Doug Kimmelman Says AI Power Demand Could Benefit Utility Customers - With Conditions

Doug Kimmelman made a striking claim as AI data centers push electricity demand higher: the expansion can become “an enormous gift” to utility customers. The Google News headline carrying his CNBC interview captured that promise, but not its demanding conditions.

Kimmelman’s argument challenges the dominant public narrative around AI infrastructure. Data centers are often portrayed as competitors for scarce electricity, drivers of new fossil generation, and a source of higher household bills. He sees another outcome, where large customers finance new capacity and spread fixed grid costs across a larger base.

Both outcomes remain possible. The deciding contest is not simply AI companies against electric utilities. It is the promise of customer benefits against the financial reality of building power plants, pipelines, substations, and transmission before demand is fully proven.

The distinction matters because Kimmelman is not observing this market from the sidelines. He founded Energy Capital Partners, known as ECP, in 2005 and now serves as its executive chairman. His firm invests in electricity generation, infrastructure, and assets positioned to serve the data-center buildout.

What Doug Kimmelman Is Actually Promising

Kimmelman’s customer-benefit case depends on data centers bringing enough new revenue and dedicated generation to cover the infrastructure they require.

The headline quotation came from a CNBC segment focused on AI data centers and power demand. Kimmelman argued that the new load can benefit ordinary customers, rather than automatically raising their bills. That conclusion rests on how utilities recover their costs.

Electric utilities carry major fixed expenses. Power plants, substations, transmission lines, control systems, and maintenance programs cost money even when electricity demand remains flat. Utilities generally recover those expenses across their customer base through regulated rates.

A large data center can change that arithmetic. It may consume electricity continuously, sign a long-term contract, and contribute substantial revenue toward infrastructure costs. If the new customer pays its full share, existing customers can benefit from spreading fixed expenses over more electricity sales.

That is the “gift” embedded in Kimmelman’s argument. The gift is not free energy, and it is not an automatic discount. It is the possibility that a large, consistent buyer improves the economics of an electricity system whose costs were already rising.

Kimmelman’s background explains why he emphasizes this structure. According to his ECP profile, he spent 22 years at Goldman Sachs before establishing ECP. His earlier work included power generation, utility finance, electricity trading, and energy assets during market deregulation.

That experience places his claim inside a familiar utility-finance framework. Large customers can support investments that would otherwise be difficult to fund. However, they must remain connected and keep buying power long enough for those investments to earn their expected return.

The central risk appears when a utility builds for projected demand that never arrives. A data-center developer might cancel a campus, reduce its requested capacity, or move workloads elsewhere. Existing customers can then inherit costs from underused assets.

Kimmelman has previously described a structure designed to limit that danger. In a 2025 interview about ECP’s partnership with Abu Dhabi investment company ADQ, he argued that new generation should provide “additionality,” meaning capacity added specifically for new demand.

He said ECP did not want to remove electricity from the existing grid in ways that hurt reliability or consumer prices. That condition is essential. Without new supply, a large data center can compete with homes, factories, and businesses for limited capacity.

The promise, therefore, is narrower than the Google News framing suggests. AI data centers become a customer benefit only when developers add supply, accept long-term obligations, and pay for the grid investments built on their behalf.

Why AI Power Demand Is Arriving So Quickly

The problem is not only the amount of electricity AI uses, but the speed and geographic concentration of new requests.

A conventional data center contains servers, storage, networking equipment, cooling systems, and backup power. AI facilities add dense clusters of accelerators, including graphics processing units designed to run many calculations simultaneously.

The International Energy Agency estimates that data centers consumed about 415 terawatt-hours of electricity worldwide in 2024. That represented roughly 1.5 percent of global electricity consumption after five years of 12 percent annual growth.

Its energy demand analysis projects approximately 945 terawatt-hours of data-center consumption in 2030. That would be just under 3 percent of worldwide electricity use and more than double the 2024 total.

The same forecast expects data-center electricity consumption to grow about 15 percent annually through 2030. Electricity used by accelerated servers grows faster, at 30 percent annually in the agency’s base case.

Those global percentages can sound manageable. The local effects are much harder to absorb because data centers cluster near fiber networks, customers, available land, tax incentives, and existing power infrastructure.

A gigawatt of demand arriving in one utility territory creates a different challenge from the same consumption distributed across a country. Grid planners must deliver electricity at the location and time requested, not merely produce an equivalent amount somewhere else.

Timing creates another mismatch. The IEA says a data center can become operational within two or three years. Major power infrastructure often requires longer planning, permitting, construction, and interconnection periods.

That gap pressures utilities to commit capital before they have complete information about AI adoption. It also pressures developers to find electricity before competitors secure the available capacity.

AI companies can improve chip efficiency, model architecture, and cooling. Those gains do not necessarily reduce total demand. Lower computing costs can encourage developers and customers to run more inference, train larger systems, and embed AI into more products.

This rebound effect makes forecasts unusually uncertain. A more efficient accelerator can complete a task with less electricity, while the market simultaneously creates enough new tasks to increase overall consumption.

The Electric Power Research Institute estimates that AI workloads currently represent 15 to 25 percent of data-center electricity use. Its 2026 outlook stresses that AI adoption, hardware efficiency, algorithms, and supply constraints can all change the trajectory.

That uncertainty matters to Kimmelman’s claim. Utilities must build assets with operating lives measured in decades, while AI infrastructure plans can change between corporate budgeting cycles.

Customers receive the benefit only when commercial contracts bridge that mismatch. Long-term commitments, minimum payments, security deposits, and exit fees can keep speculative demand from becoming a household obligation.

Can Google News Readers Expect Lower Power Bills?

More electricity sales can lower average system costs, but only if regulators prevent utilities from shifting data-center risks onto other customers.

Kimmelman’s case begins with economies of scale. If a utility has spare capacity, adding a large customer can increase revenue without requiring an equivalent increase in fixed investment. Existing assets produce more value, and costs are distributed across more sales.

A data center can also support investment in new generation that benefits the wider system. A new plant might provide capacity during peak demand, strengthen local reliability, or replace an older facility with higher operating costs.

The benefit becomes more plausible when a data center operates steadily. Many AI workloads can run throughout the day, producing a relatively high load factor. Load factor measures average electricity use compared with the maximum demand a customer places on the system.

A stable customer can be easier to serve than a highly variable one. It gives generators predictable demand and can produce consistent utility revenue. Some computing tasks can also shift in time when grid conditions tighten.

However, those advantages do not settle who pays. Rate design, the rules that determine how utilities divide costs among customers, controls whether the promised savings reach households.

The US Department of Energy has identified several unresolved issues in large-load rates. These include fair cost allocation, stranded-asset protection, resource adequacy, and risk sharing for newer generation technologies.

Each issue represents a condition attached to Kimmelman’s “gift.”

First, large customers must fund the direct facilities needed to connect them. Those facilities can include substations, transformers, transmission upgrades, and generation capacity.

Second, their contracts must cover the risk that expected consumption falls short. A utility should not assume that every announced campus will reach its requested capacity.

Third, data centers must contribute to capacity costs, not merely pay for the electricity they consume. Capacity refers to resources kept available for periods when the system faces its highest demand.

Fourth, regulators need transparent forecasts. Developers sometimes submit similar connection requests in multiple territories while evaluating possible sites. Counting every request as firm demand can lead utilities to overbuild.

Fifth, contracts need meaningful duration. A short commitment cannot reliably support infrastructure expected to operate for several decades.

These protections can produce a genuine customer benefit. They can also slow negotiations and make some projects less attractive. That friction is not a flaw. It is the mechanism that tests whether the claimed demand is serious enough to support long-lived investment.

The outcome also varies by market. Vertically integrated utilities own generation and networks under state regulation. Competitive power markets separate some of those functions and use auctions to procure capacity.

A contract that protects customers in Texas may not fit Virginia, Ohio, or Pennsylvania. Resource availability, market rules, grid congestion, and regulatory authority differ substantially.

Google News readers should therefore resist a national yes-or-no conclusion. Data centers do not universally lower bills, and they do not universally raise them. The contract, location, supply plan, and regulatory decision determine the result.

The Ratepayer Reality Is Already Challenging the Promise

Kimmelman describes a credible upside, but current power-market signals show that customers can face higher costs before new supply arrives.

The PJM Interconnection provides the clearest pressure test. PJM coordinates a large regional electricity market covering all or parts of 13 states and the District of Columbia. Its territory includes Northern Virginia, the country’s largest data-center cluster.

PJM uses a capacity market to ensure enough generating resources will remain available during periods of high demand. Generators receive capacity payments for being ready, separate from the electricity they sell.

In July 2026, PJM’s capacity auction cleared at the applicable cap of $325 per megawatt-day. The resulting capacity obligations totaled $16.4 billion, according to an auction analysis.

The grid operator said the price would have reached $554.72 without the cap. PJM’s independent market monitor estimated that data centers accounted for about $6.3 billion of the total capacity cost.

Those figures do not prove that every household increase came from AI. Generator retirements, transmission limits, fuel costs, market rules, and broader demand growth also influence electricity prices.

They do show that load growth can raise the value of scarce capacity before developers add enough new supply. That timing problem is the strongest challenge to Kimmelman’s argument.

A new data center can obtain servers and begin construction faster than a utility can complete a major transmission line. New gas generation also requires turbines, pipeline access, permits, financing, and an interconnection agreement.

During that gap, existing generators gain pricing leverage. Customers can pay more for reliability while investors work on capacity that will arrive later.

There is another uncertainty: announced demand may exceed realized demand. AI companies have strong incentives to reserve power early because electricity access can determine which data-center projects proceed.

Utilities have a different incentive. They earn regulated returns on approved infrastructure investments, although regulators review whether those investments are prudent.

The combination can inflate forecasts. Developers reserve more capacity than they ultimately need, and utilities propose construction based on those requests. Ratepayers face risk if contracts do not hold developers accountable.

Political resistance follows when residents see higher bills but not the promised long-term savings. Local communities also weigh water use, noise, land requirements, air emissions, tax incentives, and the number of permanent jobs created.

That resistance has moved beyond neighborhood planning hearings. State officials increasingly question whether data centers receive favorable treatment while households absorb system costs.

The customer-benefit case needs more than a forecast showing higher electricity sales. It needs evidence that residential and small-business rates fall below what they would have been without the project.

That counterfactual is difficult to establish. Bills can rise even when a data center reduces the size of the increase. Conversely, a utility can claim broad economic benefits while assigning significant infrastructure costs to other customers.

Regulators should therefore publish assumptions, contract protections, and cost-allocation methods. Confidential commercial terms may require limits, but the public still needs enough information to understand who carries the downside.

Kimmelman’s claim is best treated as a testable proposition, not a settled result. The evidence must appear in rate cases, capacity prices, construction schedules, and customer bills.

ECP Is Betting on New Supply, Especially Near the Load

ECP’s strategy attempts to solve the grid conflict by building generation beside large customers, but that approach introduces fuel and emissions tradeoffs.

ECP and ADQ announced a partnership in March 2025 aimed at investing in power generation and infrastructure for data centers and other industrial consumers. The structure linked international capital with ECP’s experience in US energy markets.

The partners said they intended to provide a combined initial commitment of $5 billion. Their broader ambition covered more than $25 billion of investment and a potential project pipeline totaling 25 gigawatts.

Kimmelman told Axios that a major objective involved gas-fired generation co-located with data centers. Co-location places generation near the customer, reducing dependence on distant power plants and congested transmission paths.

The power partnership planned to focus first on the United States. Kimmelman identified Texas and Ohio as potential candidates because projects need gas access, skilled workers, and a supportive regulatory environment.

This approach strengthens his argument about additionality. A data center paired with a newly constructed plant does not simply claim electricity from the existing fleet.

It can also reduce pressure on the transmission system, depending on the project’s configuration. A campus that generates most of its electricity on-site may require a smaller grid connection than one relying entirely on utility service.

However, co-location does not remove every system cost. Data centers still need backup arrangements, and generation facilities require maintenance. A campus may depend on the grid when an on-site plant is unavailable.

Utilities must price that backup service correctly. If a facility pays only for occasional electricity but expects the grid to reserve enough capacity for its full demand, other customers can subsidize its reliability.

Natural gas offers operational advantages for AI facilities. Gas plants can provide firm output, meaning operators can schedule their availability rather than depend entirely on weather conditions.

Gas generation also creates exposure to fuel prices, pipeline constraints, methane leakage, and carbon emissions. Projects can combine gas with renewable generation and batteries, but storage duration remains important when a facility expects continuous power.

Kimmelman has said gas can be supplemented selectively with renewables and storage. The language matters because he is not claiming that wind, solar, or batteries alone will support every planned AI campus.

That position differs from technology companies’ broader carbon-free energy goals. Some hyperscalers are pursuing nuclear power, advanced geothermal systems, renewable contracts, and longer-duration storage alongside gas generation.

The competition among these routes will shape the real cost of AI infrastructure. Gas can arrive faster in some markets, although turbine backlogs and pipeline approvals can cause delays. Nuclear and geothermal projects offer lower operational emissions but often carry longer development timelines.

None of these options escapes the need for firm contracts. Investors financing a dedicated plant need confidence that the data-center buyer will remain for the asset’s useful life.

ECP’s strategy therefore converts AI demand into a financing problem. The firm is betting that hyperscalers value reliable power enough to support new generation through long-term commitments.

If that bet succeeds, new supply can relieve pressure on constrained markets. If demand weakens, investors, utilities, or ratepayers will dispute who must absorb the remaining costs.

Three Signals Will Decide Whether Customers Receive the Gift

The next phase must be judged through executed contracts, completed generation, and measured rate outcomes rather than larger project announcements.

The first signal is the quality of large-load contracts approved by state regulators. Minimum payments, exit fees, credit requirements, and long terms would strengthen Kimmelman’s case.

Those provisions make data-center developers carry the financial consequences of changing their plans. Weak commitments would leave more risk with utilities and existing customers.

Watch whether regulators establish separate rate classes for data centers. A dedicated class can make infrastructure and capacity costs easier to identify, although the label alone does not guarantee fair allocation.

The second signal is the amount of new generation that reaches commercial operation near planned campuses. Announcements and interconnection requests are not operating capacity.

Projects must secure turbines or other equipment, fuel supply, permits, financing, and construction resources. Delays would extend the period when new demand competes for existing power.

The relevant comparison is not simply gigawatts announced against gigawatts requested. Observers should compare firm data-center load with generation that has financing, permits, equipment, and a credible completion schedule.

The third signal is the direction of residential rates in the most concentrated markets. PJM capacity prices, state utility cases, and customer bills can reveal whether new revenue is spreading costs or intensifying scarcity.

A higher bill does not automatically disprove the customer-benefit argument. Analysts must compare it with the bill expected without the data center, while accounting for fuel costs and other investments.

Still, repeated increases tied to new capacity and transmission would weaken Kimmelman’s claim. Lower average system costs, combined with strong developer protections, would support it.

Readers should also distinguish electricity demand from electricity availability. AI developers can announce ambitious computing plans, but the power system decides how quickly those plans become operational.

That constraint affects more than utilities. Developers choosing cloud regions, businesses budgeting for AI services, and workers using AI products all depend on the cost and reliability of the underlying infrastructure.

Tracking those connections requires more than saving individual headlines. A structured personal knowledge system can connect project announcements with later permits, rate decisions, delays, and completed capacity.

Google News surfaced a provocative claim, but the headline is only the opening evidence. The useful question is whether each new campus brings enough generation, contractual protection, and durable revenue to improve the system around it.

Over the next several months, follow approved utility contracts first, operating generation second, and residential rate evidence third. If all three align, Kimmelman’s “gift” will look like an investable model. If they diverge, customers will know who received the benefit and who received the bill.