Financial institutions are rapidly expanding their use of artificial intelligence, including agentic systems capable of executing activities with increasing autonomy. But how well do firms understand the physical and commercial infrastructure supporting these services?
The resilience of AI services depends on much more than the resilience of the AI provider itself.
When an AI provider’s biggest risk isn’t the AI
Recent reporting about Anthropic’s arrangements for gas-fired electricity-generation equipment illustrates the extraordinary infrastructure requirements of frontier AI. Reports described commitments worth approximately $3.4 billion, covering around 1.7 gigawatts of generating capacity. These reports should not be treated as independently verified confirmation of the transaction.
AI providers increasingly need to secure not only computing equipment but also reliable electricity, data-centre capacity, cooling and construction resources. This expands the risk ecosystem well beyond traditional technology subcontracting.
Power: a potential constraint on AI growth and service delivery
In remarks at a G20 innovation event in September 2026, Elon Musk warned that electricity availability could constrain AI expansion, citing an anticipated power shortfall of around 15 gigawatts in 2027. He contrasted rapid growth in AI-chip production with slower growth in electricity infrastructure outside China. These are Musk’s estimates, not independently established forecasts.
For financial institutions, the question is not simply whether AI development slows. Could competition for scarce electricity and computing capacity affect the cost, performance, availability or reliability of AI services already embedded in business processes?
Geopolitics and domestic politics add another layer
Semiconductor export controls, competition between the US and China, energy security and national AI sovereignty influence where AI infrastructure can be built and operated. At the same time, proposed data centres face local scrutiny over electricity prices, grid capacity, water consumption and environmental effects.
Political debate around data-centre expansion has included opposition from Democratic candidates in the United States, but local concerns cross party lines. Development delays, permitting restrictions, supply-chain constraints and changing policy priorities can all translate into upstream operational risks.
Illustrative dependency layers, not a complete subcontracting chain. Political and geopolitical risks cut across multiple layers.
From infrastructure challenge to political controversy
Data-centre expansion is becoming a political and community issue, not merely a technical or environmental one. Reporting in November 2026 described coordinated protests against data-centre development across approximately 15 European countries, with campaigners raising concerns about electricity and water consumption, environmental impacts and pressure on local infrastructure. In the Netherlands, protesters reportedly entered a Microsoft data-centre construction site, temporarily interrupting work.
Governments face increasingly difficult trade-offs between AI investment and competitiveness, energy security, environmental objectives and public opposition. Planning objections, permitting restrictions, community protests and changing political priorities can delay projects, increase costs or constrain where capacity can be developed. These are potential risk pathways, not evidence that any particular AI service is currently unreliable.
For financial institutions, these developments are relevant to both third-party risk management and operational resilience. An AI provider may be technologically sophisticated and financially well-resourced yet still depend on infrastructure projects whose development or operation is exposed to factors outside its control. Constraints on new capacity could also reinforce concentration in established cloud providers and existing data-centre locations.
The key question is whether firms understand the resilience of the wider AI dependency ecosystem, including the political and community factors that may affect future infrastructure availability.
The hidden supply chain: beyond subcontractors
Conventional TPRM frameworks focus on direct providers and material subcontractors. Yet AI services may depend on cloud platforms, semiconductor manufacturers, electricity generators, fuel suppliers, grid connections, cooling systems, telecommunications, specialist construction and infrastructure financing.
Not every upstream dependency is a subcontractor or fourth party in a legal or regulatory sense. Nevertheless, its failure or constraint may affect the service received by a financial institution. Firms therefore need to distinguish contractual outsourcing chains from the broader ecosystem of operational dependencies.
Why agentic AI raises the stakes
An unavailable AI assistant may inconvenience staff. An autonomous agent executing steps in customer service, financial crime, payments or operational workflows could create a more direct service disruption. Dependencies on common foundation models or cloud infrastructure may also undermine apparent diversification between different AI solutions.
Operational resilience assessments should consider service degradation as well as complete outages, human override, tested fallback arrangements and the ability to maintain important business services within impact tolerances.
Six practical questions for TPRM
- Materiality and criticality: Which AI applications support important business services, critical or important functions, or other material activities?
- Dependency visibility: Which upstream infrastructure dependencies are material, and what assurance is available?
- Concentration: Do apparently independent providers share computing platforms, geographic locations or other infrastructure?
- Capacity and continuity: How will providers manage power constraints, computing shortages, service degradation and competing customer demand?
- Geopolitical and external risks: How might export controls, political decisions, energy markets and infrastructure delays affect service availability?
- Fallback and exit: Can the firm continue operating if the AI service is unavailable, and are alternatives genuinely independent and viable?
These questions belong throughout the 3 Os of TPRM: On-Boarding, OnGoing oversight and Off-Boarding, supported by initial and continuing materiality and risk assessments. Proportionate oversight does not mean attempting to assess every supplier of every supplier.
JADEtc Academy perspective
AI offers substantial opportunities for financial services, but the infrastructure supporting it is complex, interconnected and potentially constrained. The challenge is not simply whether the direct AI provider has sound security and continuity arrangements. It is whether firms sufficiently understand the material dependencies, vulnerabilities and concentration risks that could affect the AI services on which they increasingly rely.
As agentic AI takes on greater operational responsibilities, understanding the resilience of the wider service-delivery ecosystem becomes an essential part of effective TPRM and operational resilience.
Companion Insight: Could Electricity Become AI’s Biggest Constraint? Elon Musk’s G20 Warning →
- Financial Times — reporting on Anthropic’s electricity-generation arrangements (November 2026).
- Le Monde — reporting on European protests against data-centre development (9 November 2026).
- International Energy Agency: Energy and AI.
- DORA: Regulation (EU) 2022/2554.
Statements attributed to Musk are estimates, not independently verified industry forecasts. Political and commercial developments are discussed as risk scenarios, not predictions of particular provider failures.