
Data Centers, Energy Use, and Local Backlash
Electricity demand and infrastructure
POWER DEMAND: DEFINE THE BOUNDARY
SCOPE Data centers are not synonymous with AI. · Separate historical estimates, forecasts, and proposals.
Three different questions
Electricity analysis must distinguish how much energy a facility consumes, how much power the grid must deliver at a critical moment, and who pays for the infrastructure and services required. These questions interact, but none can be answered by a single campus capacity announcement. Annual consumption describes an energy quantity; peak demand describes a rate of use; a tariff determines financial obligations.
The IEA’s 2026 outlook covers electricity consumed by data centers, including services beyond artificial intelligence. Its central global projection is approximately 950 TWh in 2030, compared with an estimated 485 TWh in 2025.[27] Presenting the entire figure as AI consumption would misstate the boundary. Enterprise computing, storage, networking, and conventional cloud services remain part of the sector.
An evidence hierarchy
This report treats measured or modeled historical consumption separately from forward projections.
A utility forecast is a planning input, not a promise that every proposed campus will operate. A developer’s announced maximum demand is not the same as commissioned load, and an interconnection request is not a completed electrical connection.
For each claim, record geography, year, publication vintage, unit, facility coverage, and status. Comparisons become meaningful only after these fields align. A global estimate and a United States forecast cannot be combined into a growth series. Nor should a server-only estimate be compared directly with a whole-facility total.
The analysis that follows uses published evidence to establish scale, then develops practical tests for forecast credibility, reliability, and cost allocation. Illustrative calculations are labeled. They explain mechanisms rather than predict a particular household bill or certify a particular development as beneficial.
MW, MWh & THE CAPACITY ILLUSION
MEASUREMENT Power is a rate. Energy is an accumulated quantity. · Illustrative calculations use 8,760 hours per year.
Translate the announcement
A megawatt measures power at a moment; a megawatt-hour measures energy delivered over time. A hypothetical facility drawing 100 MW continuously would consume 876,000 MWh, or 0.876 TWh, in a non-leap year. At an average draw of 60 MW, its annual consumption would be 0.526 TWh after rounding. The same advertised connection can therefore support very different annual energy outcomes.
Specify whether a number refers to information-technology equipment, total facility demand, contracted service, transformer capacity, or a future campus buildout. Cooling and electrical losses make facility demand different from IT demand. Redundant equipment ratings can also exceed the actual simultaneous load. Adding every equipment nameplate can create a misleading impression of grid requirements.
Peak coincidence matters
A steady 100 MW customer and a 100 MW customer operating only during system peaks impose different energy volumes but can create similar peak obligations.
Conversely, a large annual consumer that reliably reduces demand during constrained hours may require less incremental capacity than its energy total suggests. This is why multiplying annual TWh by a retail rate cannot reveal the full cost of service.
EIA distinguishes generation, consumption, and electricity prices in its statistical products; its capacity-addition releases describe generating capability, not annual energy output.[43] The report’s practical rule is to keep separate columns for MW, MWh, and dollars, with explicit conversion assumptions.
For a campus approaching service, request an hourly ramp profile, expected utilization, contingency imports, and the contractual maximum. Record the date each phase becomes electrically usable. These details turn a publicity figure into a planning input and reveal whether the utility is being asked to reserve infrastructure long before the customer can use it.
GLOBAL OUTLOOK: LARGE GROWTH, LOCAL CONSTRAINTS
GLOBAL EVIDENCE 2025: 485 TWh estimated. · 2030: 950 TWh projected in IEA’s 2026 outlook.
Read the current vintage
The IEA’s 2026 update projects data-center electricity use roughly doubling from 485 TWh in 2025 to 950 TWh in 2030, approximately 3% of global electricity demand. It projects AI-focused data-center consumption growing faster, roughly tripling over that interval. These are different measures: the sector total includes non-AI activity, while the AI-focused category describes a subset.[27]
The earlier 2025 Energy and AI report estimated 415 TWh in 2024 and projected about 945 TWh in 2030.[28] The similar endpoint does not mean that nothing changed. The observed starting point, near-term constraints, and evidence base were updated. A chart should identify which vintage supplies each series rather than silently combining forecasts.
Global percentages do not allocate capacity
A small global share can still create a large local challenge when demand concentrates behind the same substations or transmission corridors.
World electricity supply cannot be delivered indiscriminately to any campus. The relevant constraint may be a local transformer, regional reserve requirement, construction schedule, or permission to operate backup equipment.
A useful comparison therefore has two levels. The global view establishes the scale of the industry and the range of possible growth. The local view examines the additional load on a specific network and the resources available at the hours when that network is constrained.
Neither level alone settles affordability. A globally modest load can require expensive local reinforcement; a large new customer can also contribute substantial revenue where spare capability exists. The analytical question is not whether data centers are large in the abstract, but whether their particular location, timing, and service requirements create costs that their payments cover.
UNITED STATES: THE UPDATED REFERENCE CASE
U.S. EVIDENCE 2024: 192 TWh and 4.7%. · 2030 reference: 649 TWh and 11.8%.
Replace obsolete shorthand
Berkeley Lab’s 2025 Update, published in 2026, estimates a 2030 reference case of 649 TWh, or 11.8% of total United States electricity use. A November 2026 laboratory presentation reports the historical 2024 estimate as 192 TWh, or 4.7%. The update’s scenarios span 9.5% to 15.3% of total electricity use in 2030.[29][30]
These figures should not be presented as a guaranteed outcome or a confidence interval. They reflect modeled assumptions about equipment and operation. The earlier 2024 report’s widely cited 325–580 TWh range referred to 2028, a different endpoint and forecast vintage.[31] A responsible chart retains these distinctions rather than treating every published number as an interchangeable estimate of the same future.
What the range does and does not show
Scenario variation reveals that energy demand is sensitive to assumptions, but it does not supply a probability for each outcome.
A utility may face a very different exposure depending on where the projected activity occurs. National totals do not identify which projects arrive, their hourly profiles, or the infrastructure required to connect them.
The practical implication is to use national research as a reason to strengthen regional planning, not as a substitute for project verification. Require developers to connect their requested service to phased construction, equipment procurement, and credible occupancy or computing demand.
For public discussion, keep three statements separate: the United States has already experienced substantial data-center electricity use; credible models project further growth; the magnitude and geography remain uncertain. This formulation recognizes the investment challenge without converting a scenario into an inevitability or dismissing real demand because individual projects may slip.
AI LOAD IS NOT ALL COMPUTING LOAD
ATTRIBUTION Separate workload, hardware, and facility boundaries. · No universal electricity-per-prompt factor is used here.
Why attribution is difficult
AI training, AI inference, conventional cloud computing, storage, and communications can share facilities and supporting systems. A building’s meter measures their combined demand, including cooling and electrical overhead. Allocating that total to AI requires additional information about equipment, utilization, workload scheduling, and the allocation of common infrastructure.
Berkeley Lab’s modeling framework estimates sector electricity demand through equipment and operating assumptions, while the IEA separately discusses AI-focused facilities within broader data-center demand.[32] These approaches are useful but do not make every campus megawatt an AI megawatt. A facility designed for accelerated computing can also operate below design load during deployment or serve several types of work.
Avoid the prompt multiplication shortcut
A fixed electricity estimate per chatbot request can be misleading when requests differ in model, output length, modality, hardware, batching, and system utilization.
Multiplying a convenient average by a speculative request count compounds uncertainty. It can also omit training, idle equipment, networking, and facility overhead, or count them twice.
For an operator, a more defensible accounting system measures facility electricity and reconciles it with equipment-level telemetry. Assign shared overhead using a stated method and show the unallocated remainder. Track changes in useful work as well as energy, because improved efficiency can coexist with higher total consumption if activity grows faster.
For a policymaker, the immediate grid question is generally the facility’s actual and contracted electrical behavior, regardless of which application caused it. AI attribution matters for industrial strategy and environmental accounting, but protection of the power system should not depend on a marketing label. Apply comparable technical and financial requirements to comparable large-load behavior.
FORECASTING: FROM REQUESTS TO CREDIBLE LOAD
FORECAST QUALITY An inquiry is not a commitment. · Track project maturity and actual ramps.
Build a project ledger
Large-load forecasting needs more than a list of requested megawatts. Record the customer, location, requested phases, expected energization, construction status, financial security, and whether alternative sites represent the same underlying project. Otherwise, competing location inquiries can appear as additive demand even when only one site will be built.
PJM’s January 2026 forecast reduced near-term peak expectations relative to its prior forecast through 2032, citing updated economic and electric-vehicle assumptions and improved vetting of large-load adjustments.[33] This is evidence that screening changes the planning result. It is not evidence that data-center growth disappeared or that every earlier request was fictitious.
Use observable milestones
Assign project categories such as inquiry, executed service agreement, financed construction, commissioned phase, and operating load. The categories should describe evidence, not subjective enthusiasm.
Update them when deadlines slip or deposits expire. Keep the original forecast so reviewers can see whether changes reflect new information or unexplained optimism.
Avoid a single discount applied indiscriminately to all proposed demand. Two projects with identical requested capacity may have very different completion risks. A physically advanced campus with secured equipment is not equivalent to an undeveloped land option. Nor does a strong sponsor eliminate risks from unavailable transmission or uncertain tenant demand.
A credible forecast reconciles each new vintage with actual load growth. Publish aggregate additions, withdrawals, delays, and revisions without disclosing sensitive customer details. This allows regulators and communities to distinguish genuine growth from a constantly moving pipeline. The purpose is not to force certainty where none exists, but to make uncertainty visible before it becomes an irreversible infrastructure commitment.
UNCERTAINTY: DESIGN FOR MORE THAN ONE FUTURE
SCENARIO DESIGN Test delayed load and accelerated load. · National projections are not local probabilities.
Separate the sources of uncertainty
Demand can diverge because fewer campuses are built, because construction is delayed, or because operating utilization differs from expectations. These mechanisms have different financial consequences. A delayed project may eventually repay infrastructure costs but create years of carrying costs; a canceled project may leave an asset with little alternative use.
PJM’s 2026 forecast anticipates average annual summer peak growth of 3.6% over ten years, substantially above the approximately 0.3% projected in 2021.[34] That contrast shows how rapidly planning expectations can change. It does not establish an error distribution or justify assuming the latest forecast is exact.
Stress the investment sequence
Test a base case, a slower-ramp case, and a concentrated high-demand case. Vary commissioning dates separately from ultimate capacity.
Examine whether the proposed network can serve existing customers if new load arrives faster than generation, and whether those customers face stranded costs if load arrives later.
The analysis should also consider common causes. Several projects may depend on the same chip supply, financing conditions, or transmission upgrade. Treating their completion probabilities as independent can understate correlated delay. Conversely, geographic substitution can reduce demand in one service area while leaving national demand unchanged.
Stage investment where technically practical. Early work can preserve options without committing every downstream asset immediately. Identify which milestones justify larger expenditures and which costs become unrecoverable. A useful decision is not simply the cheapest outcome under the central forecast. It is an investment sequence that remains acceptable across plausible outcomes while maintaining reliability obligations. The most important output may be the conditions that trigger the next stage, rather than a single precise megawatt forecast.
LOCATION: THE GRID IS NOT A NATIONAL SOCKET
GEOGRAPHY Study the relevant electrical boundary. · Regional averages can hide a local constraint.
The connection point changes the answer
Two identical campuses can have different system impacts because their connection points differ. One may sit near existing generation and uncongested transmission; another may require a new substation and reinforcement across several jurisdictions. Local land availability therefore says little about electrical readiness.
DOE’s transmission-needs work evaluates constraints and congestion at regional and interregional scales.[35] Its significance for data-center siting is methodological: assess the network that must actually carry the power. National capacity totals do not establish that a specific location can receive a specified quantity under stressed conditions.
Compare delivered service, not headline price
A site-selection comparison should include the timing and firmness of service, upgrade obligations, congestion exposure, and the cost of interruptions.
A low published energy rate can be outweighed by a long wait for electrical infrastructure or restrictions on operating during critical hours. Conversely, a more expensive site may provide a more credible path to usable capacity.
The developer should identify which requirements are essential and which can change. Latency, customer proximity, water availability, workforce access, and network connectivity may limit geographic flexibility. Grid planners should not assume all computing can relocate freely, but developers should not assume the grid must accommodate every preferred location on the same terms.
Communities need the same locational clarity. Separate on-site equipment, nearby network work, and broader regional upgrades. Explain which assets primarily serve the campus and which provide shared benefits. This prevents a debate about one visible building from obscuring the larger electrical footprint and makes it possible to compare alternative siting choices fairly.
TRANSMISSION: CAPACITY, BENEFITS & TIME
NETWORK INVESTMENT A transmission project can serve multiple purposes. · Cost allocation should follow demonstrated benefits.
Planning beyond a single customer
FERC’s Order No. 1920 framework calls for long-term regional transmission planning using at least three distinct scenarios, a horizon of at least twenty years, and reassessment at least every five years. Its cost-allocation framework connects assigned costs to estimated benefits rather than assuming every project has one beneficiary.[36]
For data-center growth, this matters because a network upgrade can address several needs at once: delivery to a new load, relief of existing congestion, replacement of aging equipment, and improved access to generation. Allocating the entire project to one customer can overstate that customer’s responsibility; allocating none can shift customer-driven costs to others.
Separate need, selection, and payment
First establish the electrical need. Then compare technically credible alternatives. Finally determine which beneficiaries should pay under the applicable rules.
Reversing this sequence can turn cost allocation into a negotiation over a project whose necessity has never been demonstrated.
Construction schedules also require scrutiny. Study completion, siting approval, land access, equipment procurement, and commissioning are separate milestones. An approved plan does not make a line immediately available, and a funding agreement does not resolve every permitting issue. Maintain a schedule with dependencies rather than one optimistic completion date.
The report’s proposed test is to ask what would be built without the data center, what changes because of it, and what benefits the changed project produces for others. Document those answers before assigning costs. Update the assessment if the campus changes size or timing. Shared infrastructure can be valuable, but the claim of shared benefit should be supported by analysis rather than used as a blanket justification for spreading costs.
INTERCONNECTION: TWO QUEUES, ONE TIMELINE
DELIVERY RISK Generation requests are not completed plants. · Large-load service rules remain region-specific.
Supply and demand must meet in time
A campus may obtain land and begin building before the generation and network needed to serve it are ready. Conversely, a proposed generator may spend years in studies without entering service. Berkeley Lab’s Queued Up 2026 edition reports that, in regions with available data, projects completed in 2025 had a median interval of more than five years from interconnection request to commercial operation.[37]
That statistic describes completed generating projects, not data-center connections and not a guaranteed waiting time for every future plant. It nevertheless demonstrates why queued capacity cannot be treated as immediately available supply. Withdrawn projects and incomplete studies need separate treatment.
Regulatory reform is not instant capacity
On June 18, 2026, FERC issued six regional show-cause orders asking grid operators to justify or reform large-load connection rules.[38] Those proceedings should not be described as a single nationwide final tariff or as proof that every proposed customer now has accelerated service.
Current project terms depend on the relevant orders, compliance filings, and effective tariffs.
For planning, align three schedules: load commissioning, generation availability, and network readiness. Identify the earliest date all required elements are available together. If interim service is proposed, specify its restrictions, duration, and fallback when the permanent solution slips.
A credible agreement also defines what happens when only one side is ready. The operator may need to limit the campus ramp, provide conditional service, or defer energization. Those choices should be explicit before capital is committed. A faster administrative process is useful, but it does not itself manufacture transformers, complete transmission lines, or supply dependable generation during a system shortage.
RESOURCE ADEQUACY: ENERGY IS NOT ENOUGH
RELIABILITY Annual energy balance does not guarantee peak service. · Accredited capacity differs from nameplate capacity.
Cover the difficult hours
Resource adequacy asks whether enough dependable supply and demand-side capability will be available during periods of system risk. A system can generate sufficient annual energy yet face shortages during a cold evening, a heat wave, or an extended outage. Data-center service commitments therefore require more than an annual purchase of an equivalent quantity of renewable electricity.
PJM’s December 2025 capacity auction for the 2027/28 delivery year procured 134,479 MW of generation and demand-response capacity on its reported unforced-capacity basis, with a clearing price of $333.44 per MW-day at the approved cap.[39] The auction result is a forward capacity-market measure, not an electricity-consumption figure or a retail rate.
Evaluate the portfolio
A reliable supply plan examines the joint performance of resources under stress.
Weather can affect demand, renewable output, fuel availability, and equipment outages simultaneously. Simply adding nameplate ratings ignores these dependencies. Storage duration, fuel assurance, transmission deliverability, and demand-response performance all influence the usable contribution.
For a new campus, identify the obligation imposed on the serving utility or supplier. Does the customer require uninterrupted firm service, accept curtailment, or provide resources that are recognized under the applicable adequacy rules? A private claim to be self-sufficient does not automatically reduce a public system obligation.
Treat adequacy assessments as risk analyses, not predictions of a particular blackout. A reserve shortfall signals an elevated planning concern; it does not identify the exact hour an outage will occur. Conversely, an acceptable aggregate reserve margin does not eliminate local delivery constraints. Decision-makers need both the regional supply picture and the campus-specific electrical study.
DYNAMIC RELIABILITY: WHEN LOAD DISAPPEARS
OPERATING BEHAVIOR Sudden load loss can destabilize balance. · Test ride-through and reconnection behavior.
A data center is an active electrical system
Large electronic loads can change demand rapidly during a disturbance. Protective equipment may transfer a facility to backup supply or disconnect it from the grid. When many facilities respond together, the power system suddenly has more generation than load. Reliability concerns therefore include abrupt demand reductions as well as insufficient supply.
NERC’s 2026 State of Reliability overview identifies a February 2025 Eastern Interconnection event involving 1,800 MW of data-center consequential load reduction following a transmission fault.[41] This is evidence of a specific operating phenomenon, not proof that all data centers behave alike or that every disturbance causes service failure.
Model the behavior that matters
Connection studies should use appropriate dynamic models and validated assumptions about protection settings, voltage response, backup transfers, and restoration.
A static hourly demand forecast cannot describe a rapid disturbance. The system operator needs to understand how much load may disappear, how quickly, and under what conditions it returns.
Reconnection deserves equal attention. Several facilities restoring demand simultaneously can create a second large change. Recovery procedures should coordinate the timing of return to normal service and verify that controls behave as expected after an abnormal event.
The report’s operating recommendation is to require evidence through commissioning tests, model validation, and periodic updates when equipment or settings change. Confidential technical information can be protected while aggregate reliability risks are communicated publicly. The central distinction is between a facility’s internal uptime and its behavior as part of the wider grid. A strategy that protects servers can still create a system-level challenge if its interaction with other facilities has not been studied.
FLEXIBILITY: VALUABLE ONLY WHEN DELIVERABLE
DEMAND RESPONSE Define the amount, duration, and recovery. · Technical potential is not a contracted grid service.
Identify the flexible component
Some computing work can be delayed, shifted geographically, or interrupted with limited consequences. Other work has strict response-time or continuity requirements. Cooling, networking, and safety systems also impose operating limits. A facility’s theoretical ability to reduce load is therefore not equivalent to an obligation that a grid operator can rely upon.
A practical flexibility product specifies the baseline, available megawatts, response time, maximum event duration, notice period, and restrictions. It must also describe rebound demand when deferred work resumes. Without these terms, a statement that a campus is flexible is difficult to evaluate and impossible to price consistently.
Match compensation to verified service
PNNL’s 2026 analysis examines battery storage as a means of reducing data-center cost shifts, emphasizing the relationship between incremental system costs and the new customer’s contribution.[42] The useful lesson is conditional: flexibility can improve the economics when it changes the actual cost to serve, not merely when it moves energy behind a different meter.
Test the proposed service under realistic conditions. Confirm telemetry, dispatch authority, baseline integrity, and performance during the hours of greatest need. Avoid paying twice for the same capability through overlapping programs or counting backup capacity that cannot legally or technically operate as assumed.
A flexible tariff should allocate both rewards and consequences. If the customer receives cheaper or earlier service in exchange for curtailment, failure to curtail cannot silently become an obligation imposed on everyone else. Clear enforcement protects reliability and the credibility of the arrangement. Properly designed flexibility can create a middle ground between immediate unrestricted service and an indefinite connection delay, but only if its limits are measurable and enforceable.
NEW SUPPLY: THE MIX AND THE MISMATCH
GENERATION PIPELINE Planned additions are not completed additions. · Storage shifts energy; it does not create fuel.
Interpret the capacity pipeline
EIA reported that developers planned 86 GW of utility-scale capacity additions for 2026, a record if realized. Solar represented 51% of the planned total, batteries 28%, and wind 14%.[43] These percentages describe nameplate additions in that release, not dependable capacity during every hour and not the share of a data center’s electricity that each technology will supply.
The distinction matters because different technologies solve different problems. Solar adds daytime generation. Storage can move energy across time within its power and duration limits. Dispatchable plants can provide output on demand only when fuel, equipment, permits, and operating conditions allow. No single nameplate total describes all these capabilities.
Match the load profile
A continuously operating campus needs a portfolio that covers low-generation periods as well as annual energy volume.
Evaluate hourly supply, transmission access, outages, and reserve obligations. When storage is part of the plan, include charging demand, conversion losses, degradation, and the possibility of consecutive stress events.
New generation can reduce the pressure created by new load, but the relevant question is additional deliverable capability on the required schedule. A contract for an existing plant may change who pays for its output without increasing system supply. A proposed plant can add supply only after it is financed, built, connected, and available.
The report recommends comparing portfolios on delivered service and risk, not on the cheapest isolated technology metric. Include the cost of backup, balancing, transmission, and delay. A low energy price is useful, but it is not a complete cost estimate for serving a demanding customer reliably throughout the year.
CO-LOCATION & BEHIND-THE-METER POWER
SERVICE BOUNDARY Physical proximity does not erase grid dependence. · Specify imports during generator outages.
The connection still matters
Locating a data center beside a generator can change its electrical and commercial arrangement, but it does not automatically eliminate effects on other customers. If an existing generator redirects output to the campus, less output may remain available to the wider system. If the campus imports during generator outages, it may still require substantial network capability.
FERC’s December 2025 PJM co-location action directed development of clearer service arrangements; subsequent June 2026 action modified parts of that framework and addressed compliance.[44][45] These are evolving regional rules. A project should be evaluated under its actual effective tariff and approved agreements, not a generic claim that co-location avoids ordinary system obligations.
Measure all operating states
Study normal operation, generator outage, maintenance, startup, and restoration. Specify the maximum grid import in each state and the protections that prevent unauthorized withdrawals.
A nominally self-supplied campus can create a large standby requirement if its on-site plant fails while computing continues.
Financial analysis should include standby service, fuel infrastructure, emissions controls, maintenance, and the consequences of generator unavailability. Physical separation from ordinary retail service does not guarantee economic independence. Nor does a private power contract establish that existing customers are protected from every network or market effect.
The appropriate question is what services the facility actually receives and what costs or benefits it creates. Where the arrangement genuinely reduces network requirements, charges should recognize that reduction under applicable rules. Where it continues to depend on the grid, the dependence should be visible and funded. This approach avoids treating either co-location or conventional service as inherently superior without examining the specific electrical configuration.
RETAIL RATES: UNPACK THE BILL
PRICE EVIDENCE Wholesale energy is one component. · A rate increase is not a complete causal explanation.
Follow the cost layers
Electricity bills recover several categories of expense. EIA identifies fuel, power-plant investment and operation, transmission and distribution, weather, and regulatory arrangements among the factors affecting prices.[46] Data-center growth can influence some of these channels, but it is not the only possible driver. Aging infrastructure, storm recovery, financing costs, and changes in fuel markets can also matter.
Separate the energy charge from capacity, network charges, riders, taxes, and fixed customer charges where the bill structure permits. A change in a wholesale auction price does not translate one-for-one into a household’s total bill. Suppliers may hedge purchases, tariffs may update with a lag, and customers have different usage patterns.
Rates and bills are different outcomes
A household bill can rise because the unit price increased, because consumption increased, or both.
Weather can change usage even if the tariff is unchanged. Comparing average monthly bills without controlling for consumption can therefore confuse price pressure with a hotter summer or colder winter.
For public reporting, show the applicable tariff, a constant-usage example, and the actual average bill separately. State whether figures are nominal or inflation-adjusted and whether they include all mandatory charges. Compare the same utility, customer class, and billing period rather than mixing statewide averages with an individual supply offer.
The evidence needed to attribute an increase to data centers is more demanding than observing that both grew. Trace the approved costs, market exposure, and allocation decisions. A transparent decomposition can identify genuine cost shifts without blaming the sector for unrelated expenses or overlooking effects that arrive through regional wholesale markets rather than a nearby substation.
CAPACITY PRICES: A SIGNAL, NOT A HOUSEHOLD BILL
MARKET EVIDENCE 2027/28: $333.44 per MW-day. · 2028/29: $325.00 per MW-day.
Understand the auction unit
PJM’s 2028/29 Base Residual Auction reported a capacity clearing price of $325 per MW-day, compared with $333.44 for 2027/28.[40][39] These are prices for capacity obligations in specified future delivery years. They are not dollars per megawatt-hour of energy, not national electricity prices, and not a direct percentage increase in residential bills.
Capacity-market outcomes reflect the interaction of demand forecasts, eligible supply, accreditation, auction parameters, and administrative constraints. A price at a cap does not reveal an unconstrained market-clearing price. Comparing years requires attention to rule changes and the relevant region, not merely a line connecting headline prices.
Trace the pass-through
An illustrative customer with a one-MW capacity obligation at $325 per MW-day would have a simple annualized capacity amount of $118,625 over 365 days before other adjustments.
This arithmetic is not a retail quote. Actual obligations, allocation factors, supplier contracts, and billing treatment determine what appears on a bill.
To assess data-center effects, compare a documented counterfactual with and without the relevant load while holding other assumptions explicit. Removing load can change the clearing price paid across a market, not just the quantity purchased for that customer. Conversely, new supply or different auction rules can change the result substantially.
Report the scope of any modeled attribution. A capacity-market estimate is not an estimate of total household cost, and a result for one delivery year does not establish a permanent burden. The defensible conclusion is narrower: large-load growth can affect capacity procurement and prices, while the amount passed to each customer class depends on market design and retail recovery.
COST CAUSATION: WHO BENEFITS, WHO PAYS?
ALLOCATION Incremental costs and shared benefits both matter. · Dedicated assets and regional assets are different.
Start with the counterfactual
Cost causation asks which expenditures or obligations arise because a customer connects or expands. A dedicated line to a campus may be straightforward to identify. A regional upgrade that also relieves existing congestion requires a broader benefit analysis. The appropriate allocation is not always all or nothing.
FERC’s long-term transmission framework emphasizes evaluating benefits and assigning costs in a manner roughly commensurate with them.[36] Retail allocation and customer-specific service terms operate through additional jurisdiction-specific processes. A federal transmission principle should not be presented as an automatic answer to every state retail-rate question.
Revenue can offset costs
A large customer can help spread existing fixed costs when spare infrastructure is available and its payments exceed incremental service costs.
The opposite can occur when serving the customer requires expensive additions, creates wholesale price pressure, or leaves stranded assets. The direction is an empirical question, not a property of the phrase data center.
Use a transparent bridge from the no-project case to the project case. Include capital costs, operating costs, energy and capacity purchases, incremental revenues, and benefits to other users. Show who bears costs during construction, during the load ramp, and after an early exit.
Do not count general economic-development claims as electricity-rate revenue unless money actually reaches the relevant utility accounts. Local tax receipts, jobs, and procurement may matter to a community, but they do not automatically compensate a residential customer for higher network charges. Keeping these ledgers separate makes both the economic-development case and the ratepayer-protection case more credible.
LARGE-LOAD TARIFFS: PROTECTION BY DESIGN
CONTRACT TERMS Minimum commitments can reduce stranded-cost risk. · The tariff text matters more than the headline.
Make commitments match exposure
Special large-load tariffs can address the risk that infrastructure is built for demand that arrives late, operates below expectations, or disappears. Relevant provisions include minimum billing demand, contract duration, financial security, ramp schedules, termination charges, and responsibility for dedicated facilities. Berkeley Lab’s 2026 rate-design update examines evolving practices for these customers.[47]
The Ohio Consumers’ Counsel reports that the July 2025 [*REDACTED*] settlement requires new large data centers to pay for at least 85% of contracted electricity capacity for up to twelve years even if they use less.[48] This is a specific jurisdictional example, not a national standard and not a requirement that every customer buy 85% of its forecast annual energy.
Test the protection in adverse cases
A minimum bill protects only the obligations it actually covers.
Determine whether it funds distribution assets, transmission charges, generation commitments, or a narrower set of costs. Examine credit support and the ability to collect after bankruptcy, project transfer, or sponsor restructuring. A long contract with weak security may offer less protection than it appears.
The customer also needs predictable terms. Define when the ramp begins, how delays caused by the utility are treated, and whether unused capacity can be released or reassigned. A tariff should discourage speculative reservations without penalizing customers for infrastructure they cannot yet access.
Assess incentives as well as recovery. If a minimum charge is poorly designed, the customer may have little financial reason to reduce peak demand after committing to capacity. Coordinate the tariff with verified flexibility benefits. The goal is to recover justified costs and reward useful behavior, not merely to create a larger guaranteed revenue stream regardless of system value.
STRANDED ASSETS & THE FINANCING GAP
ILLUSTRATIVE ECONOMICS No figures here describe a specific utility. · Security must cover the exposure it is intended to protect.
Costs arrive before full utilization
Consider a hypothetical dedicated network investment with a $20 million annual revenue requirement. If expected campus sales are two billion kWh annually, recovering that amount entirely through energy sales would require one cent per kWh. If sales reach only one billion kWh and the same costs remain, the required amount becomes two cents per kWh. This simplified example excludes other charges and illustrates volume risk, not a proposed tariff.
The risk is greatest when the utility commits capital early but the customer ramps slowly. A forecast can ultimately prove correct while still leaving a substantial interim recovery gap. Financing and depreciation schedules do not necessarily wait for server utilization to catch up.
Allocate delay and exit risk explicitly
A service agreement can reduce this mismatch through milestones, upfront contributions, minimum charges, or secured termination obligations.
Each mechanism transfers a different risk. An upfront payment reduces financing exposure; a minimum charge addresses utilization; collateral improves collectability; a termination payment addresses early departure.
Test the contract against a delayed opening, a partial buildout, and a complete cancellation. Calculate the remaining unrecovered obligation after applying security. Then identify who pays any residual amount. Do not assume a recognizable corporate name guarantees every project entity’s obligations.
Also consider alternative use. An asset that can serve several future customers has a different stranded-cost profile from a highly specialized facility. Crediting realistic reuse value can avoid overcharging the initial customer, but speculative reuse should not be treated as certain. The financial test should be conservative, explicit, and periodically refreshed as the project and surrounding system develop. Ratepayer protection depends on enforceable cash flows, not simply an optimistic demand curve.
AFFORDABILITY: DEMAND A CAUSAL RECORD
PUBLIC ACCOUNTABILITY Separate evidence from attribution. · Show the no-project case and the project case.
A useful public explanation
Communities deserve more than a claim that growth will pay for itself or that every increase is caused by AI. Publish a service-area analysis showing new investment, expected customer payments, exposure to wholesale energy and capacity costs, and the treatment of delays. Summarize uncertainty in understandable terms without disclosing confidential commercial details.
EIA makes electricity sales and price data available by state and utility.[49] These datasets can establish what happened to average prices and sales. They do not, by themselves, identify the causal contribution of a particular development. That requires utility filings, allocation decisions, and an explicit counterfactual.
Measure who is exposed
An aggregate finding of net benefit can conceal uneven outcomes. Residential customers, small businesses, industrial users, and customers on competitive supply contracts may experience different changes.
A utility’s revenue position and a household’s affordability problem are not the same measure.
Show bill effects at several usage levels and explain fixed-charge changes separately from energy charges. Identify whether protections apply to existing projects, new projects, expansions, or only a defined customer class. Track the timing of benefits and burdens; a future tax base does not necessarily help a household facing a near-term electric-bill increase.
Independent review should examine the assumptions most capable of reversing the result: customer ramp, infrastructure cost, wholesale prices, contract security, and allocation methodology. Where evidence is insufficient, present the uncertainty rather than a precise unsupported percentage. A credible public process can acknowledge substantial digital-infrastructure benefits while insisting that the people financing shared systems understand their exposure and have a meaningful opportunity to challenge it.
A DECISION STANDARD FOR RESPONSIBLE GROWTH
OPERATING STANDARD Approve evidence-backed service, not speculative megawatts. · Review the forecast, contract, and physical system together.
Five conditions for a credible connection
A responsible large-load decision establishes a credible demand ramp, deliverable electrical service, enforceable financial responsibility, acceptable operating behavior, and transparent treatment of other customers. Failure in any one can undermine the project even if the others appear strong. A well-funded campus cannot operate without power; a technically feasible connection can still create an unfair financial exposure.
Recent federal proceedings emphasize clarity, consumer safeguards, and region-specific large-load integration.[50] Their existence reinforces the need to distinguish proposed reforms from effective requirements. Before making a binding decision, verify the applicable tariff, order, service agreement, and technical study rather than relying on a general policy announcement.
Maintain a living evidence record
Record the forecast vintage, project milestones, committed capacity, measured utilization, and infrastructure dependencies. Reconcile operating data with the forecast at regular intervals.
Escalate material changes in size, timing, financial backing, or equipment behavior before they invalidate planning assumptions.
Use staged approvals where technically and legally appropriate. Tie each expansion to demonstrated readiness and the availability of required supply and network capability. If flexible or conditional service is part of the arrangement, test it and document the consequences of nonperformance. A promise to cooperate is not a substitute for dispatchable, measurable behavior.
The central conclusion is conditional rather than ideological. Data centers can support investment, digital capability, and utility revenue, while also creating concentrated demands and material cost risks. Neither benefit nor harm is automatic. The strongest projects make their electrical requirements visible, pay for justified obligations, and remain workable under plausible delays and stress conditions. The strongest public oversight measures outcomes and updates decisions when the evidence changes.
Exhibit 07 · Global data-center electricity demand
The IEA’s central 2030 projection includes all data centers, not AI only. The 2025 figure is an estimate, not a measured sum of every facility. The hatched bar is a projection.
IEA, Key Questions on Energy and AI: primary source
Exhibit 08 · U.S. data-center electricity demand
Berkeley Lab’s 2025 update estimates 192 TWh in 2024 (4.7% of U.S. electricity) and projects 649 TWh in its 2030 reference scenario (11.8%). The projection is not a committed interconnection queue or an AI-only total.
LBNL, 2025 Update; laboratory seminar, Nov. 4, 2026: primary source
Exhibit 09 · Residential electricity prices vary
Selected states and national average. Average revenue per kWh is not a household bill, approved tariff schedule, or estimate of the effect of data centers. State selection is illustrative.
EIA, Electric Power Annual, 2025 state retail price table: primary source
Exhibit 10 · PJM capacity price is not a retail rate
Illustrative comparison of the cited Base Residual Auction RTO prices. Capacity procurement is separate from energy, retail tariffs and final utility bills; this chart does not identify how much any one data center caused.
PJM, 2028/2029 Base Residual Auction results, July 14, 2026: primary source
Browse: White Papers · Reports · Downloads
