Who Holds Correlated AI Losses? Concentration and Risk Allocation in the AI Supply Chain
Abstract
Firms that use artificial intelligence buy their models from very few providers, so choices of provider that each firm makes on its own tie the liability risk of many firms to a few common shocks. I study how this structure of the supplying industry decides who bears the losses. However many policies an insurer writes, the AI component of its book carries the systematic variance of only 1/H independent risks, where H is the Herfindahl index of provider shares. Diversified capital could carry this accumulation cheaply if slices of it could be traded, but liability wordings do not identify the provider behind a claim, and insurers have begun to exclude AI as such. An excluded loss stays with the firm that deployed the model and, if the firm defaults, passes to its bank. I then measure what supervisors do not observe. Of 313 firms whose annual reports name a provider, 89 say that their products run on its models, 73 of them on OpenAI’s; 22 of these firms and four AI-infrastructure borrowers owe USD 78 billion in syndicated loans, held to a large extent by the same banks. When a provider has an incident, its clients’ shares fall by 1.4 percentage points over two days relative to the clients of its competitors. Their loans carry no spread premium and their bonds do not move. Contracts filed with the SEC show how the factor is allocated: providers are protected against their suppliers’ failure, deployers report neither recourse nor cover, and the suppliers’ lenders hold the provider’s credit under covenants tied to its rating. For 104 European banks, a defect in the largest provider’s model that harms a tenth of its clients would cost the median bank 0.3% of its CET1 capital, or 0.7% if its AI-using borrowers all rely on that provider. Equity prices register the provider factor; loan and bond prices, credit files and capital rules do not. I propose recording provider dependence in credit and underwriting files and running one provider-failure scenario across insurance and banking supervision.
- Keywords
- artificial intelligence; AI adoption; foundation models; market concentration; accumulation risk; liability insurance; exclusions; syndicated loans; event study; financial stability
- JEL
- G14, G21, G22, G28, K13, L86, O33
- Version
- November 2026
Cite as: Aslan, Ilker (2026). Who Holds Correlated AI Losses? Concentration and Risk Allocation in the AI Supply Chain. Working paper, November 2026. https://ilkeraslan.github.io/papers/who-holds-correlated-ai-losses.html