Central Bank AI: Who Writes the Rulebook Wins

The precedent: Herstatt, 1974

In June 1974, Bankhaus Herstatt failed.

German regulators closed the bank at day's end in Frankfurt. Correspondent banks in New York had already released Deutsche Marks, and the dollar leg they expected vanished into the gap between time zones.

The mechanism was plain: a settlement timing mismatch riding on a shared payment rail. The outcome arrived within months. The G10 central bank governors founded the Basel Committee on Banking Supervision that same year, and the architecture of modern financial regulation began there.

Central banks absorbed one lesson from the wreckage. Shared infrastructure manufactures shared fragility. In 2026, that structure has returned in a new dress, and the rail this time is compute.

Central bank AI reaches the agenda

The shift became explicit in Singapore.

At the 31st Meeting of Governors of the East Asia and Pacific Central Bank Group, held on July 23, 2026, senior officials elevated artificial intelligence from a technology afterthought to a matter of financial stability. The tone was vigilance, adaptability, and coordination.

Their questions were precise. Who answers for an algorithmic decision, and what safeguards the payment system requires when models allocate capital at machine speed.

The choice of Singapore matters. The city-state has spent a decade building itself into a regulatory laboratory, and hosting the governors places it at the pen.

These are governance questions, rather than engineering ones. The officials grasp a hard truth: the code will outrun the rulebook, and the drafting of principles has become the decisive contest. The world watches model capability; the durable power sits in the rules.

The mechanism: concentration is the risk

AI infrastructure sits in few hands.

A small cluster of technology corporations controls the compute, and financial institutions lean harder each quarter on their cloud. That dependency concentrates operational risk at a layer that supervisors barely reach, and the exposure grows quietly.

Algorithm-driven trading compounds the problem. When many desks run correlated models on the same three providers, a single technical incident propagates across the system at a speed human oversight fails to match.

The interconnection that raises efficiency also widens the blast radius. This is the Herstatt logic, updated. The failure point has migrated from settlement timing to compute dependency, and the cascade travels faster than in 1974 by orders of magnitude.

What AI changes inside the central bank

The technology cuts both ways.

Inside monetary policy, AI processes vast information in near real time. It can detect shifts in inflation pressure, consumer behavior, the labor market, and activity faster than quarterly cycles allow. That speed reshapes how a central bank reads its own economy.

On the supervision side, analytical tools flag anomalies early. They trace complex links between institutions, surface fraud and money laundering, and read declining liquidity or rising credit risk before it metastasizes.

The same tooling that sharpens oversight hands the supervised institutions equal power. A regulator armed with pattern detection faces counterparties armed with the identical models. The arms race runs on both sides of the table, and the governance question follows from that symmetry.

The semiconductor layer underneath

Governance rides on hardware.

My second standing thesis holds here: the US-China contest over AI turns on access to semiconductors, rather than the elegance of any model. Whoever governs the fabs and the compute governs the outcome. Models replicate; fabrication plants resist replication.

Central banks have begun to read this map. A stability framework that ignores where the chips and the cloud physically reside is a framework built on sand.

By my own reserve tracking, the dollar's share of global reserves slid from 71.4 percent in 2001 toward 58 percent in 2024. The same lesson applies. Power concentrates in the layer few people watch, and it drains from the layer everyone watches.

My position, and what would reverse it

Here is the thesis, stated plainly.

The central banks that codify central bank AI governance first will export their standard globally, the way Basel exported capital rules after 1988. The jurisdiction that writes the template acquires durable authority over cross-border finance.

Asia is moving ahead of the Atlantic. EMEAP's coordination signals intent, and Singapore has positioned itself as the drafting room for the region.

The evidence that would overturn my view is concrete: a binding EU or US framework enacted before any Asian equivalent, carrying extraterritorial reach comparable to GDPR. Absent that, the center of gravity shifts east. I weight this as structural, playing out across a decade, rather than a quarter.

Three implications for the capital

The consequences are allocatable today.

Family offices and sovereign funds: treat cloud and compute concentration as a portfolio risk factor across the next 36 months. Exposure to the few providers governing AI infrastructure carries a regulatory tail risk that standard equity screens miss. Diversification across providers becomes a fiduciary question, rather than a procurement one.

Chief risk officers: a correlated AI-provider outage sits outside most VAR frameworks, and the Herstatt precedent shows shared-rail failures cascade. Add the scenario to the model set this year.

CFOs and boards: the narrative of AI as pure margin expansion will age poorly once supervisors impose governance costs. Reprice that assumption across an 18-month horizon, before investors do it for you.

The prediction

One claim, dated and verifiable.

An EMEAP member central bank, led by the Monetary Authority of Singapore, will publish a formal AI governance framework or supervisory guideline for financial institutions before the close of 2027. That instrument precedes any comparably binding G7 rule.

Confidence: 68. Horizon: 365 days. Verification: a published supervisory document that assigns named accountability for algorithmic decisions.

What to watch

Three leading indicators will confirm or break the thesis. Watch for a Singapore or Hong Kong consultation paper on model risk, a BIS Innovation Hub project on AI supervision, and any G7 statement that lags the Asian timeline. Read the full record of the Singapore meeting and the broader strategy desk analysis.

The divergence between technology speed and regulatory speed resolves in every documented case. The question is who writes the resolution.