Estimated share of U.S. electricity used by data centres in 2023
The electricity story is not simply that AI uses more power. It is about where new demand lands, who finances grid upgrades and how reliable capacity arrives on time.
Distinguish observed facts from projections, arguments and interpretations.
What the numbers establish
A DOE-supported study estimated that U.S. data centers accounted for about 4.4% of U.S. electricity use in 2023, up from 1.9% in 2018. It projected a potential 6.7% to 12% share by 2028. Those latter figures are scenarios, not meters recording future consumption. The large range reflects uncertainty in computing investment, server efficiency, cooling and operational patterns. The correct takeaway is a potential major load-planning problem, not a claim that every community will see the same increase.
The geography of electricity is the crucial detail
National annual electricity figures conceal a local infrastructure problem. A new cluster of data centers can create demand in a particular grid territory much faster than transmission lines, transformers or power plants can be permitted and built. Even when sufficient electricity exists nationally, congestion at a substation or connection point can delay a project. The DOE has identified transmission bottlenecks and long interconnection queues as central constraints. This is why regional utility filings matter at least as much as national forecasts.
Who pays for the build-out?
The commercial question is whether large industrial customers bear the incremental cost of the connections and resources they require, or whether part of the cost flows through to the wider customer base. Different utilities and state regulators can make different decisions. Any article claiming AI automatically raises every household bill ignores tariff design, existing generation reserves and regulatory cost-allocation decisions. Watch approved utility rate cases and the contracts used for new large loads.
Potential alternatives — and their limits
Power companies and data-center operators are considering firm generation, storage, on-site energy and flexible computing loads. Shifting a noncritical computing task away from peak hours could reduce pressure on the grid. But not every AI or cloud workload can be interrupted: service-level commitments, latency and system reliability impose constraints. Flexibility should be measured by dependable peak reduction, not by marketing claims about green power purchases.
What would change this assessment?
Track whether projected facility applications become signed utility contracts, the timing of new grid connections, commission-approved cost recovery and realised data-center load. DOE projections are useful for setting a range; local regulatory documents reveal who will actually see the costs and benefits. Without that second layer, a national demand forecast cannot tell an individual family what its next power bill will be.
What the headline leaves out
The national figure of 4.4% is a historical estimate of US data-center electricity consumption in 2023, while the widely quoted 6.7–12% is a scenario for 2028. Neither number establishes what a specific county’s power bill will be. The October 2026 regulatory dispute makes the harder question urgent: when a data center requests hundreds of megawatts, what infrastructure must be built, which obligations are legally binding, and who pays if projected demand fails to appear?
What the original numbers actually say
Read the period, units and definition in each row. These are distinct measurements, not personal forecasts.
| Measurement | Earlier period | Latest reference | What it means — and does not mean |
|---|---|---|---|
| US data-center electricity, 2014 | — | 58 TWh | Historical estimate in the DOE-supported Berkeley Lab study, not an AI-only figure. |
| US data-center electricity, 2023 | — | 176 TWh / about 4.4% of national consumption | Observed-year estimate covering all data-center uses, not just generative AI. |
| US data-center use, 2028 scenario | — | 325–580 TWh/year | Projection range published by DOE/LBNL, not metered 2028 electricity demand. |
| Share of national electricity in 2028 scenario | — | 6.7–12% | Percentage depends on the denominator and possible growth in the wider electricity system. |
| US grid operators addressed by FERC June 2026 action | — | Six jurisdictional RTOs/ISOs | FERC directed operators to justify or propose tariff changes for integrating large electrical loads. |
| PJM area population mentioned in FERC deliberations | — | About 67 million residents | PJM spans several states; a regional system statistic, not the number of data-center users. |
| October 9, 2026 PJM order | — | Large-load registry accepted in part; curtailment proposal rejected in part | The regulator approved tracking information, but did not approve a sweeping proposed emergency shutoff rule. |
| Hypothetical continuous facility load | — | 100 MW running all year → 876,000 MWh | Pure arithmetic at 100% utilisation, not a figure from an individual proposed data center. |
The deeper investigation
1. First separate measured demand from a forecast for 2028
The Department of Energy released Berkeley Lab’s assessment of rising US data-center electricity requirements in December 2024. It estimated 176 terawatt-hours of consumption in 2023, or roughly 4.4% of total US electricity, against about 58 TWh in 2014. Its 2028 range is 325–580 TWh, equivalent to approximately 6.7–12% of projected national usage. Those future figures embody uncertainty about chip efficiency, servers installed, cooling technology, data-center utilisation, and overall demand in the rest of the country. They are not a meter reading from the future. Nor should the entire 2023 consumption figure be described as electricity consumed only by generative AI: data centers support ordinary cloud storage, video distribution, enterprise computing and numerous other functions. Data-center energy demand has genuinely become a larger planning issue, but rigorous reporting describes the baseline year, model assumptions and range, instead of treating the highest 2028 figure as an inevitable outcome.
2. A power grid is constrained locally even when national generation looks sufficient
A national estimate aggregates hundreds of locations where computing campuses are connected to different power utilities and transmission systems. One region can have extensive available capacity while another suffers transformer bottlenecks, limited local generation, a full connection queue or delays securing rights to build new lines. Electricity must physically reach a substation with enough capacity at the right moment, not simply exist somewhere else in the national annual ledger. This is why the same 100 MW load application can be routine in one location and very expensive in another. Connections require studies of steady demand, contingency conditions, voltage and reliability standards, and may involve new substations, conductors and generation arrangements. Local electric utility and state commission documents are essential for understanding those costs. A national figure such as 4.4% is useful to assess scale; it cannot show who needs a local upgrade or whether a customer in a particular ZIP code will pay more.
3. The June 2026 FERC action was a regulatory instruction, not a uniform new tax
On June 18, 2026, the Federal Energy Regulatory Commission issued show-cause proceedings requiring six regional electricity operators under its jurisdiction to defend or reform the tariffs through which large loads connect to the grid. FERC framed the action as an attempt to speed access for large electricity users while protecting existing customers from unjustified cost transfers. A show-cause order does not automatically establish one nationwide dollar charge per megawatt or make every proposed data-center project financially responsible for every new power station. The operators, utilities and regulators must work through applicable tariffs, proceedings and jurisdictional responsibilities. Retail rate structures often involve states, while FERC oversees specified wholesale and interstate transmission questions. Describing this action as a federal ban on AI infrastructure is as mistaken as describing it as permission for unlimited new load with costs socialised across all households. The actual effect lies in filed tariff terms and subsequent orders.
4. October’s PJM decision illustrates the difference between a registry and a shutdown policy
PJM Interconnection covers a major part of the mid-Atlantic and Midwest and has faced exceptional demand requests from large computing facilities. On October 9, 2026, FERC issued an order in docket ER26-3515 accepting certain proposed tariff changes but rejecting others. Contemporary reporting described approval for a registry of large electricity loads while rejection of PJM’s proposed interim curtailment structure. The difference is fundamental. A database recording where large loads plan to operate, when they expect service and their projected requirements can improve forecasting and oversight. It does not by itself give a regional grid operator unrestricted authority to disconnect any particular customer. Curtailment rules require careful jurisdictional and contractual treatment, and important questions remain about how to prioritise hospitals, homes, manufacturers and other customers in an emergency. A responsible October article reports exactly what was accepted and rejected, and does not write as if the rejected mechanism were already law.
5. The argument about bills is really about recovering costs if a forecast is wrong
Consider a utility asked to install a specialised transmission connection and substation because a developer predicts a large, continuous demand. If that developer signs a robust long-term service agreement and pays the costs that only it causes, a significant portion of the financial risk may be kept away from existing ratepayers. If the developer later abandons or downsizes the project and contracts do not fully cover costs, some investment may become stranded. The applicable tariff, guarantees, minimum monthly demand, security collateral and exit charges determine who pays. Meanwhile, other customers may benefit from a line that adds reliability or from increased system sales that spread fixed costs. Both directions are possible. The Commission’s debate therefore concerns cost causation, risks and benefits, rather than an automatic identity between AI load growth and a precise percentage increase in residential prices. Assess utility rate cases, capital expenditures and customer contracts before assigning household costs.
6. Capacity in megawatts and electricity in megawatt-hours answer different questions
A facility requesting 100 megawatts has described a power level, not its annual energy use. If it draws a constant 100 MW for 8,760 hours in an ordinary year, it uses 876,000 megawatt-hours, or 0.876 TWh. But if average electrical demand is only 60 MW, usage is approximately 525,600 MWh. Billing rules may include both consumed energy and charges associated with reserved capacity or maximum demand. A grid operator must also consider coincidence with regional peak load; a data center whose load is steady throughout summer afternoons can affect system planning differently from a manufacturing plant that can turn down during emergencies. The typical headline about a “one-gigawatt data center” therefore needs a clear definition: connected requested power, actual observed power, planned ultimate build-out and hourly profile may be very different quantities. This technical distinction is central to avoiding inflated comparisons with household annual energy consumption.
7. Can batteries, onsite generators or flexible computing solve the problem?
Some computing jobs can be scheduled for periods when network demand is lower, while a customer may use a battery or onsite power plant to reduce peak grid imports. Whether that actually saves network infrastructure depends on reliability. The operator needs evidence of how much power can be reduced, for how many hours and under what conditions; a marketing promise of renewable procurement does not make a server farm dispatchable. Generators produce emissions and have fuel-supply, noise and permitting implications, while batteries have finite energy storage and may need charging from the grid before and after a peak event. Critical online services may be unable to accept interruptions without violating service guarantees. Flexible AI training tasks may be easier to shift than real-time inference or essential cloud systems, but even those possibilities must be established by site-specific engineering and operating agreements. A credible proposal quantifies dependable peak reduction instead of quoting an annual renewable-energy certificate total.
8. Why the economic-development case deserves consideration alongside residents’ concerns
Large computing projects can bring construction activity, property-tax revenue, investment in transmission facilities and demand for local services. They may also compete for scarce power capacity, water, land and financing while generating relatively modest numbers of permanent on-site jobs after construction. The mix of benefits and costs depends on location and the developer’s obligations. Local officials should ask whether promised tax revenue is net of incentives and public infrastructure costs, whether there are clawback provisions if a project is cancelled and what employment commitments are contractually enforceable. Residents are entitled to examine ratepayer exposure, reliability, emergency plans and environmental impacts. Supporters are entitled to present evidence about investment and improved infrastructure. Neither position should be dismissed through national stereotypes about AI. The practical question is whether the public agreement makes the project’s real beneficiaries pay the incremental costs and receive clearly specified services.
9. What a useful 2027 update would actually verify
The DOE 2028 projection will be tested by future national energy data, but local commitments should be monitored sooner. Report requested versus executed electrical-service agreements, the megawatts of new load actually energised, substation construction progress, reserve margins and approved regional capacity procurements. Check whether PJM or other operators file revised curtailment or registry rules and whether FERC accepts them, paying attention to docket numbers and the exact language of final orders. For households, follow the local utility’s approved tariff and rate-case decisions instead of deriving individual electricity inflation from national terawatt-hour forecasts. For the environment, compare local generation mix, firm capacity, cooling-water demands and verified energy use. Future reporting must distinguish forecasts, signed commitments and actual measured consumption. When new evidence supersedes the 2024 Berkeley Lab study or the October 2026 regulatory record, date the revision and explain precisely which earlier assumption changed.
A 100 MW data-center proposal: capacity, energy and who might pay
Hypothetical engineering and financial examples, not facts about any company, power project or approved rate case. The 176 TWh DOE figure is historical, but all plant-level assumptions and prices below are invented solely for transparency.
Contracted or requested power capacity is distinct from electricity actually consumed.
0.876 TWh/year, assuming no maintenance or reduction in electrical draw.
An example load profile; not a model of a particular AI training centre.
Illustrates how an assumed energy price and usage could form one piece of a bill. Excludes demand charges, taxes and infrastructure recovery.
This is a made-up capital cost, NOT a quotation from PJM or a utility.
Illustrates one possible contract outcome, not the standard legal rule for every grid connection.
Allocation across customers, regulatory timing, utilisation and offsets from new revenues would all matter.
The worked numbers are not US electricity rates, PJM price forecasts or utility tariffs. Megawatts are power; megawatt-hours are energy. Cost allocation is decided in applicable contracts and regulatory proceedings, not by this mathematical scenario.
The arguments, and what evidence they need
These are possible interpretations, not invented quotations from people we did not interview.
Residents concerned about electricity affordability
A large campus should pay for the facilities that its own demand makes necessary.
What must still be established: Utility tariffs, incremental-cost studies and binding customer guarantees are needed to establish whether that is happening.
Data-center operator seeking a predictable connection
Long delays and unclear curtailment rules create business risk and could slow investment.
What must still be established: Claims about public benefits should be matched against independently verifiable employment, tax and network commitments.
Grid regulator prioritising reliability
Better forecasting and cost-causation rules can protect existing users and allow viable large loads to connect.
What must still be established: A load registry is only a planning tool; operating and emergency authority must be decided separately.
Frequently misunderstood details
Did American data centers use 12% of US electricity in 2023?
No. DOE estimated about 4.4% in 2023. The 12% figure is the high end of a projection for 2028.
Are those DOE figures only about ChatGPT or generative AI?
No. They cover all data-center electricity, including cloud and conventional computing, not solely generative AI.
Does a 100 MW project always use 876,000 MWh per year?
Only if its electrical draw remains at 100 MW continuously for 8,760 hours. Actual loads may differ.
Will my household bill rise because of new data centers?
Not necessarily. Check the local utility’s approved tariffs, cost allocation and the project’s signed service obligations.
Did FERC approve disconnecting all PJM data centers in October 2026?
No. Its October 9 order accepted tariff changes in part, including a large-load registry, while rejecting other proposed measures. It was not a blanket power-disconnection order.
Can a data center operate without drawing grid power during peak demand?
Sometimes, depending on onsite supply, storage and actual workload flexibility. Only verified dependable performance counts for reliability planning.
What to monitor
Conclusions can change as new numbers and decisions arrive. Monitor these indicators.
- 01Regional utility interconnection queues and committed megawatts
- 02State utility rate-case decisions on large-load customers
- 03New grid capacity and demonstrable peak-demand flexibility
Source documents
Read the underlying records to check units, reporting dates and subsequent corrections.