AI Risk to Financial Stability: What the Fed Sees15 min read3,008 words

AI Risk to Financial Stability: What the Fed Sees

The Fed's survey puts AI ahead of other financial risks. See why valuations, cyber exposure and compute debt matter—and what leaders should do next.

AI financial stability riskFederal Reserve AI surveyAI infrastructure financingAI operational riskAI governance
AI Risk to Financial Stability: What the Fed Sees

AI Is Now a Financial Stability Risk. Here's Why

Artificial intelligence has stopped feeling like a niche technology story and started showing up in conversations about what could rattle the financial system. In the Federal Reserve's latest survey of market contacts, 87% cited AI as a salient risk - up from 50% in the spring. That doesn't mean a crisis is around the corner. It does mean financial professionals are treating AI as something they need to plan for, not something they can safely ignore.

The anxiety isn't only that some companies may overestimate what AI can do. It's that AI now links together several parts of the modern financial machine: stock valuations, cybersecurity, data centers, chip supply, and debt markets. And we've already seen how quickly an incident involving AI - whether or not the model itself is the culprit - can turn into an operational problem, a reputational headache, or both.

This article unpacks what the Fed's numbers do (and don't) tell us, how AI risk can move through the financial system, and what the FAA incident and Anthropic's reported compute commitments suggest about real-world stakes. It ends with a practical way to assess exposure without treating every AI rollout as either harmless or apocalyptic.

What does the Fed's AI risk survey actually tell us?

The headline number is hard to miss: 87% of the Fed's market contacts cited AI as a salient risk, up from 50% in the spring survey. But the result needs a careful read. This is a survey of market contacts, not a prediction that AI will trigger a financial crisis - and not a direct measure of how likely any specific failure is.

What it captures most clearly is attention. More of the people the Fed talks to now see AI as relevant to financial stability. That shift matters because it can change how institutions think about exposures, how lenders price risk, and how regulators frame the questions they ask.

Three risks behind one headline

The Federal Reserve's Financial Stability Report points to a few different channels.

First, high stock valuations can leave markets vulnerable if expectations about AI-driven growth get ahead of reality. When a large slice of a company's valuation depends on future gains, disappointing results can lead to a fast repricing. AI isn't the only thing that can do this - but the scale of investment and the amount of attention can make the swings feel bigger.

Second, the report highlights cyberattacks, including the concern that frontier AI models could help increase damage. AI may help defenders spot vulnerabilities and respond faster. It can also make some attacks more efficient or speed up how quickly weaknesses are exploited. The outcome depends on who has access, what safeguards exist, and how quickly defenders improve.

Third, the Fed flags "complex forms of AI-related financing." Translation: AI ambitions often rely on a chain of companies - technology firms, data-center operators, chip suppliers, lenders, and investors. If one part of that chain breaks, the impact may not stay neatly contained in the company that started the plan.

Attention is not the same as a prediction

Treat the survey as a prompt to look closer, not a verdict. Risk perception can rise before something happens, and a high survey share doesn't prove AI is more likely than every other threat. Still, it signals that AI has become "material" enough - at least in the minds of market contacts - to sit alongside familiar financial vulnerabilities.

For executives, the useful question usually isn't "Is AI a bubble?" That invites an either/or answer when the evidence is messy. A better, more operational question is: Which assumptions about AI growth, infrastructure, and security are already built into our plans and our counterparties? That way, the survey becomes actionable without pretending it can forecast the next market shock.

How can AI risk spread through the financial system?

AI becomes a financial stability issue when exposure to AI is big, interconnected, and hard to unwind. One company running a model experiment is unlikely to matter to the system. The story changes when many firms make similar bets, finance them through overlapping channels, and depend on the same limited infrastructure.

Valuations can turn expectations into exposure

Investors price companies based in part on expected future earnings. If AI is expected to deliver major productivity gains, that optimism can support high valuations for software firms, chipmakers, and infrastructure providers. But expectations aren't the same thing as cash flow. If adoption takes longer than planned, customers push back on costs, or the productivity gains aren't as large, investors can revise their assumptions quickly.

That doesn't automatically make high valuations irrational. It does mean the margin for error can shrink - especially when the business case depends on rapid adoption, sustained demand, and dependable access to computing capacity all at once. A company can make sensible technology decisions and still feel financial pressure if the market has priced in a rosier outcome than what actually arrives.

Data centers and chips require major, long-lived commitments. Those investments often happen before demand is fully visible. That creates a chain of exposure: model developers need compute; cloud providers and data-center operators finance capacity; chip suppliers rely on continued orders; and lenders and bond investors fund parts of the expansion.

JPMorgan, according to the reporting, expected US investment-grade bond sales to reach a record $2.1 trillion in 2027, with AI among the biggest drivers. The estimate included about $225 billion from hyperscalers and another $225 billion for data centers and chips. Those numbers are projections, not proof of stress. But they show why the Fed is paying attention not just to model performance - also to financing structures and concentration.

Correlation matters more than a single failure

The system-level question is whether multiple firms are vulnerable to the same shock. If several companies depend on similar financing, suppliers, or demand assumptions, one change - higher borrowing costs, a supply disruption, or weaker customer uptake - can hit many balance sheets at once.

That's why "AI exposure" shouldn't be treated like one neat number. A bank's loan to a data-center operator, a technology company's long-term cloud commitment, and an investor's position in an AI-heavy stock aren't the same risk. They run on different timelines and fail in different ways. The practical job is to map where exposures overlap, who would absorb losses, and how quickly obligations could be reduced.

What do the FAA and OpenAI incidents show about operational risk?

AI risk isn't only about markets or model outputs. It also includes the surrounding systems - people, procedures, and controls - used when AI is deployed. Two examples in the underlying reporting make that distinction clear: a nationwide FAA ground stop after an AI tool was introduced at regional control centers, and a video call in which OpenAI national security staff spoke with pranksters impersonating Ukrainian officials.

The FAA case is about attribution and escalation

The FAA halted flights nationwide days after its SMART AI system went live across regional air route traffic control centers. Transportation Secretary Sean Duffy attributed the ground stop to a communication issue affecting all centers. DHS Secretary Markwayne Mullin said officials were looking into the system. The system's builder, Air Space Intelligence, said the stop was not caused by SMART.

Even so, the causal question remains unresolved. The timing alone doesn't prove the AI tool caused the disruption, and the vendor's denial doesn't automatically settle the matter. That difference matters. If blame gets assigned before an investigation is complete, it can blur the real failure - whether it's technical, procedural, or tied to communications between systems.

Still, the incident offers a clear operational lesson. When a new tool goes into a critical environment, an outage can trigger questions about the tool even if it isn't the cause. Teams need a straightforward account of what the system can influence, what it can't, how humans review its outputs and recommendations, and what happens operationally when the tool is unavailable. In safety-critical settings, those details aren't "later paperwork" - they're part of the deployment.

The OpenAI call shows that people are part of the attack surface

The Intercept reported that Russian pranksters Vovan and Lexus posed as Rustem Umerov, a Ukrainian intelligence official, and spoke with three OpenAI national security staff for 17 minutes. The conversation concerned OpenAI's military support for Ukraine and potential offensive capabilities. OpenAI described it as a prank and said it took steps to prevent it from happening again.

This wasn't a model failure. It was an identity and verification breakdown involving people, sensitive discussions, and a convincing impersonation. That's relevant to any business using AI in high-trust settings. Even a strong technical stack can't compensate for weak checks on who's making a request or what information can be shared on an unverified call.

The response isn't to ban sensitive conversations or assume everyone is hostile. It's to set proportionate verification for sensitive topics, define what staff can disclose, and make it easy to pause and escalate unusual requests. The same controls also help with executive impersonation, supplier fraud, and social engineering that isn't really "about AI" at all.

Why are AI compute commitments becoming a business risk?

Frontier AI depends on access to large amounts of computing capacity. That makes planning tricky. Reserve too much and you carry expensive capacity you can't use. Reserve too little and you may not be able to serve customers or train models when demand spikes.

Anthropic's reported IPO filing makes the tension concrete. It lists $900.9 billion in compute commitments running through 2045. At the time of the reporting, the company was due to price its IPO on Wednesday and begin trading on Nasdaq on Thursday under the ticker ANT. The specifics here should be understood in that reporting context, not as a general template for every AI company.

Long-term capacity is both an asset and an obligation

A big commitment can secure access to infrastructure that's hard to obtain on short notice. For a fast-growing model provider, that can support reliability and future expansion. But long-dated commitments can also reduce flexibility. If demand drops, model efficiency improves, or strategy shifts, the company may still owe money for capacity that no longer fits its needs.

Anthropic CEO Dario Amodei described compute as the company's biggest business dilemma in a conversation with JPMorgan CEO Jamie Dimon. His reported approach was to buy on the lower end up front and add capacity in the spot market when available. That can reduce overcommitment - but it depends on spot capacity still being available when you need it. A strategy that works in one market cycle can fail in the next.

Financing risk can sit outside the AI company

The financing chain doesn't stop with model developers. Data centers, chips, and power infrastructure require capital, and some of that money comes through borrowing. That means AI investment can affect lenders and bond markets even if the model companies themselves are private or well funded.

The question for management isn't just how much a company wants to spend. It's how much of that spending is fixed, what assumptions make the plan viable, and who else depends on those assumptions. Procurement might focus on the cloud contract; finance might focus on the long-term obligation; risk teams might focus on concentration and continuity exposure. Those perspectives have to line up before a commitment is made - not after forecasts change.

The reported value of Anthropic's early investment from Alameda Research is another reminder that technology financing can have a long tail. The Financial Times reported that the $500 million investment made in 2021 would have been worth about $37 billion at the expected IPO price, while FTX's bankruptcy estate sold it in 2024 for about $1.3 billion. That comparison isn't a prediction of future returns. It shows how ownership, timing, and an investor's financial condition can determine who ultimately captures value.

How should companies manage AI risk in practice?

A useful AI risk process connects the model to the business system around it. It should cover the use case, data, people, suppliers, fallback options, and financial commitments. Checking model accuracy on its own won't tell you whether your organization can keep operating if a provider changes terms or a service goes dark.

Use a five-part exposure review

For each material AI deployment or investment, document five things:

  1. Business dependency: What process, revenue stream, or customer promise relies on the system? Separate an optional productivity tool from something whose outage stops core operations.
  2. Failure modes: What can go wrong technically, operationally, or through misuse? Include incorrect outputs, unavailable services, compromised credentials, data exposure, and changes in supplier access.
  3. Human control: Who reviews consequential outputs, who can override the system, and who is accountable when the system is wrong? A human in the loop is only useful if that person has the information and authority to act.
  4. Concentration: Which providers, models, chips, cloud regions, or financing partners are shared across critical processes? Even if products differ, they may rely on the same underlying infrastructure.
  5. Exit and recovery: How quickly can the organization switch providers, revert to a manual process, or restore service? Spell out the cost and time required instead of assuming an alternative is readily available.

This review should lead to decisions, not just a risk register. For example: if a customer-support assistant has low business impact and a workable manual fallback, monitoring may be enough. If an AI system shapes decisions in a regulated process or a critical operation, you'll likely need stricter validation, staged rollout, audit trails, and recovery procedures that have actually been tested.

Match controls to consequence

Controls should fit the consequences of failure. A low-stakes internal summarizer shouldn't require the same approvals as software that affects access to credit, or an operational tool used in aviation. Over-applying heavy controls everywhere wastes time and can push teams toward unapproved workarounds. Under-applying controls everywhere leaves serious exposures unmanaged.

A practical baseline includes a named business owner, a record of data and model dependencies, clear limits on sensitive uses, a process for incident reporting, and a tested fallback plan. For external providers, look closely at contract terms covering service availability, data usage, change notifications, and termination. And for systems with material consequences, test not only whether the model performs well - also how the workflow behaves when outputs are wrong or the service is unavailable.

Treat incidents as feedback, not proof of a single theory

When something goes wrong, separate what you know from what you're assuming. Record the timeline, systems involved, human decisions, provider communications, and any evidence about causal links. Don't assume AI caused the incident just because it was recently deployed. Also don't dismiss the possibility just because a vendor says it's unrelated.

The point is to improve the system and preserve trust. That might mean changing tools, revising procedures, adding verification, reducing dependency, or concluding that AI isn't the right fit for the job. A mature risk process should be able to accommodate all of those outcomes.

Key takeaways

  • The Fed survey reflects rising concern among market contacts; it doesn't predict an AI-driven financial crisis.
  • The main systemic channels include high valuations, cyber exposure, and financing links between AI firms and infrastructure providers.
  • The FAA ground stop's cause was unresolved in the reporting. Timing raises questions, but it doesn't establish causation.
  • The OpenAI impersonation incident highlights why identity verification and staff procedures still matter.
  • Long-term compute commitments can support growth while limiting flexibility and shifting risk through debt markets.
  • Companies should map dependencies, concentration, human control, and recovery options before scaling AI systems with real operational impact.

Conclusion

The Fed's survey is best read as a change in the risk landscape - not as a claim that AI will trigger the next financial crisis. AI is now woven into market valuations, cyber operations, and a fast-growing infrastructure financing chain. That combination makes it both a business and financial issue, not just a technical one.

The answer isn't to treat every AI deployment as automatically dangerous or to assume rapid adoption automatically justifies every investment. Map dependencies, understand fixed commitments, verify who is making sensitive requests, and test how operations continue when a model or provider fails. Those are the practical controls that should hold up whether AI growth speeds up, slows down, or takes a different turn than expected.

Frequently Asked Questions

In the latest survey described in the reporting, 87% of the Fed’s market contacts cited AI as a salient risk, up from 50% in the spring. Concerns include high valuations, cyberattacks and complex financing tied to AI infrastructure. The result reflects growing attention, not a prediction of a crisis.
No. The FAA halted flights after a communication issue affected air route traffic control centers, but the reporting did not establish that SMART caused the issue. Its builder said the stop was not a result of its system. The timing makes the incident worth investigating, but timing alone is not proof.
Compute commitments can secure capacity for growth, but long-term obligations are difficult to reduce if demand falls or technology changes. Because data centers and chips also rely on substantial financing, risks can extend to lenders, suppliers and bond investors—not just the AI company making the commitment.
Map the business dependency, likely failure modes, human oversight, provider concentration and recovery options. Then match controls to the consequences of failure. A critical workflow needs tested fallbacks and clear accountability; a low-stakes tool may need lighter controls and routine monitoring.

Related articles