AI governance: dependence, bottlenecks and state leverage

AI power / The institutions behind the interface

Who governs the systems that institutions cannot do without?

AI governance changes when the distribution of power changes. A model becomes useful; a firm gives it discretion; customers build around its outputs; a state treats access as strategically important. Rules then face actors with stronger incentives and greater leverage than the original policy anticipated. Ethics statements may stay unchanged while practical authority moves elsewhere.

There are two distinct risks. Human owners can use AI to consolidate control over markets, administration and access to expertise. Alternatively, sufficiently capable systems could become difficult for those owners to direct or stop. The first does not require machine consciousness or disobedience. The second requires evidence about sustained autonomous behavior, not merely a large market valuation.

The Bank of England/FCA survey published on 26 October 2026 gives a concrete view of institutional exposure. The top three cloud/compute providers accounted for 79% of reported provider relationships; the top three model/API providers accounted for 54%. These are respondent-reported relationships, not shares of global compute or proof of monopoly. They nevertheless show how many apparently separate users can depend on a narrow supplier base.

Only 2% of the survey’s respondents reported broadly autonomous agents. Supplier dependence and independent AI authority therefore require separate assessments.

The race is for control over bottlenecks.

Frontier systems depend on capital, advanced hardware, electricity, cooling, connectivity and skilled organizations. Control of those inputs can create bargaining power even when model access looks inexpensive. A customer buying many model brands may still rely on a common cloud provider. Counting interfaces can therefore overstate the diversity of the underlying system.

The IEA’s 16 April 2026 energy report treats power infrastructure and supply chains as constraints on data-centre expansion. Inference: owners of scarce connections, sites and equipment can influence where capability grows and who captures its returns. This is a political economy question, not evidence that physical limits have disappeared or that every proposed project will be built.

AI-assisted R&D could tighten the feedback between capability and ownership. Better tools may increase research throughput, helping their owners produce better tools and attract further capital. But the loop has to close through validated results. A flood of plausible experiments can consume compute and reviewer attention without producing a better model. Physical research and manufacturing add further delays.

Competition pulls in both directions. It can reward speed and encourage firms to accept weak controls. It can also create substitutes and make reliability valuable. States may restrict access for strategic advantage, while shared infrastructure risks give them reasons to cooperate. “Arms race” describes a pressure; it does not establish that all safeguards will be abandoned.

Adaptability without accountability can become permission to drift.

The 2026 International AI Safety Report identifies uncertainty about future loss of control, alongside concerns about economic and societal effects. That uncertainty does not make every future equally likely. It means policy should distinguish observed dependencies, demonstrated failures and assumptions about capabilities that do not yet have adequate evidence.

Effective oversight must survive success. As a service becomes more valuable, the constituency against interrupting it grows. Independent review, enforceable limits on delegated authority and credible alternatives matter most when replacing the system is inconvenient. Otherwise a formally cautious institution can progressively approve the very dependence its rules were meant to constrain.

Lower switching costs, independent suppliers, successful continuity exercises and gains without expanded authority weaken the dependence pathway. Governance determines whether capability and capital accumulation also transfer effective institutional control.