Audit
Who Pays for AI’s Electricity?
Data centres can increase infrastructure costs or spread existing costs more widely. A comparison of electricity policies in Ohio, Ireland and Britain shows why the contract matters more than blanket promises about cheaper power.
Who Pays for AI’s Electricity?
The decisive question about an artificial-intelligence data centre is not simply how much electricity it uses. It is what the electricity system must build for it, who promises to pay, and what happens if the promised customer does not arrive.
The International Energy Agency’s October 2026 policy review makes an important distinction: new demand can raise other consumers’ bills when it requires investment whose costs are shared broadly. But a large customer that pays appropriately and uses spare infrastructure can spread fixed costs over more consumption. Neither outcome follows automatically from installing servers. IEA, Powering Artificial Intelligence.
That distinction changes the test for public policy. An announcement that a company will buy its own electricity does not settle the price of its connection, its reservation of network capacity, or its obligations when plans change. The following comparison examines documented safeguards, not a measured claim that any one policy has already reduced household bills.
Size matters, but location changes the calculation
The IEA estimated that data centres consumed 415 terawatt-hours in 2024, approximately 1.5% of global electricity consumption. Those figures cover data centres generally, not AI alone. Its April 2025 analysis also found that nearly half of United States data-centre capacity was concentrated in five regional clusters. A modest global percentage can therefore coexist with substantial local pressure. IEA, Energy and AI.
The distinction between electricity and capacity is equally important. Electricity is the energy used over time; capacity describes the power that must be available at a particular moment. An annual purchasing commitment cannot, by itself, demonstrate that a network can serve the site at its busiest hour.
For this audit, the useful comparison is consequently not countries’ total electricity consumption. It is the obligation attached to an additional connection: whether a developer pays for reservation, construction, continued availability and premature departure. Those obligations address different risks and should not be mistaken for interchangeable charges.
Ohio: a customer cannot simply walk away
Ohio Power’s data-centre tariff took effect on 23 July 2025. Its published process requires a customer that cancels, or delays a project by more than 12 months before the target energisation date, to reimburse 100% of buildout costs under the relevant agreement. The initial service term comprises a ramp-up period of no more than four years followed by eight years. Ohio Power, Data Center Tariff.
The same published terms provide minimum-demand charges, financial security requirements and restrictions on early exit. They also permit a conditional reassignment of some contracted capacity, but require protection against stranded investment being recovered from other ratepayers. This is more precise than a general promise that technology companies will pay their way.
The mechanism is intelligible: the customer receives a commitment to serve a large load, and the utility receives obligations that survive lower consumption or withdrawal. It is not proof that every possible cost has been insulated from households. It does show how a tariff can address the risk of building assets for a customer whose demand later falls short.
Ireland: require additional supply, not just a buyer
Ireland’s regulator chose another lever. Its December 2025 connection-policy decision requires new data centres to provide generation or storage, onsite or nearby, matching requested maximum import capacity. The associated capacity must participate in the wholesale market. The policy also requires at least 80% of annual demand to be matched by additional renewable electricity generated in Ireland, allowing a six-year development glide path. Commission for Regulation of Utilities, connection-policy decision announcement.
Those are separate requirements. Renewable electricity over a year is not the same thing as dependable capacity when the system needs it. Additionality also matters: the regulator excludes renewable generation already supported through specified existing schemes from satisfying the requirement.
This approach attaches supply obligations to permission to connect. It does not establish that every associated network cost disappears, or that an annual renewable target ensures renewable supply every hour. Its consumer-protection logic is to ask what the new customer brings to the system, rather than examining only what it takes.
Britain: screen the queue before building for it
A different risk arises before construction. In July 2026, Ofgem reported that contracted demand offers had increased from 41 gigawatts to 125 gigawatts between November 2024 and June 2025, largely driven by data centres. It consulted on a commitment fee of £237,500–£712,500 per megawatt, refundable on energisation and forfeitable on early withdrawal, alongside stronger project milestones. These were proposals, not evidence of money already collected. Ofgem, 29 July 2026 consultation announcement.
The policy problem is not just unpaid equipment. An implausible project can occupy a queue position and distort the picture of future demand. A financial commitment and evidence of progress help distinguish a serious reservation from an option taken cheaply against an uncertain future.
A large fee nevertheless creates its own trade-off: it may screen out speculative demand, but it also ties up capital for genuine developers. Consumer protection therefore requires proportionate conditions, not simply the highest possible barrier to entry.
A promise of lower bills is not a bill calculation
Britain’s Data Centre Covenant, published on 28 October 2026, says data centres should cover the energy costs they create and help spread system costs across a larger customer base. It links additional government support for strategic projects to tests including local benefits, environmental standards and security. UK Government, Data Centre Covenant.
That is a policy framework, not a quantified household saving. Its promise needs to be examined alongside the actual connection agreement, charging rules and infrastructure plan. Local community investment, valuable as it may be, is not the same account as electricity-network cost recovery.
There is also an explicit evidence gap. Asked about reinforcement costs attributable to new data centres, the government told Parliament on 10 September 2026 that it was not tracking those costs specifically. It identified Ofgem as responsible for the charging framework. UK Parliament, written answer 26555.
That answer does not prove costs are uncontrolled. It does mean the cited government record cannot supply a national total or a developer-versus-household split. An audit should not manufacture either from a political commitment.
Building beside a power station does not end the question
Co-location, placing a major customer alongside generation, introduces another boundary. In December 2025, the United States Federal Energy Regulatory Commission found PJM’s rules lacked clarity and consistency for these arrangements and directed revisions. Its order addressed transmission-service choices, including options reflecting limits on withdrawals from the wider system. FERC, co-location order fact sheet.
The significance is narrower than a claim that co-location is either a loophole or a complete solution. A site’s physical proximity to generation does not alone specify its entitlement to network support. The contracted service and the limits placed on that service remain material. This historical order demonstrates the regulatory problem; it is not presented here as a complete account of subsequent PJM implementation.
The finding: follow the liabilities
These examples identify distinct protections: Ohio’s contractual payment obligations, Ireland’s supply conditions, and Britain’s effort to improve the credibility of demand reservations. They cannot be ranked as universal winners without comparable evidence on costs, enforcement and outcomes.
For an individual project, the decisive documents are the connection offer, tariff, security arrangements and approved investment plan. Together they should identify the customer’s liability if it cancels, uses less capacity, needs backup support or leaves early. Evidence of those obligations is stronger than an investment announcement or an assurance that electricity purchases will be green.
A useful public assessment would therefore disclose two scenarios: the project operating as expected, and the project falling materially short. The difference exposes the remaining liability. Without that comparison, even a detailed estimate of construction expenditure can leave readers unable to tell whether a promised benefit rests on an enforceable payment or an optimistic forecast.
The defensible conclusion is conditional. AI infrastructure does not inevitably raise everybody’s electricity bill, and expansion does not guarantee savings either. The public exposure depends on which costs are created, which risks remain, and whether enforceable obligations assign them to the party taking the investment decision.
Sources & editorial status
- IEA, Powering Artificial Intelligence ↗Source
- IEA, Energy and AI ↗Source
- Ohio Power, Data Center Tariff ↗Source
- Commission for Regulation of Utilities, connection-policy decision announcement ↗Source
- Ofgem, 29 July 2026 consultation announcement ↗Source
- UK Government, Data Centre Covenant ↗Source
- UK Parliament, written answer 26555 ↗Source
- FERC, co-location order fact sheet ↗Source
Source note
By The Kopje Journal & Review. Original editorial research; sources linked in the article.