Central Bank AI: Whoever Controls Compute Wins

A precedent: 1973, the birth of SWIFT

In 1973, 239 banks from fifteen countries founded SWIFT in Brussels. The mechanism was elementary: standardise payment messages between institutions. Whoever controlled the channel controlled the flow.

For five decades that channel appeared neutral. Reality operated on another plane.

In 2012 the United States pushed to exclude Iranian banks from the system. In 2022 they repeated the move against part of Russia's banks. The structural lesson remains identical: the technical infrastructure of money is geopolitical power, and every new layer of that infrastructure redefines who commands.

The precedent deserves a finer reading. SWIFT was born as a technical cooperative, not as an instrument of foreign policy. Its neutrality was the premise of its universal adoption. But a channel used by everyone becomes, by definition, a point of control. The moment exclusion from SWIFT produces a real cost for the counterparty, the channel's operator acquires a power that no treaty had assigned it. It is the mechanism that repeats at every new layer of monetary infrastructure: first adoption for efficiency, then the discovery that efficiency has a master.

The current pattern: central banks embrace AI

The Bank for International Settlements launched its Innovation Hub in 2019, with centres in Basel, Hong Kong and Singapore. The stated aim was to test technologies for the global financial system.

Today the use cases of central bank AI are concrete and documented.

• Nowcasting of growth and inflation
• Automated banking supervision
• Textual analysis of supervisory documents
• Detection of anomalies in cross-border payments

The Federal Reserve, the ECB and the People's Bank of China are strengthening internal analytical capabilities. The topic moved from research to production in less than five years. This pace has a precise precedent: the digitisation of payments in the 1990s followed the same curve, and whoever governed it first dictated the standards.

It is worth clarifying what distinguishes these four use cases from a mere technological upgrade. Nowcasting anticipates rate decisions. Automated supervision shifts the judgment on solvency from the official to the model. Textual analysis filters what supervisors read. The detection of anomalies in cross-border payments bears directly on the boundary between lawful flow and blocked flow. In each case, the model does not accompany the decision: it structures it. Whoever calibrates the model orients the outcome before the outcome is decided.

The mechanism: whoever controls compute controls money

The power of AI applied to monetary policy depends on three physical resources: data, semiconductors, energy. Models replicate. Chip fabs, on the other hand, remain concentrated in very few hands.

This is my second founding position applied to money. The US-China competition on AI is a problem of access to fabs, beyond the capability of the models. Whoever controls TSMC and Samsung controls the outcome.

The causal mechanism is linear. A central bank that manages nowcasting, surveillance and liquidity through models becomes dependent on the compute that runs them. That compute comes from a small number of foreign producers. Monetary sovereignty thus shifts towards whoever dominates the semiconductor chain. The market has already priced AI as an equity theme. The market has yet to price AI as a vector of monetary power.

The distinction between the two pricings is the point. To price AI as an equity theme means discounting the revenues, margins and growth of the producers. To price AI as a vector of monetary power means discounting the possibility that a central bank's autonomy depends on a foreign supply chain. The first is already in the multiples. The second appears in no sovereign risk model. A dependence on compute is a political dependence disguised as a technical choice: the central bank chooses the model for efficiency, but inherits the geography of whoever manufactures the hardware. That geography is not diversified.

Energy is the hidden constraint

AI devours energy and hardware. Every model requires compute, and compute requires stable power and efficient storage.

The evolution of batteries signals the direction of the chain. On 11 August 2026 ZDNET reviewed a $49.99 semi-solid-state power bank, a marker of an energy density advancing across the entire sector (source).

The lesson holds on an industrial scale. Whoever controls the fabs, the power grids and energy storage controls the monetary output of the AI era. A central bank lacking sovereign energy and accessible chips effectively delegates a part of its own autonomy. The energy constraint will remain the decisive factor of the coming decade.

The energy constraint is hidden because it operates upstream of every other variable. A model can be trained elsewhere, data can be purchased, a chip can be imported. The energy to run inference at continuous scale, on the other hand, must be produced and distributed within the border. A saturated grid, or one dependent on foreign supplies, adds a second layer of fragility on top of the semiconductor one. Whoever controls neither the chips nor the energy does not control their own compute, and whoever does not control their own compute does not govern the analytical instruments on which they base monetary policy.

My position, and what would refute it

My thesis is clear-cut. The central banks adopting AI are shifting monetary power towards whoever controls semiconductors and energy, beyond the models themselves.

This surpasses the ordinary technological cycle. It is a regime change in monetary architecture.

The reasoning rests on a structural fact I have followed for years: according to IMF COFER data, the dollar's share of global currency reserves fell from 71% in 2001 to 58% in 2024. A fragmentation of monetary infrastructure is already underway. AI accelerates it, because it adds a layer of technological dependence on top of the financial one. What would change my reading? A real diversification of fabs outside Taiwan and Korea by 2028, with documented operational capacity in Europe and the United States. In that case the constraint loosens and the thesis must be revised.

The limits of the evidence must be declared. The decline from 71% to 58% is a datum on currency reserves, not a direct measure of dependence on compute: what links the two phenomena is a line of reasoning, not a historical series. The thesis remains conditional on the current concentration of foundries, a fact that decisive industrial policy can modify within a handful of years. For this reason the condition of refutation is precise and dated: without a horizon and a verifiable indicator, the position would be an opinion, not a prediction.

Three implications for capital

1. Family offices and sovereign funds. In the next 36 months it is advisable to reallocate towards the entire compute chain: foundries, producers of lithography equipment, dedicated energy infrastructure. Geographic risk must be priced by country, as well as by sector.

2. Chief Risk Officers. The scenario of a G7 central bank with analytical capabilities compromised by a chip blockade remains absent from VAR models. It should be inserted as a tail risk with systemic impact.

3. CFOs and Investor Relations. The narrative of a purely deflationary and neutral AI risks appearing naive in eighteen months. Whoever brings it to investors today will have to correct it tomorrow. Better to anticipate the revision now, with a thesis that integrates the geopolitical risk of semiconductors into the macro case.

The prediction

Here is the explicit call. By 31 December 2027, the BIS Innovation Hub will document at least three central bank AI projects in operational production, beyond the pilot phase.

Confidence: Medium (68%). Horizon: 31 December 2027. Verification: the official Innovation Hub reports and the BIS Quarterly.

Kill signal: BIS documents record zero projects in production by that date. In that case the thesis of accelerated adoption falls and must be rewritten from the ground up.

What to watch

Three lead indicators will tell whether the pattern holds.

• New semiconductor fab capacity operational outside East Asia
• Extensions of US restrictions on the export of advanced GPUs
• Central bank publications on language models in production

These signals arrive before the headlines. Whoever reads them now maps the regime, rather than merely suffering it. The rest will follow prices, as always.