Banks have spent years preparing their data for artificial intelligence. They have migrated workloads to the cloud, built data platforms, strengthened governance and invested heavily in making information accessible to increasingly sophisticated models. Yet as AI moves from experimentation into the everyday machinery of banking, the industry’s next constraint may emerge somewhere few banking executives have traditionally needed to look: the electricity grid. AI requires enormous computing infrastructure, that infrastructure requires power, and the arrival of AI agents could multiply the amount of computation behind seemingly ordinary financial interactions. The question for banks may therefore shift from whether they have enough data to deploy AI to whether the economics of turning electricity into intelligence continue to make sense at banking scale.
The numbers behind the AI infrastructure build-out are becoming difficult to ignore. The International Energy Agency estimates that data centres consumed about 485 terawatt-hours of electricity globally in 2025, with consumption projected to roughly double to around 950 TWh by 2030. More importantly for financial institutions, electricity consumption from AI-focused data centres is expected to triple over the same period.
That does not mean banks are suddenly going to start building power stations beside their headquarters. Most will experience the issue indirectly through cloud providers, AI platforms, software vendors and the economics of computing capacity.
But indirect exposure does not make it irrelevant.
Financial services could become one of the world’s most intensive institutional users of AI. Fraud detection, anti-money-laundering monitoring, credit assessment, customer service, software development, investment research, compliance, document processing, personalisation and financial advice are all candidates for significantly greater AI involvement.
Today, many of these applications remain relatively contained. A bank might deploy a generative AI assistant to several thousand employees or introduce an AI chatbot for customers. The computational requirements are meaningful, but manageable.
The equation begins to change when AI stops simply answering questions and starts performing work.
Agentic AI is particularly important because an AI agent may require substantially more computation than a simple prompt-and-response interaction. Instead of receiving a question and generating an answer, an agent might interpret a customer’s request, retrieve information from several systems, reason about the appropriate action, call external tools, verify the result, execute a transaction and document what happened.
The IEA has specifically identified reasoning, agentic tasks and video generation as substantially more energy-intensive AI applications, noting that some can consume hundreds or even thousands of times more energy per query than simple text generation. At the same time, efficiency per AI task is improving rapidly, making the ultimate trajectory difficult to predict.
That combination creates an interesting problem for banking.
AI could become dramatically more efficient while total electricity consumption continues rising because institutions use vastly more of it.
The same phenomenon has occurred repeatedly in computing. When the cost of performing an activity falls, organisations often do much more of it. Banks may therefore save enormous amounts of compute on individual AI tasks while simultaneously introducing AI into millions of processes that previously required little or no AI computation.
Imagine a future retail bank serving ten million customers. Instead of customers occasionally interacting with a chatbot, each customer could eventually have an AI financial assistant continuously monitoring cash flow, upcoming payments, savings goals, unusual transactions, credit positions and investment opportunities.
At the same time, employees could have their own AI agents preparing meetings, reviewing documents, monitoring compliance obligations and executing administrative workflows.
The bank would effectively move from occasionally calling AI to running an enormous layer of continuous machine intelligence across the organisation.
At that point, the cost of intelligence becomes an operating variable.
Banks already obsess over cost per transaction, cost per account and cost-to-income ratios. It is conceivable that AI introduces another metric into that vocabulary: compute cost per decision.
A highly sophisticated reasoning model may be capable of handling a customer request, but that does not necessarily mean it is economically sensible to use it. A smaller model might perform the task adequately at a fraction of the computational cost. Conventional software might be cheaper still.
This could produce an entirely new discipline inside financial institutions: deciding not simply where AI can be used, but where intelligence is economically worth deploying.
A complex wealth-management conversation may justify considerable computation because the economic value of the relationship is high. Detecting a sophisticated fraud pattern may justify almost any reasonable computational expense because the potential loss is significant. Using an advanced reasoning model to answer a basic question about branch opening hours clearly would not.
The AI architecture of a bank could therefore increasingly resemble a hierarchy of intelligence, with different models and computational resources allocated according to the complexity and economic value of each task.
Electricity is also becoming important because the infrastructure supporting AI cannot expand as quickly as software.
A new application can be deployed in months. Power generation, transmission infrastructure and grid connections can take years. The IEA says data centres can be operational within two to three years, while the wider energy infrastructure required to support them generally has much longer planning and construction cycles. Transformers, turbines, grid connections and other physical infrastructure have consequently become potential bottlenecks in the AI expansion.
This is where an infrastructure problem begins to become a financial-services problem.
Cloud providers ultimately have to recover the cost of building and operating enormous computing estates. Electricity is only one component of that cost, alongside chips, networking equipment, cooling systems, land and capital, but the availability and price of power increasingly influence where computing capacity can be built and how quickly it can expand.
Banks may never see an electricity line item attached to an AI transaction. They could nevertheless feel its effects through cloud pricing, model pricing, capacity availability and the economics offered by technology vendors.
There is another side to the story that makes the relationship between banking and AI electricity demand even more significant: financial institutions will help finance the infrastructure creating that demand.
Data centres, power generation, transmission networks, renewable projects, gas generation and potentially new nuclear technologies require enormous amounts of capital. Banks, institutional investors, private-credit providers and asset managers are increasingly positioned to participate in financing that build-out.
That connection is already visible. In September 2026, a consortium of ten banks was reported to be providing $22 billion of financing for an AI cloud venture backed by Blackstone and Alphabet, illustrating the extraordinary scale of capital now flowing into computing infrastructure.
Financial services therefore occupies an unusual position in the AI-energy equation.
The industry will consume AI, finance the infrastructure supporting AI and invest client capital into companies building that infrastructure.
It will also have to reconcile those activities with sustainability commitments. The IEA expects renewables to provide a substantial share of additional electricity required by data centres, but natural gas, nuclear and other sources will also play important roles.
That could make the physical footprint of digital banking increasingly difficult to ignore. The cloud has always sounded almost weightless. AI is revealing how physical it really is.
Behind every model sits a data centre. Behind the data centre sits a grid connection, cooling infrastructure, semiconductor supply chain and ultimately a source of electricity.
For banks, none of this suggests that AI adoption is about to stop. The economic value of automation, fraud prevention, personalisation and productivity could vastly outweigh the additional computing costs involved.
But it does suggest that the next stage of AI adoption will require a more sophisticated understanding of economics.
The first era of banking AI was about whether models worked. The next is increasingly about whether they can be governed and integrated safely. The era after that may be about determining where intelligence produces enough value to justify running it continuously at enormous scale.
Banks have spent decades learning how to allocate capital efficiently.
They may now have to become equally disciplined about allocating computation.
What it means for the industry
- AI economics could become part of banking economics. Financial institutions may increasingly measure the computational cost of automated decisions rather than treating AI simply as another technology licence.
- AI agents could materially change demand. Agentic workflows involve multiple reasoning and system interactions, potentially making them significantly more computationally intensive than today’s simple AI assistants.
- Not every banking task needs the most powerful model. Banks are likely to develop tiers of AI capability, matching model complexity and compute expenditure to the economic value and risk of individual tasks.
- Financial institutions will sit on both sides of the infrastructure boom. Banks and asset managers will consume increasing amounts of AI while simultaneously financing data centres, grids and power generation supporting its growth.
- The cloud is becoming visibly physical. As AI scales, electricity availability, grid infrastructure and computing capacity could become indirect strategic considerations for financial institutions that previously treated cloud resources as effectively unlimited.
