The Bank for International Settlements has built a sustained body of analysis over the past two years identifying artificial intelligence as a source of both structural opportunity and measurable financial stability risk — and the weight of its most recent work tilts toward caution.

The BIS's December 2025 quarterly report put high AI valuations alongside resilient risky assets as part of a cluster of mounting financial stability challenges. That framing — valuations as a stability variable, not merely an equity market curiosity — reflects how far the conversation has moved since the June 2024 Annual Economic Report, which opened with the broader question of AI's opportunities and risks for the macroeconomy and central banks. The institution has progressively tightened its focus from the general to the specific: from growth and productivity dynamics to concentration risk, leverage, and oversight gaps.

Concentration and the Earnings Question

One of the sharper findings in the BIS's November 2025 analysis is that AI-related equity activity is disproportionately concentrated among established firms with actual earnings, rather than spread across speculative early-stage names. That matters structurally. It means the AI trade is not, at least by that measure, a classic bubble in the sense of capital flooding into entities with no cash flows. But concentration itself carries tail risk: a re-rating of a handful of mega-cap names — triggered by earnings disappointment, regulatory action, or a competitor shock — can transmit rapidly across passive vehicles and correlated portfolios.

The December 2025 quarterly review noted that tech stock prices began recovering in January 2025, with renewed appetite linked partly to domestic AI developments following the DeepSeek release. DeepSeek's emergence rattled consensus assumptions about the compute intensity required for frontier model training, briefly destabilizing the capex narratives that had underpinned Nvidia and hyperscaler valuations. That recovery, while it resolved near-term price dislocations, did not resolve the underlying valuation question the BIS had flagged.

The Oversight Gap

The more operationally urgent thread in the BIS's work concerns financial services specifically. A June 2025 Financial Stability Institute report on the financial stability implications of AI concluded that AI use in financial services without appropriate controls and oversight could amplify certain financial vulnerabilities. The FSI's framing is precise: this is not a claim that AI is inherently destabilising, but that the absence of controls is the risk vector. That distinction matters for how institutions calibrate their governance response.

The channels of concern are familiar from earlier technology adoption cycles — procyclicality, herding, model opacity, third-party concentration — but AI introduces them at greater speed and scale. Algorithmic trading strategies informed by large language models, credit underwriting driven by models that cannot be fully interrogated, and fraud-detection systems that share common architectures across competing institutions all create correlated failure modes that traditional microprudential frameworks were not designed to catch.

The Financial Stability Board moved in the same direction. Reuters reported in October 2025 that global financial watchdogs had ramped up monitoring of AI, with the FSB among those expanding their surveillance of AI-related risks in the financial system. Regulatory bandwidth is finite; the question facing supervisors is whether monitoring translates into enforceable standards before the exposures compound.

What Central Banks Are Watching

The 2024 Annual Economic Report chapter on AI and the economy, published by the BIS in June 2024, framed the monetary policy dimension: AI-driven productivity shifts could alter the neutral rate, change the transmission of policy through labour and credit markets, and complicate inflation forecasting. Those are slow-moving variables. The financial stability concerns catalogued in the December 2025 and June 2025 reports operate on much shorter horizons.

The practical synthesis is that central banks and supervisors face a two-speed problem. Structural macroeconomic effects from AI diffusion will take years to manifest clearly in data. Meanwhile, the financial system is already accumulating AI-related exposures — in equity portfolios, in model-driven lending, in the operational dependencies of systemically important institutions — that require oversight responses now. The BIS's accumulated output maps both timescales, but the recent trajectory of its publications signals where the institution sees the more pressing near-term risk.

Whether existing regulatory frameworks can adapt quickly enough is a separate question the BIS does not answer — and which no one can answer with confidence yet.