Who Pays For The Power Infrastructure Behind AI?

Building an AI data centre is one thing. Building the electricity system needed to keep it running is another. Those two investments are increasingly difficult to separate. The International Energy Agency (IEA) expects global data centre electricity consumption to almost double from 485 TWh in 2025 to 950 TWh by 2030. 

Electricity use from AI-focused data centres is expected to triple over the same period. In the US, the numbers are even harder to ignore. Lawrence Berkeley National Laboratory estimates data centres could account for 11.8 per cent of US electricity consumption by 2030, although uncertainty around future demand leaves a wide range of possible outcomes. 

Someone has to build the generation, transmission and distribution capacity needed to support that growth. Someone also has to finance it. Increasingly, regulators and utilities are asking a third question: what happens if they build it and the expected demand never arrives? That changes the conversation around AI infrastructure costs. 

The price of electricity is only part of the equation. The harder question is who carries the financial risk of making that electricity available in the first place.

AI Demand Creates Costs Before The Compute Arrives

There’s a timing problem sitting underneath AI’s growing demand for power. Technology moves quickly. The IEA estimates that a data centre can become operational within two to three years. The wider energy system works on a much slower schedule, with new infrastructure often requiring extensive planning, high upfront investment and long construction periods. 

Utilities therefore have to plan around expected demand rather than simply responding once that demand exists. A proposed data centre may require additional generation, new transmission capacity or upgrades to the local distribution network before its servers start consuming electricity. 

That would be relatively straightforward if every proposed facility arrived on time and consumed exactly as much power as expected. They don't. Goldman Sachs Research forecasts US data centre power demand rising from 31 GW in 2025 to 66 GW in 2027. Yet it expects only around 60 per cent of capacity scheduled for the next year to come online on time, falling to roughly 50 per cent when looking two years ahead. 

Developers may pursue several possible locations before choosing one, while construction, labour and supply chain constraints can also delay projects. For utilities, that creates an awkward situation. They need enough infrastructure to serve enormous new loads if they arrive, but building around every projected load creates its own risk. 

A substation, transmission upgrade or generating asset doesn't disappear because a data centre was delayed, downsized or cancelled. The investment has already been made, and its cost still needs to be recovered from somewhere.

The Real Question Is Who Carries The Investment Risk

Traditionally, many electricity infrastructure costs are spread across a utility's customer base. That model becomes harder to defend when one unusually large customer creates the need for billions in additional investment, particularly if there’s uncertainty around how much of that capacity it will eventually use. The problem isn't simply who pays upfront. 

It's who remains responsible for the cost if the assumptions behind the investment change. US regulators are already responding to that distinction. In June 2026, the Federal Energy Regulatory Commission (FERC) ordered the country's six regional grid operators under its jurisdiction to justify or reform rules for connecting data centres and other large electricity users. 

Among the proposed protections are Cost Recovery Agreements. These are designed to make large customers responsible for costs incurred to serve them even if they don't eventually come online as planned. They can also address the gap between utilities beginning to recover infrastructure costs and a data centre reaching full operation. 

The issue has now reached Congress too. On 16 September 2026, the US House passed the Ratepayer Protection Act by 417 votes to three. Rather than imposing a single national tariff, the legislation would require state utility regulators to consider approaches under which large-load customers cover the incremental generation, transmission and distribution costs required to serve them, alongside appropriate financial assurances. 

The details differ, but the underlying principle is becoming clearer. If one customer creates the need for additional infrastructure, regulators are increasingly looking for ways to tie more of the resulting financial exposure to that customer. How they do it, however, varies considerably.

Different Markets Are Moving The Risk In Different Ways

There isn't one established model for financing the power infrastructure behind AI. Instead, different markets are changing the relationship between utilities and large-load customers according to their own regulatory structures, available capacity and appetite for investment risk. Broadly, two approaches are becoming visible. 

Utilities can continue building the infrastructure while demanding stronger financial commitments from the customers requesting it. Or more of the responsibility for providing additional power can move towards the data-centre developer itself.

Making large loads commit to the capacity they request

Ohio offers a useful example of the first approach. The Public Utilities Commission of Ohio approved a tariff requiring new large data centres served by AEP Ohio to pay for at least 85 per cent of their contracted electricity capacity for up to 12 years, even when they use less. Michigan has taken a similar route with different terms. 

Changes approved in August for Indiana Michigan Power require customers drawing 50 MW or more to sign contracts lasting 15 years. They also introduce a 90 per cent minimum monthly billing demand, upfront collateral and charges if customers terminate contracts early or significantly reduce their capacity. Georgia has gone further in linking large customers to the infrastructure built around them. 

Customers using more than 100 MW can be charged for site-specific infrastructure as well as upstream generation, transmission and distribution costs. Longer contracts and minimum billing requirements are intended to reduce the chance of other customers inheriting those costs if expected demand disappears. 

The utility can still build and operate the infrastructure under these models. What changes is the commitment expected from the customer whose demand justified the investment. But utilities don't necessarily have to carry the entire expansion themselves.

Moving power investment closer to the data centre

Saskatchewan and Bell are testing a different balance. Bell's planned Saskatchewan AI infrastructure hub retains an existing 300 MW allocation from the provincial grid. However, the proposed expansion adds up to another 900 MW under Saskatchewan's Bring Your Own Power principle, creating a possible 1.2 GW hub. 

That creates a hybrid arrangement. Some capacity still comes from the existing electricity system, while responsibility for securing much of the additional power needed for expansion moves closer to the company creating the demand. The distinction is important. Instead of asking a utility to expand enough to meet the entire future load and then deciding how those costs should be recovered, part of the infrastructure requirement moves outside the traditional utility investment model. 

For AI infrastructure developers, these aren't simply different ways to buy electricity. They change the economics of where and how new capacity gets built.

Power Financing Is Becoming Part Of AI Infrastructure Strategy

Power procurement has traditionally been something infrastructure teams could consider once they knew where a facility would be built and how much capacity it needed. AI is making that sequence increasingly difficult. 

If securing hundreds of megawatts also means accepting a 12 or 15-year commitment, providing collateral, contributing towards infrastructure upgrades or developing dedicated generation, the power arrangement becomes part of the investment case itself. 

That means AI infrastructure strategy has to consider more than whether enough electricity is technically available at a location. Leaders also need to understand the conditions attached to accessing it:

• How much capacity needs to be guaranteed?
• Who finances the infrastructure required to provide it?
• What happens if demand grows more slowly than forecast?
• What commitments remain if the project is downsized?
• And how much flexibility does the organisation retain once those agreements are signed?

Those questions can change the total economics of an AI project long before the first workload reaches production. They can also make two apparently similar locations very different propositions. One may offer grid capacity backed by long-term minimum payments. Another could require infrastructure contributions. 

Somewhere else, securing capacity may depend on bringing additional generation with the project. Electricity availability is therefore only one part of the calculation. The financial structure behind that availability can be just as important.

The Cheapest Megawatt Isn't Necessarily The Lowest-Risk One

It would be easy to take this argument one step too far and assume the safest answer is simply to make data centres pay for everything themselves. Electricity systems are more complicated than that. Shared grid infrastructure can spread fixed costs across more users, while large customers can sometimes improve the utilisation of existing assets. 

BCG argues that data centres designed to operate more flexibly can potentially lower overall system costs by making better use of available grid capacity rather than simply adding another fixed source of demand. Dedicated generation brings different trade-offs. 

It can reduce dependence on constrained grid capacity, but it also moves more infrastructure responsibility towards the developer and changes where long-term operational and financial exposure sits. So the useful comparison isn't simply grid power versus private power, or cheap electricity versus expensive electricity. 

Infrastructure leaders need to look at the whole arrangement: the cost of the electricity, the investment needed to make it available, how much capacity the organisation must commit to, how flexible that commitment remains and who carries the risk if future demand looks different from today's forecast. 

The cheapest megawatt on paper may look very different once those obligations are included.

Final Thoughts: AI Capacity Will Depend On Who Is Willing To Carry Its Infrastructure Risk

AI companies can order servers, secure chips and plan new data centres relatively quickly. Power infrastructure doesn't move at the same speed. That gap means someone has to commit capital today based on an expectation of how much electricity AI infrastructure will need years from now. 

As projected loads grow, utilities and regulators are becoming less willing to leave existing electricity customers carrying the risk if those expectations prove wrong. The response isn't one universal funding model. We're seeing long-term contracts, minimum payments, financial guarantees, direct infrastructure contributions, dedicated generation and hybrid arrangements that divide responsibility between the grid and the customer. 

For CIOs and infrastructure leaders, this makes power financing part of the real cost of AI capacity. Where an organisation can secure electricity, the conditions attached to that capacity and who carries the long-term investment risk can all influence whether a project remains viable. 

As AI becomes more deeply tied to physical infrastructure, EM360Tech will continue examining how these changing relationships between compute, power and infrastructure are reshaping the decisions enterprise technology leaders have to make.