Who pays for AI’s electricity bill?

Who actually pays for the electricity used by AI?

AI companies pay their direct electricity bills, but households, businesses and sometimes taxpayers also pay part of the cost of powering the AI boom. The difference comes down to who pays for the electricity itself and who finances the infrastructure needed to supply it.

Microsoft, Google, Amazon and other major operators buy electricity from utilities, independent generators or specialized data center providers. Those expenses become part of the cost of running their cloud infrastructure and AI services. Power bills sit inside the cost of running AI; we compare who makes money and who burns cash in our AI infrastructure report.

But an electricity bill does not necessarily cover every cost a new data center creates.

Utilities may need to build substations, expand transmission lines or secure additional generating capacity. Some expenses are charged directly to the data center. Others are recovered through electricity rates paid by millions of customers.

The distinction is already shaping regulation across the United States. Ohio now requires qualifying data centers to make long-term payment commitments. Virginia has approved a separate electricity rate class for very large customers, while Georgia has introduced financial protections for existing ratepayers.

Why has AI electricity become such a big financial issue?

AI electricity costs are becoming a major financial issue because data centers are adding enormous demand to power grids that cannot expand nearly as quickly.

The International Energy Agency estimated that data centers consumed approximately 415 terawatt-hours worldwide in 2024, equivalent to 1.5% of global electricity consumption. Its base-case forecast reaches 945 TWh by 2030, an increase of 128%.

That additional 530 TWh is larger than Japan's annual electricity consumption. AI-optimized facilities are expected to drive much of the increase, although ordinary cloud services and other computing activities also contribute.

The United States has already experienced a sharp acceleration. According to Lawrence Berkeley National Laboratory, American data centers consumed 58 TWh in 2014 and 176 TWh in 2023. Its 2028 projections range from 325 to 580 TWh.

For perspective, U.S. electricity demand barely grew for much of the previous two decades. The Energy Information Administration reported average annual growth of approximately 0.1% between 2005 and 2019, compared with 1.7% between 2020 and 2025.

Timing makes the situation harder. A data center can be built in two or three years, while major transmission projects and power plants often take considerably longer.

Indicator	Earlier level	Recent level or forecast	Change
Global data center electricity	415 TWh, 2024	945 TWh, 2030 forecast	+128%
Global electricity share	1.5%, 2024	Just under 3%, 2030 forecast	Almost doubles
U.S. data center electricity	58 TWh, 2014	176 TWh, 2023	+203%
U.S. data center electricity	176 TWh, 2023	325–580 TWh, 2028 forecast	+85% to +230%
U.S. data center electricity share	4.4%, 2023	6.7–12%, 2028 forecast	+2.3 to +7.6 percentage points
U.S. annual electricity demand growth	0.1%, 2005–2019	1.7%, 2020–2025	17 times faster
Irish data center electricity	1,240 GWh, 2015	7,663 GWh, 2025	+518%
Irish data center electricity share	5%, 2015	23%, 2025	+18 percentage points

How much does electricity cost a large AI data center?

A large AI data center can easily spend more than $100 million a year on electricity, and the biggest campuses can face annual bills several times higher.

The U.S. Energy Information Administration reported an average industrial electricity price of 8.62 cents per kilowatt-hour for 2025. Individual data centers may pay considerably more or less depending on their location, contracts and network charges, but this provides a useful benchmark.

A facility consuming an average of 100 megawatts continuously would use 876 million kilowatt-hours annually. At 8.62 cents per kWh, its electricity bill would reach approximately $75.5 million.

A 500 MW facility would spend approximately $378 million, while a 1 GW campus would approach $755 million annually.

These are illustrative calculations based on average operating loads, not verified bills from particular facilities. A project with 500 MW of maximum capacity could consume substantially less electricity if its average utilization were lower.

The numbers also exclude GPUs, buildings, networking equipment, financing and maintenance.

For a continuously operating 500 MW facility, every additional cent per kilowatt-hour adds $43.8 million to the annual bill.

Are Americans already paying higher electricity bills because of AI?

Some Americans are already exposed to higher electricity costs partly driven by data center expansion, even when those facilities pay their own bills. AI does not explain every increase in residential power prices, but its contribution is becoming harder to ignore.

The strongest evidence comes from PJM, the regional electricity market serving approximately 67 million people across 13 states and Washington, D.C.

PJM operates a capacity market that pays generators to make electricity supply available when needed.

In its auction for the 2025–2026 delivery year, the main capacity price jumped from $28.92 to $269.92 per megawatt per day, more than nine times the previous level. Prices subsequently reached approximately $329 for 2026–2027 and $333 for 2027–2028.

Those costs can eventually reach households and businesses through their electricity suppliers.

Data centers are an important source of rising demand in the region, but the increase also reflects retiring power plants, changing reliability requirements and insufficient new generating capacity.

Virginia provides a more direct estimate of what the expansion could mean for households.

In 2024, the state's Joint Legislative Audit and Review Commission estimated that data center growth could increase a typical Dominion Energy residential customer's monthly generation and transmission costs by approximately $14 to $37 by 2040, depending on the scenario.

Those figures are projections rather than bills customers are already paying. They also predate Virginia's newer large-customer tariff protections.

The mechanism behind these increases involves both electricity supply and grid investment.

When demand rises faster than available generation, wholesale power becomes more expensive. Utilities may also need to build substations, transmission lines or additional generating capacity, with regulators deciding how those costs are divided among customers.

A dedicated connection can often be billed directly to the developer. Shared transmission infrastructure is more complicated because it may also serve nearby homes and businesses.

Virginia's JLARC investigation found that existing utility tariffs generally allocated current costs appropriately, while warning that rapid data center growth could still increase future electricity prices.

There is also a potential benefit. A large customer using previously underutilized infrastructure can contribute substantial revenue and help spread fixed costs across more electricity sales.

For example, a 100 MW facility connecting to a grid with spare capacity may benefit existing customers. A 500 MW project requiring new power plants and transmission lines could have very different economics.

Are AI data centers really responsible for most new electricity demand?

AI data centers are becoming one of the biggest sources of additional U.S. electricity demand, and Ireland shows just how dominant computing facilities can become in a concentrated market.

The IEA expects data centers to account for nearly half of the increase in American electricity consumption through 2030. Data centers account for nearly half of U.S. demand growth, and we cover what changed over the past year in our report on AI infrastructure.

That still leaves substantial demand growth from manufacturing, electric vehicles, buildings and other industries.

Data centers are unusual because one campus can require hundreds of megawatts at a single location. A utility might spend years adding residential and commercial customers without encountering the electricity requirement of one very large computing facility.

Ireland gives us an especially useful historical comparison.

The country's Central Statistics Office reported that data centers consumed 1,240 GWh in 2015, representing 5% of metered electricity. By 2025, consumption had reached 7,663 GWh, representing 23%.

We can go further by calculating how much of Ireland's total electricity growth came from data centers.

Between 2015 and 2025, data center consumption increased by 6,423 GWh. Total metered electricity consumption rose from 24,600 GWh to 32,986 GWh, an increase of 8,386 GWh.

Data centers therefore accounted for approximately 76.6% of Ireland's entire net increase in metered electricity consumption over that decade.

That's a remarkable concentration of demand in one industry. The latest annual figures also show that data center consumption increased by 10% in 2025, compared with 2% for all other metered customers.

These facilities include traditional cloud computing as well as AI workloads, so the entire increase cannot be assigned to artificial intelligence.

Is Ohio making AI data centers pay more for their electricity?

Ohio has introduced one of America's strongest financial protections against speculative data center projects, requiring qualifying developers to pay for substantial contracted capacity even when they use less electricity.

In July 2025, the Public Utilities Commission of Ohio approved AEP Ohio's special data center tariff.

The tariff includes minimum billing commitments tied to 85% of contracted capacity, along with long-term agreements, financial assurances and protections against early termination. The arrangements generally cover 12 years, including a four-year ramp-up period.

If a developer reserves hundreds of megawatts, the utility may need to build infrastructure before the facility opens. Without a binding commitment, the developer could cancel the project and leave other customers exposed to those costs.

AEP Ohio's February 2026 filing provides evidence that the policy is producing binding commitments.

The utility reported 5,642 MW of contracts signed under the new tariff, in addition to 12,219 MW contracted before the tariff took effect.

Together, those agreements represented 17,861 MW, equivalent to almost 18 large 1 GW facilities in contracted capacity.

But only approximately 32% of that total was signed under the newer tariff. Contracts signed under earlier arrangements still represent most of the combined capacity and do not automatically benefit from identical protections.

The Ohio Consumers' Counsel supported the special tariff to reduce the risk that ordinary customers would finance infrastructure associated with uncertain developments.

Data center industry groups challenged the decision, showing how much money is potentially involved when utilities require developers to guarantee capacity they may never fully use.

There is another important distinction: contracted capacity is not the same as electricity currently being consumed.

Utilities may need to order transformers, expand substations and arrange generation years before a facility opens. If the project disappears, those investments cannot always be canceled or reused elsewhere.

This risk becomes more serious when developers request large amounts of power for several possible projects without having secured all the financing or customers needed to complete them.

Minimum billing commitments, security deposits and exit fees help protect against those cancellations, although developers can still become insolvent and contracts can be disputed.

Has Virginia protected households from AI data center costs?

Virginia has made large data centers financially responsible for more of their contracted electricity infrastructure, but it has not guaranteed that households will avoid every price increase linked to data center growth.

Northern Virginia is one of the world's largest data center markets, making its electricity policies especially important.

In November 2025, the Virginia State Corporation Commission approved a new GS-5 rate class for customers demanding at least 25 MW. The class takes effect in January 2027.

Qualifying customers must pay minimum charges tied to 85% of contracted distribution and transmission demand and 60% of contracted generation demand, subject to the detailed tariff rules.

The commission also introduced credit and security requirements intended to reduce exposure if large customers fail to meet their commitments.

These measures address the risk that a utility builds infrastructure for a proposed data center that consumes much less electricity than expected.

However, broader electricity market conditions can still affect prices. If regional demand pushes up wholesale generation or capacity costs, existing customers may continue to face higher bills.

Virginia's earlier JLARC study had already identified this risk, estimating potential monthly residential cost increases of $14 to $37 by 2040 under different data center development scenarios. Those estimates should not be treated as forecasts of the newly approved GS-5 tariff.

Is Georgia forcing AI companies to pay for new power plants?

Georgia is placing more of the financial risk from data center expansion on Georgia Power and large electricity customers, rather than automatically passing it to existing ratepayers.

In January 2025, the Georgia Public Service Commission approved rules allowing Georgia Power to apply special contract terms to new customers requiring more than 100 MW.

The state subsequently approved a major expansion of electricity generation in December 2025.

The commission described the approval as covering approximately 10 GW of new capacity, with roughly 80% expected to support data center demand.

That means close to 8 GW of the planned capacity was associated with anticipated data center requirements. It's a huge amount of electricity to plan around demand that has not yet fully materialized.

The arrangement includes financial protections if the expected customers fail to materialize. Georgia Power agreed to backstop costs associated with the new generation through 2031.

Regulators also identified measures that could reduce costs if demand forecasts prove too optimistic, including stopping unstarted projects, ending certain electricity purchase arrangements or selling excess power.

Georgia also has a base-rate freeze through 2028, although that should not be interpreted as a permanent guarantee against every component of an electricity bill increasing.

Are Microsoft, Google and Amazon building their own power plants for AI?

Microsoft, Google and Amazon are increasingly supporting new power generation through long-term contracts and investments, particularly in nuclear energy, but their AI infrastructure still depends heavily on shared electricity grids.

Microsoft offers one of the clearest examples.

In September 2024, Constellation Energy announced a 20-year agreement with Microsoft supporting the planned restart of Three Mile Island Unit 1 in Pennsylvania, now renamed the Crane Clean Energy Center. Microsoft signed a 20-year deal to restart Three Mile Island, and we look at who signs the giant contracts in our AI infrastructure business models report.

The facility is expected to restore approximately 835 MW of nuclear generating capacity.

Microsoft's long-term electricity purchase commitment helps make the investment commercially viable. The company is effectively supporting new electricity supply rather than relying entirely on existing generating capacity.

Google has also pursued nuclear agreements, including a partnership with Kairos Power for advanced reactors.

Amazon has invested in advanced nuclear development through partnerships involving X-energy and utilities. Its nuclear initiatives include plans for several projects that could add substantial generating capacity over the coming years.

Buying electricity through a long-term agreement is different from owning a power station. Financing reactor development is different from having an operating reactor supplying a data center.

Many announced advanced nuclear projects also depend on licensing, construction and successful commercial deployment.

Microsoft's Pennsylvania arrangement is tied to an existing nuclear facility, while small modular reactor projects involve newer technologies and potentially longer timelines.

Some data center operators are also exploring onsite generation to reduce their reliance on public grids.

A campus can use gas turbines, solar generation, batteries or other systems to supply some of its electricity requirements. For developers facing years of connection delays, these options can be commercially attractive.

Ireland has gone further by making local energy resources part of its data center connection policy.

In December 2025, the Commission for Regulation of Utilities adopted a framework requiring new data center applicants to provide generation or storage capacity onsite or nearby, matched to their maximum requested grid import capacity.

However, onsite generation introduces additional expenses. Gas turbines need fuel and maintenance. Solar generation varies throughout the day, while batteries need an energy source for recharging.

A campus may also keep a substantial grid connection for backup electricity even when most normal consumption comes from onsite equipment.

Will AI keep pushing up electricity prices even as companies add new power plants?

AI demand can continue pushing electricity prices higher because new power plants often arrive too slowly or in the wrong places to relieve existing shortages.

The IEA expects renewable generation to supply nearly half of the additional global electricity required by data centers through 2030. Natural gas and coal together are projected to provide more than 40% of that increase.

But electricity produced in one location cannot always be delivered efficiently to another.

Transmission capacity, connection queues and the availability of generation during peak hours all influence prices.

PJM offers a recent example.

Its 2027–2028 capacity auction fell short of the region's reliability requirement by 6,517 MW. The grid operator identified the shortfall as the first of its kind since its capacity market began in 2007.

The result reflects growing demand and insufficient eligible supply, alongside retirements and other market factors.

Data centers are a significant part of the demand story, especially in northern Virginia and neighboring regions.

For an AI operator, signing a contract for future generation may improve long-term supply. For households, the immediate electricity market can remain tight until that generation actually becomes available.

Corporate renewable electricity purchases also need careful interpretation.

A company can buy enough renewable electricity over a year to match its annual consumption while continuing to draw power from gas- or coal-fired generation during hours when renewable output is low.

Are taxpayers helping pay for AI electricity infrastructure?

Taxpayers are helping support some electricity projects and tax incentives connected to AI infrastructure, although public financial support should not automatically be treated as a direct subsidy or loss.

Government loans, guarantees and tax exemptions can lower the cost of building infrastructure or encourage companies to locate data centers in particular regions.

Nuclear projects provide one example.

The Associated Press has reported federal financing support of up to $4.4 billion across three U.S. nuclear restart projects. Such loans can reduce the financing barriers associated with bringing generating facilities back into service.

The projects can benefit a wide range of electricity customers, including data centers, manufacturers and households.

A government loan also differs from a grant. If the borrower repays the money, the taxpayer has not necessarily incurred a loss. Nevertheless, public financing can expose the government to credit risks that private lenders might otherwise carry.

Tax incentives provide another route.

Virginia offers qualifying data centers exemptions from sales and use taxes on eligible equipment.

According to Virginia's economic development incentive reporting, those exemptions represented approximately $2.73 billion in estimated tax savings over fiscal years 2015–2024.

These exemptions are broader data center incentives rather than direct payments for electricity.

Who pays for the pollution from AI electricity use?

The environmental costs of AI electricity are partly paid by power companies and their customers, while some consequences fall on the wider public.

According to the IEA, coal supplied approximately 30% of the electricity used by data centers worldwide in 2024. Natural gas provided about 26%, renewables 27% and nuclear power around 15%, with the remainder reflecting other sources and rounding.

The electricity mix varies enormously between countries.

American data centers rely heavily on natural gas, which accounted for more than 40% of their electricity supply. In China, coal represented almost 70%.

Those differences influence the emissions associated with running similar computing equipment in different locations.

The IEA projects that global carbon dioxide emissions from electricity generation serving data centers could reach approximately 320 million tonnes around 2030 in its base case.

These figures cover all data center activities, including conventional cloud computing. They should not be presented as emissions created exclusively by AI.

Carbon taxes, emissions trading and environmental rules can incorporate some pollution costs into electricity prices. Where these mechanisms are absent or incomplete, the financial and health consequences of emissions are spread more widely.

Corporate clean-energy agreements can help reduce those effects, especially when they support additional low-carbon generation. Yet matching annual electricity consumption with renewable certificates does not necessarily eliminate the emissions associated with operating a data center around the clock.

Will ChatGPT and other AI services become more expensive because of electricity?

AI services could become more expensive as electricity costs rise, but GPU expenses, competition and computing efficiency currently matter far more to many providers' pricing decisions.

Data centers pay for electricity and infrastructure. Cloud providers sell computing capacity. AI developers either operate their own infrastructure or purchase capacity from providers such as Microsoft Azure, AWS and Google Cloud.

Consumers and enterprise customers ultimately help finance those expenses through subscriptions, API charges and software contracts. Subscriptions and API charges end up covering these costs, and we compare which business models are winning in our report on AI infrastructure.

How much additional electricity expense reaches customers depends on each company's margins and pricing strategy.

A cloud provider generating billions in profit may absorb some higher power costs. A heavily indebted infrastructure operator with expensive computing equipment may have much less flexibility.

At a continuously operating 100 MW facility, an additional one cent per kilowatt-hour raises annual electricity expenditure by $8.76 million.

At 500 MW, the increase reaches $43.8 million annually.

Those amounts need to be compared with the cost of advanced processors, data center construction, financing and the revenue generated by the infrastructure.

Could more efficient AI models solve the electricity problem?

More efficient AI models will reduce the energy needed for individual tasks, but they are unlikely to eliminate the growth in total data center electricity demand.

AI developers are improving hardware utilization, model architecture, quantization and inference software. Smaller models can also handle tasks that previously required much larger systems.

The difficulty is that lower costs encourage more usage.

When inference becomes cheaper, companies can introduce AI features into more products, process larger volumes of information and automate tasks that were previously too expensive.

Suppose an AI service handles one billion requests per year. If the electricity required per request falls by 50% while total requests triple, annual electricity consumption still increases by 50%.

The IEA's projections show the range of possible outcomes. Its high-efficiency scenario allows for faster technological improvements, while its faster-adoption scenario produces stronger electricity demand.

Historical data also suggest caution.

American data center electricity use tripled between 2014 and 2023 despite substantial improvements in processor and server efficiency. Ireland's data center consumption increased every year between 2015 and 2025.

One potential way to reduce grid costs is to make some AI workloads more flexible.

Tasks such as model training and batch processing can sometimes be shifted away from peak electricity periods, unlike interactive applications that need immediate responses.

If operators can reliably reduce demand during expensive hours, utilities may need less backup capacity or fewer infrastructure upgrades.

Who should pay for AI's new electricity infrastructure?

AI companies should pay the additional infrastructure costs and financial risks directly created by their projects, while shared investments should be allocated according to who benefits from them.

Ohio, Virginia and Georgia are already moving toward stronger financial commitments for large electricity users.

The basic mechanisms are minimum billing obligations, enforceable long-term contracts, financial guarantees and charges that reflect the additional infrastructure required by large customers.

These arrangements are particularly important when a project reserves enormous amounts of electricity years before it begins operating.

At the same time, some infrastructure genuinely benefits more than one customer.

A transmission upgrade originally prompted by a data center may also improve regional reliability. A new power station can serve households, manufacturers and other businesses.

Regulators therefore need credible estimates of the additional cost caused by each project rather than simply assigning every new network expense to AI companies.

We can separate the main financial responsibilities.

Cost or risk	Who currently pays?	What stronger protection would involve
Electricity consumed	Data center operator	Full payment under applicable supply tariffs
Dedicated grid connections	Developer or utility, depending on tariff	Direct project-specific connection charges
Regional transmission upgrades	Utilities and allocated network users	Charges reflecting incremental costs caused by large customers
Additional generating capacity	Utilities, generators and investors	Long-term supply commitments and cost recovery from benefiting users
Wholesale electricity price increases	Electricity buyers across the market	Additional supply, demand flexibility and better market planning
Canceled data center projects	Developers, utilities or wider customers	Security deposits and enforceable exit fees
Unused contracted capacity	Depends on electricity contracts	Minimum billing and take-or-pay commitments
Public loans and tax incentives	Governments and taxpayers	Transparent fiscal accounting and measurable economic returns
Pollution and emissions	Generators, customers and wider society	Effective environmental pricing and enforceable emissions rules

Who pays for AI's electricity bill in the end?

AI companies pay their direct electricity bills, but the broader cost of powering artificial intelligence is still shared with households, businesses and taxpayers. Current regulations do not guarantee that the technology industry covers every additional cost its expansion creates.

For electricity consumed inside a data center, the operator is responsible. Those expenses are ultimately supported by cloud customers, AI developers, enterprise software buyers and consumers.

For new infrastructure and tighter electricity markets, the answer is more complicated. Utilities can recover costs from wider customer groups, and rising regional demand can push up electricity prices even when individual data centers pay all their contractual charges.

PJM's capacity market shows how electricity costs can increase across an entire region. Virginia's independent analysis identified potential long-term household bill increases linked to data center expansion.

Meanwhile, Ohio's new data center tariff, Virginia's GS-5 rate class and Georgia's financial protections show that regulators are taking the risk of cost shifting seriously enough to change existing rules.

These reforms should make developers more responsible for their electricity requirements, particularly when projects are delayed, canceled or underused.

But the scale of future demand remains enormous. Global data center electricity consumption is projected to more than double by 2030, and much of the necessary generation and transmission capacity still needs to be built.

For now, AI companies are paying for the electricity they buy, but they are not yet demonstrably paying the full economic cost of the power systems required to support their expansion.

Whether households continue subsidizing part of that growth will depend largely on how effectively regulators enforce long-term cost commitments and prevent speculative electricity demand from becoming everyone else's financial problem.