, cov-lite exposures the public market doesn't price daily.
How stress could spread
- Demand undershoots or pricing compresses as competition intensifies; unit economics narrow for compute-heavy models.
- Data centre REIT re-ratings trigger covenant pressure and higher refinancing costs.
- PPA counterparties wobble; merchant exposures face price volatility or curtailment.
- GPU resale values drop; recoveries on equipment-backed loans disappoint.
- Private credit NAV marks lag; LP redemptions tighten funding; banks pull back on unfunded commitments.
- Equity drawdowns force deleveraging of margin loans and structured products.
What risk teams should do now
- Map exposure: look through facilities, funds, and suppliers to quantify AI-linked concentration (by counterparty, region, asset type).
- Underwrite with shorter half-lives: lower LTVs on GPUs and specialized plant; faster amortization; mandatory upgrade reserves.
- Stress PPAs: test 20-40 percent power price swings, curtailment, interconnection delays, and counterparty downgrades.
- Tighten covenants: cash sweep triggers, capex gates tied to pre-sold capacity, step-ups for delayed energization.
- Limit single-name exposure: set aggregate caps for hyperscalers and for top 10 correlated credits.
- Reprice liquidity: dynamic margining on AI-heavy collateral; higher haircuts for correlated tech baskets.
- Operational resilience: reduce cloud concentration; maintain tested exit plans and viable multi-cloud footprints.
- Model risk management: validate AI-driven credit and market models; track drift and data lineage.
Scenarios worth running
- AI equity correction of 30-50 percent; tech credit spreads +200-300 bps; REITs -25 percent; equity-linked margin calls.
- Power shock: +30 percent regional price jump, 9-12 month interconnection delay, and partial curtailment for 2 quarters.
- Tech shift: new architecture reduces value of prior-gen GPUs by 40 percent in secondary markets.
- Policy shock: export controls or sanctions disrupt supply; lead times extend; capex schedules slip.
What corporate finance should change in deal terms
- Capacity-first, not logo-first: sign customers to firm commitments before locking capex and long PPAs.
- Flex leases and PPAs: embed expansion/contraction rights, indexed pricing bands, and termination provisions tied to energization milestones.
- Hurdle rates that reflect grid reality: include delay penalties, curtailment probabilities, and maintenance capex for upgrades.
- Avoid long, exclusivity-heavy cloud contracts; keep optionality to pivot providers or architectures.
Early warning indicators
- Hyperscaler capex guidance vs. revenue trajectory and gross margin trends.
- Vacancy and pre-lease rates for new data centre capacity; interconnection queue times.
- PPA pricing spreads and rising collateral requirements from utilities.
- Private credit inflows/outflows to digital infrastructure and software borrowers.
- Secondary prices for prior-gen GPUs and liquidation outcomes on failed operators.
Regulatory context worth tracking
Supervisors have flagged AI and big-tech concentration as a financial stability issue. Use their frameworks to justify tighter internal limits and more conservative collateral treatment.
Bottom line
AI may keep growing. That's not the question. The question is whether funding is priced for missteps in demand, power, and tech cycles. Treat AI like any concentrated theme: cap exposures, shorten risk, and make recovery math your first slide.
Useful resource
If you're inventorying exposure to vendors and use-cases, a practical starting point is an overview of tools by job function: AI tools for finance. Use it to cross-check where your teams depend on third-party providers and where single points of failure may exist.
Your membership also unlocks:
