A collaboration between Lewis McLain & AI

It’s Not Only the Gallons — It’s the Peak: Peak Demand, Utility Capacity, and the Terms That Protect a Community

A plain-English technical paper · July 2026

Executive Summary

Headlines about data centers quote annual totals — millions of gallons, megawatts “equal to a small city.” Those totals are real, but they point at the wrong risk. Utility systems are built, and financed with decades of debt, to survive their single worst day. Capacity built for that day must be paid for all year, every year, whether or not the water or power is ever sold. So the first question about any data center is not “how much will it use?” It is “how big must we build — and who pays for what stands idle?”

Three findings follow, developed one step at a time in the body of the paper:

• On electricity, data centers are among the steadiest, cheapest-to-serve customers a utility can sign. The real electric-side issues are sheer size, capacity reserved but never used, and split-second power swings from AI computing — all manageable by contract.
• On water, the opposite. The newest cooling designs cut annual gallons dramatically but concentrate demand into a few hot weeks — so a facility that looks small on annual volume can force a community to build big-city capacity that sits idle most of the year, and its wastewater brings its own peak and its own chemistry.
• Block one path and the demand moves to another network. Denied water, a data center leans harder on the electric peak; denied grid power, it builds gas-fired generation that lands on the pipeline system’s worst winter morning — and needs water of its own. Water, electricity, and gas must be evaluated as one decision.

Two more conclusions complete the picture. First, where the underlying supply — the water in the watershed, the generation on the grid, the gas in the pipe — is genuinely short, totals do matter, and the honest remedy is curtailment shared evenhandedly by all users, not a ban on the newest arrival. Second, the benefits are just as real as the burdens: a data center can be the best-shaped utility customer a system has ever been offered and a major tax base that demands almost nothing in services. Every major objection has a practical fix that can be written into a contract — including, for the water peak, a storage tank costing the developer roughly one percent of the project. The right answer to a well-sited data center proposal is neither “no” nor an unexamined “yes.” It is “yes, if” — and this paper ends with the term sheet.

1. Introduction

Public debate about data centers has focused overwhelmingly on totals: millions of gallons of water per year, or electricity use “equal to a small city.” Totals make good headlines, but they describe the wrong risk. A utility is rarely crippled by how much its customers use over a year. It is crippled — or forced into expensive expansion — by how much all of its customers demand at the same moment, on the system’s single worst day.

The question is no longer academic, and it is not on the horizon — it is here, and policymakers at every level are moving. At least eleven states have weighed moratoriums; New York’s legislature has passed the nation’s first statewide pause, awaiting the governor’s signature; Maine’s governor vetoed a similar bill; and in Texas, the Public Utility Commission and the Texas Water Development Board are surveying existing and planned data centers about water use and treatment, with a PUC report due to the Legislative Budget Board and the Governor at the end of 2026 [15][22]. The scale in Texas alone frames the stakes:

Texas snapshot	Figure
Data centers operating (2025)	400+, concentrated in five regions: Dallas–Fort Worth (197), San Antonio (60), West Texas (59), Austin (53), Houston (48)
Water use by Texas data centers, 2025	≈25 billion gallons
Projected water use by 2030	29–161 billion gallons per year — up to ≈2.7% of all Texas water use
Electricity demand today	≈9,500 MW
Projected electricity demand, 2030	≈22,000–78,000 MW (ERCOT forecast range)
The planning gap	ERCOT has a process for interconnecting large electric loads; no parallel statewide process yet exists for water

The scale of the question — Texas snapshot [22][23].

This paper explains why load factor, not annual usage, is the master variable in utility economics — and where that rule stops, because in supply-constrained regions totals do matter; examines what published data show about data center demand for electricity, water, and natural gas; traces the consequences for community capital budgets, utility rates, and municipal credit; and then turns the argument around, examining the substantial benefits data centers offer and the engineering and contractual tools that can neutralize each major objection. The subject is genuinely technical, and this paper does not pretend otherwise; the approach is to take it one step at a time, define every term on first use, and end each section with a Key Takeaway box for the reader who wants the point without the plumbing.

KEY TAKEAWAY — The risk to a community is not how much a data center uses in a year; it is how big a system must be built — and financed — to stand ready for its single worst day. That question has now reached legislatures and governors, and Texas agencies owe the Governor a report on it at the end of 2026.

2. Utility Economics 101: Systems Are Built for the Peak, Not the Average

2.1 Two kinds of cost

Every utility has two fundamentally different kinds of cost. Operating costs — fuel, treatment chemicals, pumping energy — rise and fall with how much product is actually delivered. Capital costs are different: they are the cost of building the system itself — the treatment plants, storage tanks, pipes, generating stations, wires, and substations. Capital costs are driven almost entirely by capacity, and capacity is set by expected peak demand plus a safety margin, not by average or total demand [12].

The intuition is everyday. A church kitchen is sized for Easter Sunday, not the average Tuesday. A highway is sized for rush hour, not 3 a.m. A water system must be able to treat and deliver the maximum-day demand, sustain the peak hour within that day, and still hold reserve for firefighting during a heat wave [1]. An electric grid must meet the single highest hour of the year, typically a summer afternoon when air conditioners are running everywhere at once. All of that capacity must be built, financed, and maintained year-round, even though much of it is idle most of the time. The debt service on a treatment plant is due every month; the plant only earns its keep on the days it runs hard.

2.2 One concept, two vocabularies — and the convention used in this paper

The electric and water industries measure the same idea from opposite directions, which causes endless confusion. The electric industry speaks of load factor — average demand divided by peak demand, so higher means flatter. The water industry speaks of peaking factor — maximum-day demand divided by average-day demand, so lower means flatter. One is simply the reciprocal of the other: a peaking factor of 4.0 is a load factor of 25 percent. They are two ways of looking at one critical concept: what share of the capacity reserved for a customer is actually used. To be consistent and as simple as possible, this paper uses load factor throughout, converting water-industry peaking factors wherever sources report them (load factor = 1 ÷ peaking factor). Table 1 sorts out the terms.

Term	Definition	A perfectly flat customer
Load factor	Average demand ÷ peak demand, usually over a year. Higher is flatter and cheaper to serve. Used throughout this paper for water, electricity, and gas alike.	100%
Peaking factor	Maximum-day demand ÷ average-day demand — the water industry’s customary metric. The reciprocal of load factor: a peaking factor of 4.0 equals a 25% load factor. Converted to load factor throughout this paper.	1.0
Peak-hour factor	Peak-hour flow ÷ average flow; derived from daily figures with additional multipliers; sizes pumps and storage.	1.0
Coincident peak	A customer’s demand at the moment of the whole system’s peak — the demand that actually drives system-wide capacity.	Equal to its average

Table 1. Key terms. Load factor and peaking factor are reciprocal expressions of the same ratio; this paper standardizes on load factor [1][2][12].

2.3 A worked example: same gallons, six times the cost

Consider two industrial customers who each buy exactly 36.5 million gallons per year — identical annual volume, identical volumetric revenue to the utility. Customer A draws a steady 100,000 gallons every day: a 100 percent load factor. Customer B averages the same 100,000 gallons per day but spikes to 600,000 gallons on the hottest days of summer: a load factor of just 17 percent.

	Customer A (flat)	Customer B (peaky)
Annual purchases	36.5 million gallons	36.5 million gallons
Average-day demand	100,000 gal/day	100,000 gal/day
Maximum-day demand	100,000 gal/day	600,000 gal/day
Load factor	100%	17%
Capacity the utility must build	100,000 gal/day	600,000 gal/day
Capacity cost per gallon actually sold	1x	6x

Table 2. Two customers with identical annual usage. Customer B forces the utility to build six times the capacity for the same sales.

If the utility recovers costs mainly through volumetric rates — dollars per thousand gallons — both customers pay the same, yet Customer B caused six times the capital requirement. The difference does not vanish; it is silently financed by every other ratepayer, embedded in general rates or in system-wide debt. This is the peaking problem in one sentence: capacity is what a community pays for, while usage is merely what the utility happens to sell. Worse, the capacity built for Customer B’s few hot days stands idle roughly 360 days a year while bondholders are paid on all 365.

2.4 A necessary distinction: supply versus capacity

Before the load-factor argument can be applied honestly, two different scarcities must be kept separate, because they are routinely conflated in public debate. The first is delivery capacity — the size of the treatment plant, the pipe, the wire, the pipeline. Capacity is a manufactured thing: a community can build more of it, and what drives its cost is peak demand. That is the subject of this paper. The second scarcity is the commodity itself — the water actually in the watershed or aquifer, the gas actually deliverable from the basin, the energy actually generatable over a year. Supply is not manufactured on demand, and what strains it is total consumption, not peaks.

Where supply is genuinely short — a drought-stressed river basin in the Southwest, a pipeline-constrained region in a deep freeze, a grid short of generation — total usage is a legitimate, binding concern, and no rate design can conjure water that does not exist. But the remedy in a supply crisis is fundamentally different: curtailment and allocation applied evenhandedly to all users — industrial, commercial, and residential alike — because a shortage of supply is a community problem, not a data center problem. A large flat customer did not cause the drought, and a moratorium on one industry does not refill the reservoir; it merely reassigns who gets to consume a fixed shortfall.

The practical sequence for any community evaluating a large-load proposal therefore has two steps. Step one is a supply adequacy assessment: how much firm water, generation, and gas supply actually exists under stress conditions — drought of record, heat wave, polar vortex — and how much is already committed? If that test fails, the conversation is about curtailment rules for everyone, and it should be had openly rather than through the back door of blocking whichever large customer arrived last. Step two, where supply is adequate, is the subject of the rest of this paper: the binding economics become peak-driven capacity, and the task is to make the customer who causes the peak pay for the peak.

KEY TAKEAWAY — Capacity is what a community pays for; usage is what it happens to sell. A customer’s load factor — how flat its demand is — decides its true cost to the system. And before any of that: verify the supply exists, because a supply shortage is everyone’s problem, not one industry’s.

3. Electricity: The Surprisingly Flat Customer

3.1 What the data show

The popular image of a data center is a building running flat-out, 24/7/365. On the electric side, that image is close enough to true that it works in the data center’s favor. Utilities and analysts commonly assume load factors of 90 to 100 percent for large data centers [2]. Utility planning practice is somewhat more conservative: Duke Energy plans for new large loads at about an 80 percent load factor, and Dominion reported roughly an 82 percent actual load factor for large data centers in Virginia in 2024 [3]. Metered results vary more than the folklore suggests — one analysis of actual facilities found fewer than half exceeded an 80 percent load factor [2][3] — but even so, data centers remain far flatter than almost any other customer class. Published figures translate to electric load factors of roughly 50 to 65 percent for conventional cloud data centers and close to 100 percent for AI training facilities, whose accelerators grind at near-constant power for weeks at a time [4].

In classic rate-making terms, a high-load-factor customer is the cheapest kind to serve per kilowatt-hour: the demand-related capital cost is spread across an enormous number of kilowatt-hours. On the traditional load-factor test — the same test that condemns them on water, as Section 4 shows — data centers pass with room to spare.

3.2 The three real electric-side problems

First, scale and coincidence. A flat load contributes its entire demand to the system’s coincident peak. A 500 MW data center at a 90 percent load factor still adds up to 500 MW to the summer peak hour, and the grid must build for it. Flatness spreads the cost over more energy; it does not shrink the capacity requirement.

Second, contracted versus realized demand. Load factor is measured against a facility’s realized peak, but the utility must build to its contracted (nameplate) capacity. If a facility’s realized peak is only 80 percent of its contract and its load factor is 90 percent, its true utilization of the capacity the utility stands ready to serve is only 72 percent [2]. Facilities also routinely over-request capacity to preserve room for growth, and they ramp up over years. The gap between what is reserved and what is used is a hidden peaking problem, and it is why minimum-demand charges, ratchets, and take-or-pay provisions exist.

Third, fast transients. AI training clusters synchronize tens of thousands of processors, producing quasi-periodic power swings every 2 to 6 seconds that can oscillate by tens of megawatts within a single facility, with ramps measured in tens to hundreds of megawatts per second at job starts, checkpoints, and failures [4][5][6][7]. These sub-minute swings do not show up in a monthly demand charge at all — they are a power-quality and grid-stability problem, a new kind of “peaking” measured in seconds rather than seasons.

3.3 Proof that peaks, not energy, are the scarce resource

Researchers at Duke University quantified how much slack lives inside existing grids if new loads simply avoid the worst hours: about 76 gigawatts of new load — roughly 10 percent of the entire U.S. aggregate peak — could be added to existing systems if that load curtailed just 0.25 percent of its annual energy, about 85 hours per year in events averaging two hours [8]. Nearly all of the grid’s spare room is hiding in the hours that are not the peak. The same insight, mirrored, explains the water problem in the next section.

KEY TAKEAWAY — On electricity, data centers are among the flattest, cheapest-to-serve customers ever offered to a utility. The real electric-side issues are sheer size at the system peak, capacity reserved but never used, and split-second power swings — each manageable by contract.

4. Water: Where the Peaking Problem Lives

4.1 Why constant cooling does not mean constant water

A natural objection runs: a data center needs cooling 24 hours a day, all year — so surely its water demand is flat, like its electricity. The premise is right and the conclusion is wrong, because water is not how most of the cooling is done. It is the peaking fuel.

Physics sets the terms. Essentially every watt of electricity entering the building becomes heat, and that heat must leave the building continuously. There are only two ways to dump heat into the environment: blow outside air across coils (dry cooling — no water consumed), or evaporate water, which absorbs roughly 8,000 Btu per gallon — the same trick the human body uses when it sweats [1][12]. Dry cooling works fine when outside air is cool. It falters exactly when it is 105°F outside, because you are trying to push heat into air nearly as hot as the equipment. Evaporation keeps working in extreme heat because it exploits the humidity-adjusted (“wet-bulb”) temperature, which stays well below the air temperature. So water demand tracks the weather, not the computing load.

4.2 But is the water not recirculated?

Inside the building, yes. Modern facilities — especially liquid-cooled AI facilities — run closed loops in which the same water is chilled and recirculated indefinitely; that water is bought once [13]. But chilling the loop does not destroy the heat, it only moves it to the edge of the building, where it must still be rejected to the outdoors by one of the same two paths: fans (electricity) or evaporation (water). Even a cooling tower recirculates its water many times — but on each pass a fraction evaporates, and a further portion must be deliberately drained and replaced (“blowdown”) to keep dissolved minerals from scaling the equipment. Recirculation dramatically reduces water withdrawal; it cannot eliminate consumption where evaporation does the work [1][12][13].

4.3 The paradox: designs that use less water use it more peakily

The dominant new design is dry cooling with evaporative assist: fans handle the heat load whenever ambient temperatures allow — typically 85 to 95 percent of the hours of the year — and water sprays switch on only during the hottest hours, when dry cooling alone cannot keep up [1]. This slashes annual water use, which is what operators advertise. But it concentrates the entire year’s water demand into a few hundred summer hours. Annual gallons went down; the maximum-day demand the water utility must be built to serve did not. The water load factor collapses — from roughly 45 percent for a conventional cooling tower into the teens or below.

Facility / system	Cooling type	Water load factor
Typical residential and commercial users (benchmark)	—	40 – 67%
The Dalles, OR pressure zone dominated by a hyperscale data center (measured)	Evaporative cooling towers	45%
Phoenix, AZ colocation facility (estimated)	Evaporative cooling towers	≤ 45%
West Des Moines, IA hyperscaler (measured monthly 23%; estimated daily)	Dry with evaporative assist	≤ 15%
Planned AI data center, Indiana innovation district (planning)	Dry with evaporative assist	≤ 16%
Leesburg, VA hyperscale campus (planning, inferred)	Dry with evaporative assist	≈ 12% or lower
Data centers in Prince William Water service area, VA (measured, 2024)	Mixed fleet	10%
Northern Virginia weighted actual, many facilities combined	Mixed fleet	27 – 29%
Wisconsin AI campus: 0.7 MGD capacity requested vs. ≈23,000 gal/day average	Dry with evaporative assist (≈480 hrs/yr of water use)	< 3.5%
Leading operators’ planning values for state-of-the-art facilities	Dry with evaporative assist	12 – 17% or lower

Table 3. Documented data center water load factors (average-day demand ÷ maximum-day demand). Source documents report daily peaking factors; converted here as load factor = 1 ÷ peaking factor. All figures compiled in [1] from utility records, government filings, and planning documents.

Two rows deserve emphasis. The West Des Moines facility uses relatively little water in winter, yet its summer draw is so large that it became the utility’s single largest annual water customer in 2024 and 2025 — a low load factor and large scale compounding [1][9]. And the Wisconsin campus is the design taken to its logical extreme: in a cold climate, evaporative assist is expected to run only about 480 hours per year, so the facility averages a trivial 23,000 gallons per day while requiring the utility to stand ready to deliver 700,000 gallons per day — a water load factor of barely 3 percent. By annual volume it is a small customer. By the capacity the community must reserve for it, it is enormous. That single contrast is the central point of this paper.

4.4 The worst timing possible

These peaks are not randomly timed. Evaporative assist switches on during prolonged heat waves — precisely when lawns are being irrigated, municipal demand is at its annual maximum, drought restrictions may be in force, and fire risk is elevated. The data center’s peak lands on top of everyone else’s peak, so it adds almost gallon-for-gallon to the system’s design day. And unlike electricity, where batteries can shave short peaks, operators have rarely built multi-day water storage voluntarily [1] — under volumetric rates, there is no reason to; the utility bears the peak. Section 9 shows how sharply that economics changes once the peak is priced.

4.5 The return trip: wastewater has its own peak and its own chemistry

What a data center sends back down the sewer deserves the same scrutiny as what it takes in. Cooling blowdown is concentrated by design: the same gallons that cycled through the towers return with elevated dissolved minerals and salts, along with residues of the biocides and corrosion inhibitors used to protect the equipment — compounds containing phosphates, nitrites, and heavy metals that stress municipal treatment plants, raise treatment costs, and can threaten permit compliance downstream [22]. Regional wholesale providers are explicit on this point: the North Texas Municipal Water District advises member cities to engage its planning and wastewater pretreatment teams early — before site plans are approved, not after [22].

And the return flow is peaky too, for the same reason the intake is: blowdown scales with evaporative operation, which is concentrated in the hottest weeks. At the Wisconsin campus discussed above, the planned wastewater discharge pattern implies a load factor of roughly 7 percent [1] — the sewer plant, like the water plant, must be sized for a worst day that arrives only a few times a year. The remedies are standard industrial-pretreatment tools, applied before approval: discharge permits with concentration and flow limits, pretreatment requirements, surcharges tied to strength and peak flow, and early coordination with the wholesale treatment provider [22].

KEY TAKEAWAY — Cooling runs 24/7; water demand does not. Modern designs cut annual gallons but concentrate demand into heat waves — water load factors of 3–17% — and the wastewater comes back with its own peak and its own chemistry. Engage the treatment provider before the site plan is approved.

5. What Peaks Cost a Community

Water systems are sized by three peak-driven standards: treatment and supply for the maximum day, pumping and storage for the peak hour, and reserve capacity for fire flow — all with safety margins for heat waves and drought [1]. Every one of those design parameters is set by peak demand. None is set by annual volume.

The capital sums are not abstract. In one Indiana economic-development district, adding 25 million gallons per day (MGD) of supply and 15 MGD of wastewater treatment is expected to take about six years and cost over $1 billion — and a single technology company’s data center holds 8 MGD, roughly a third, of the new allocation [1]. In Wisconsin, a data center developer’s request for 1.2 MGD triggered over $100 million in upgrades in a town whose entire remaining system capacity was under 2 MGD [1]. Three recent water infrastructure upgrades for large technology companies together approached $1 billion; in Louisiana, one company committed up to $400 million toward public water infrastructure serving its new campuses [1]. Nationally, U.S. data centers are projected to need roughly 700 to 1,450 MGD of new water capacity by 2030, with an estimated infrastructure valuation reaching as high as $58 billion — landing on public systems that already face a $1.3 trillion-plus, twenty-year funding backlog, in a country where 12 to 19 million households already lack affordable water service [1][11].

Rate design determines who bears these costs. Volumetric rates and gallonage-based impact fees systematically under-recover from low-load-factor customers, because — as Table 2 showed — such customers cause capacity costs far out of proportion to the water they buy. The under-recovery is not returned to anyone; it is spread across the remaining ratepayers as higher general rates or added system debt. The remedies are the water-sector equivalents of tools the electric industry has used for a century: capacity or demand charges based on each customer’s allocated maximum-day capacity; ratchet provisions so that a customer who sets a peak keeps paying for it; take-or-pay contracts and developer-funded infrastructure agreements; and impact fees computed on allocated max-day demand rather than on meters or projected average gallons [1][2]. A prerequisite for all of these is disclosure: most operators publish only annual water totals, often aggregated across an entire corporate fleet, which conceals exactly the numbers — facility-level peak demand and load factor — that determine what the community must build [1].

KEY TAKEAWAY — Water systems are sized by the maximum day, not annual volume. Under purely volumetric rates and gallonage-based fees, every other ratepayer quietly finances a peaky customer’s capacity — and the number that would reveal it, facility-level peak demand, is the one operators rarely publish.

6. The Peak Does Not Disappear — It Migrates Among Three Networks

6.1 Water to electricity

There is a tempting policy response to the water problem: simply require waterless, all-dry cooling. Some operators in cool climates already do this voluntarily. But conservation of energy is unforgiving — the heat still has to leave the building, and if water may not carry it, electricity must. All-dry facilities need larger chillers and more fan power, and their cooling equipment works hardest, least efficiently, on the hottest afternoons — which is precisely the electric grid’s annual peak hour. Squeezing the water peak inflates the electric peak. Analysts now project that water availability will rival power availability as the binding constraint on data center siting within about five years, and warn that “unmet” water demand is too often counted as zero demand rather than as a constraint someone must eventually pay to relieve [1][3].

6.2 Electricity to natural gas

The migration does not stop at the electric meter. When a utility’s interconnection queue is measured in years, developers increasingly build their own power: on-site, “behind-the-meter” gas-fired generation. Industry analysis counts roughly 100 gigawatts of gas-burning capacity planned to power U.S. data centers [16]. At first glance this looks like the community’s problem solved — the data center pays for its own generation and spares electric ratepayers the capacity build-out. In reality it moves the peak to a third network that is sized, financed, and priced by exactly the same design-day logic as the other two.

Gas pipelines and distribution systems are built for their own worst day: the coldest winter morning, when every furnace and boiler runs at once. Capacity on that system is sold in two forms. Firm transportation is reserved and paid for every day of the year whether a single cubic foot flows — the purest expression of this paper’s unused-capacity economics, and the revenue stream that finances pipeline construction. Interruptible transportation is cheaper precisely because it is cut first when the system tightens [17]. A gas-fired data center plant is, like the facility it powers, admirably flat — a high-load-factor gas customer. But its winter demand lands squarely on the design day, stacked on top of heating customers, tightening capacity exactly when both sectors need it most; in pipeline-constrained regions, polar-vortex events already push gas systems to their limits [17][18]. If the data center buys firm capacity, it enlarges the design-day build-out someone must finance. If it economizes with interruptible service, its celebrated reliability quietly depends on diesel backup or on being curtailed — and if it under-contracts, the shortfall risk shifts to everyone else sharing the pipe. Add the local-acceptance problem — gas turbines bring emissions and noise that communities did not bargain for [15] — and the “bring your own power” model is revealed as the same arithmetic on a different network. And note what the water district reminds its member cities: on-site generation itself requires water — thermal plants need cooling too — so bringing your own power can also mean bringing additional water demand with it [22].

The lesson for communities is to evaluate a data center proposal as one interlocking peak-capacity problem across all three networks — water, electricity, and gas — because the developer’s cooling and power choices largely decide which system’s design day grows and, therefore, which set of ratepayers funds the standby capacity.

KEY TAKEAWAY — Deny water and the peak moves to the electric grid; deny grid power and it moves to the gas system’s worst winter morning — and the generators need water of their own. Evaluate all three networks as one decision.

7. Moratoriums, Municipal Credit, and the Pricing Alternative

The politics have begun to move faster than the rate design. At least eleven states have considered moratoriums on data center development; New York’s legislature has passed a Responsible Data Center Development Act imposing a one-year statewide pause — the nation’s first, if signed — while Maine’s legislature passed a similar measure that was vetoed by its governor, and individual municipalities have enacted their own local pauses [15]. Governors, legislatures, and regulators are now directly engaged: in Texas, the Public Utility Commission and the Texas Water Development Board are surveying existing and planned data centers on water use and treatment, with a PUC report due to the Legislative Budget Board and the Governor at the end of 2026 [22]. Communities that put their own terms in place now will negotiate ahead of the coming rulebook; those that wait will inherit whatever it says. Rating agencies, meanwhile, have supplied the underlying worry in careful language: substantially increased water and electricity demands “can strain existing infrastructure and require the build-out of new capacity,” and if those costs are not offset by new revenues, the community funds the difference to the detriment of its credit [15]. That sentence is this paper’s argument translated into credit analysis: the build-out is financed with decades of debt, while the offsetting revenue depends on rate structures that mostly track usage.

The credit risk compounds because peaky demand is also concentrated and cancelable demand. Industry tracking counts some 120 data center projects canceled since 2024, most of them recently and most due to local opposition [15]. A utility that issues thirty-year revenue bonds for maximum-day capacity sized to a single customer’s planned peak bears stranded-capacity risk three ways: the project may cancel before it draws a gallon, it may ramp to only a fraction of its reserved capacity (the contracted-versus-realized gap of Section 3.2), or the facility may be functionally obsolete in ten to fifteen years while the debt runs twice that long. A low load factor, high concentration, and cancelability is the worst available combination for a revenue bond. Seen this way, take-or-pay contracts, developer-funded infrastructure, and capacity charges that survive cancellation are not merely fairness measures — they are bondholder protection.

One bookkeeping discipline keeps the debate honest: fiscal losses are not utility losses. Texas is reported to forgo more than $1 billion per year in sales tax revenue through data center exemptions, and Virginia’s incentives have been estimated at similar annual magnitudes [15]. Those are tax expenditures — policy choices about revenue — and they belong in a different column from the capacity cost-shifting analyzed here. Conflating the two weakens both arguments and lets each be rebutted with answers to the other.

Industry’s standard defense, meanwhile, concedes the central point. The Data Center Coalition states its members expect to be billed their “full cost of service,” with no other ratepayer paying for costs “directly assigned” to them [15]. But costs are only directly assigned if the tariff assigns them. Under volumetric water rates and gallonage-based impact fees, a customer with a water load factor of 12 to 17 percent is — by construction, as Table 2 showed — subsidized by everyone else, however sincere its intentions. A moratorium, in the end, is a price-signal failure: it is what a community reaches for when its prices cannot yet say what its engineers already know. The constructive exit is the sequence this paper has argued throughout — an honest supply adequacy assessment first, then capacity-based pricing, ratchets, take-or-pay terms, and debt backstops. Operators who are committed, as the industry says, to being responsible neighbors can demonstrate it by signing.

KEY TAKEAWAY — A moratorium is a price-signal failure — what a community reaches for when its prices cannot yet say what its engineers already know. The durable fix: verify supply, then price capacity and write contracts that survive cancellation. Move before the state rulebook arrives.

8. The Other Side of the Ledger: Why a Community Might Want to Say Yes

8.1 The best possible shape of electric customer — with one condition

Everything in Section 3 deserves restating as a benefit, because it is one. An 80-to-100 percent load factor customer is the best possible shape of electric load: it buys enormous volumes of energy against the capacity reserved for it, spreading the utility’s fixed costs — poles, wires, substations, administration — across far more kilowatt-hours than any residential or commercial customer ever could. On a system with spare capacity, adding such a customer puts downward pressure on everyone else’s average rates. This is the classic load-factor-improvement argument that utilities themselves made for decades when courting industry, and data centers fit it better than almost any industrial load in history. Communities that reflexively treat a data center as a burden on the electric system have the shape of the problem backwards.

The condition is headroom. The fixed-cost-spreading logic works when the new load is absorbed by existing, underused capacity. When the load instead triggers new generation and transmission at today’s marginal costs — which now exceed embedded average costs in most regions — even a perfectly flat customer can raise average rates. The evidence is current: capacity prices in PJM, the nation’s largest grid market, jumped roughly nine-fold between the 2024/25 and 2025/26 auctions, the market monitor attributed 63 percent of that increase — about $9.3 billion in a single year — to data center demand growth, and residential bills in parts of the region rose $16 to $21 per month [20]. Neither story disproves the other: the same customer shape that lowers rates on a slack system raises them on a tight one. This is why the supply-and-headroom assessment of Section 2.4 must come first, and why the Duke flexibility finding matters so much — 76 gigawatts of room exists for loads willing to curtail a few hours a year [8]. The honest framing for a community is: a data center is the best-shaped electric customer available; whether it is also a rate-lowering one depends on the system it joins and on whether it pays the marginal cost of what must be built for it.

8.2 The fiscal engine: high taxable value, almost no service demand

The second benefit belongs to the tax assessor, not the utility. Decades of cost-of-community-services studies show that commercial and industrial property typically demands only 35 to 65 cents of local services for every dollar of revenue it generates, while residential development demands $1.15 to $1.50 — more than it pays [19]. A data center is an extreme version of the industrial case: enormous assessed value in buildings and equipment, combined with a few dozen to a hundred employees, no schoolchildren, negligible traffic, and minimal police and fire demand. The taxable value includes not just the real property but the business personal property — the servers themselves, often worth as much as the building — which is refreshed on a three-to-five-year cycle that keeps replenishing the tax base as older equipment depreciates.

Here the record needs one correction that strengthens the community’s hand. The widely cited Texas figure — more than $1 billion per year in forgone revenue — is a sales tax exemption on equipment purchases, a state-level policy choice [15]. Property tax is different: data centers generally owe it in full unless a local government chooses to abate it. That choice is the community’s single best bargaining chip, and it should be spent, not given away: an abatement worth tens of millions of dollars can be traded for precisely the protections this paper prescribes — capacity-based utility charges, take-or-pay terms, on-site peak mitigation, disclosure, and host-community payments.

But the fiscal case is not automatic, and honesty requires saying so: your mileage may vary — so measure it. Tax rates, abatement posture, school-district shares and recapture, utility ownership, debt structure, and siting jurisdiction all move the answer, sometimes decisively. In states like Texas, the school district share of property tax dominates the bill and recapture provisions export much of it beyond the community, so the locally retained benefit is smaller than the headline levy. A facility sited just outside city limits can burden a city-owned utility while paying the city nothing. And the jobs argument should be retired: construction employment is transitory and permanent staffing is small [15]; the durable benefits are fiscal and, under the right conditions, ratepayer relief — argue those instead. Because the arithmetic is so community-specific, the authors have built a companion interactive tool — the CityBase.Net Data Center Fiscal Impact Model — that walks a Texas city through the full 20-year calculation from the municipal government’s perspective: property and sales tax revenues, utility revenues and costs, municipal service costs, employment, capacity audit, and net fiscal impact under alternative scenarios [24]. The fundamental question the model asks is the right one: not “how big is the investment?” but “does the city collect more than it spends to serve the facility — and by how much?”

KEY TAKEAWAY — A data center can be the best-shaped utility customer a system has ever signed and a major tax base demanding almost nothing in services — where headroom exists and abatements are traded for protections, never given away. But every community’s numbers differ: run them. A companion fiscal model [24] exists for exactly that purpose.

9. Engineering Away the Water Peak: The On-Site Storage Option

The biggest technical objection in this paper — the collapsed water load factor — turns out to have a remarkably cheap fix, and understanding why it has not already happened everywhere is itself instructive. The fix is a tank: on-site storage that the data center fills slowly from the public system at a contractually capped rate, around the clock and across the seasons, and draws down during heat waves. The tank does for the water system exactly what a battery does for the grid — it converts a peaky customer into a flat one. Nor is it an alien imposition: high-reliability facilities already keep hours of cooling water on-site as standard practice for ride-through of utility interruptions [12]; the proposal here simply extends hours to days.

The arithmetic is worth showing. Consider an illustrative 100 MW facility using dry cooling with evaporative assist. At a planned peak water intensity of 1.15 liters per kilowatt-hour — the Leesburg planning figure [1] — its peak-day demand is about 0.73 MGD, while at a 12 percent water load factor its average day is roughly 0.09 MGD. Table 4 shows what a tank does to the capacity the public system must reserve.

Contract design	Max draw from utility	Tank size (7-day design heat wave)	Illustrative tank cost
No cap (status quo)	0.73 MGD	None	$0 — but the utility builds 0.73 MGD of capacity
Partial cap	0.25 MGD	≈ 3.4 million gallons	≈ $7–14 million
Near-flat (≈100% water load factor)	0.10 MGD	≈ 4.4 MG (≈6.3 MG for a 10-day event)	≈ $9–18 million

Table 4. Illustrative on-site storage sizing for a 100 MW data center with dry cooling and evaporative assist (peak day ≈0.73 MGD at 1.15 L/kWh [1]; average day ≈0.09 MGD at a 12% water load factor). Tank costs assume roughly $2–4 per gallon installed for multi-million-gallon ground storage [21]. All figures are order-of-magnitude planning estimates.

Three things stand out. First, full flattening costs barely more than partial mitigation — once the tank is sized for a design heat wave, capping the utility draw near the average adds little. Second, the money is small where it matters: $9 to $18 million is on the order of one percent of the capital cost of a 100 MW facility, and it is comparable to the utility-side capacity it avoids — in the Indiana case, new supply capacity is being built at roughly $22 per gallon-per-day, so the 0.6 MGD this tank avoids would cost the public system a similar $14 million [1]. The difference is who finances it and where it sits: private capital on private land, with no six-year municipal construction timeline and no revenue bond at risk. Third, a five-million-gallon ground tank is roughly 130 feet in diameter — a rounding error on a data center campus.

Why, then, are such tanks rare? Because under volumetric rates nobody has a reason to build one: the utility bears the peak, and the data center pays only for gallons. The academic literature observes that operators rarely rely on storage for multi-day peaks and attributes this to cost [1] — but the cost is modest; what is missing is the price signal. Charge for allocated maximum-day capacity, or cap the maximum draw in the service agreement, and the tank becomes the developer’s own cheapest solution. On-site storage is thus the clearest illustration of this paper’s thesis: correct pricing does not merely allocate the cost of the peak fairly — it causes the peak to be engineered away by the party best positioned to do it.

The honest caveats: the tank must be sized to the drought-of-record heat wave duration, not an average one; potable storage requires active turnover and disinfection management; the cap must be enforceable — metered, telemetered, with penalty ratchets, not aspirational; fire flow remains a separate requirement; and storage shaves the delivery peak without reducing consumption, so it solves nothing in a genuine supply shortage — the Section 2.4 assessment still comes first. None of these caveats changes the conclusion; they define the contract terms. The companion fiscal model’s capacity-audit and water-utility modules can be used to test these terms against a specific proposal [24].

KEY TAKEAWAY — For roughly one percent of project cost, on-site storage turns the peakiest water customer a system has ever seen into a flat one. Tanks are rare only because volumetric rates give nobody a reason to build them. Price the peak, and the developer builds the tank.

10. Conclusion: Not “No” — “Yes, If”

A community is rarely overwhelmed by the gallons or kilowatt-hours a data center consumes in a year — provided the supply exists. It is overwhelmed by the size of the pipe, the plant, and the wire that must stand ready for the worst afternoon of the summer — capacity that is financed every day of the year and used only on a few. The single most informative number in any data center proposal is therefore its load factor: average demand divided by peak demand — in other words, how much of the capacity reserved for the customer is actually used. It is also, not coincidentally, the number operators disclose least.

The full ledger, honestly stated, looks like this. On the benefit side: the best-shaped electric customer utilities have ever been offered, capable of putting downward pressure on everyone’s rates where headroom exists; and a fiscal engine — enormous taxable value demanding almost nothing in services — whose abatement is a bargaining chip the community controls. On the cost side: water load factors of 3 to 17 percent landing on heat-wave design days; wastewater with concentrated chemistry and its own peak; reserved-but-unused electric capacity; second-scale power transients; winter gas design-day exposure; and thirty-year debt issued against a customer that may cancel, under-ramp, or be obsolete in fifteen. The decisive fact is that every entry on the cost side has a known mitigating instrument. These are not exotic inventions; they are the standard toolbox of utility contracting, applied to a new customer class. The productive posture for a community is neither “no” nor an unexamined “yes,” but “yes, if” — and the “if” is a term sheet, not a sentiment:

• Verify supply first, under stress conditions — drought of record, heat wave, polar vortex — for water, power, and gas together. If supply fails, the conversation is evenhanded curtailment for all users, not a ban on one industry.
• Price capacity, not just volume: charges based on allocated maximum-day and coincident-peak demand, with ratchets so a customer who sets a peak keeps paying for it.
• Contract the mitigation, don’t request it: on-site storage or a telemetered max-day cap with penalties for the water peak; minimum takes for reserved electric capacity; on-site batteries for transients; firm gas transportation assigned to any on-site generation; pretreatment terms for the wastewater.
• Protect the debt: take-or-pay terms, term-matched financing, and developer-funded infrastructure that survive cancellation, under-ramping, and obsolescence.
• Trade abatements for protections — never grant both. The property tax base is the community’s leverage; spend it deliberately.
• Require disclosure as a condition of service: facility-level peak demand and load factor, published — because the political stumbling block is trust, and trust follows from terms the public can read.

A community that does this converts the peakiest water customer it has ever seen into a flat one, converts stranded-debt risk into developer capital, and keeps the two benefits — the ratepayer arithmetic and the tax base — that made the project attractive in the first place. A moratorium is what a community reaches for when its prices cannot yet say what its engineers already know. Utilities do not go broke selling water. They go broke building for water they rarely sell — and peaks, not totals, are what force them to build. Price the peak, contract the cure, and the right answer to a well-sited data center is yes.

References

[1] “Small Bottle, Big Pipe: Quantifying and Addressing the Impact of Data Centers on Public Water Systems,” arXiv preprint 2603.02705 (2026). Primary source for water peaking data, case studies (The Dalles, West Des Moines, Leesburg, Prince William, Indiana, Wisconsin), capacity valuations, and infrastructure cost figures. Reports daily peaking factors; converted to load factors in this paper (load factor = 1 ÷ peaking factor). https://arxiv.org/abs/2603.02705

[2] Norris, T., “The Puzzle of Low Data Center Utilization Rates,” Power & Policy, August 7, 2025. Distinctions among load factor, capacity utilization, and uptime; survey of industry load-factor assumptions. https://www.powerpolicy.net/p/the-puzzle-of-low-data-center-utilization

[3] Energy and Environmental Economics (E3), “Forecasting Large Loads in the Age of AI and Data Centers,” whitepaper, December 2025. Utility planning load factors (Duke ≈80%, Dominion ≈82%) and metered load-factor findings. https://www.ethree.com/wp-content/uploads/2025/12/E3Whitepaper_DataCenterForecasting.pdf

[4] “Electricity Demand and Grid Impacts of AI Data Centers: Challenges and Prospects,” arXiv preprint 2509.07218 (2025). Reports peak-to-average ratios of ≈1.0 for AI training and 1.5–2.0 for conventional data centers — equivalent to load factors of ≈100% and 50–65% respectively; transient characteristics. https://arxiv.org/html/2509.07218v3

[5] “Technical Challenges of AI Data Center Integration into Power Grids — A Survey,” Energies 19(1):137 (2026). Gradient-synchronization power oscillations and megawatt-per-second ramp rates. https://www.mdpi.com/1996-1073/19/1/137

[6] “Power Stabilization for AI Training Datacenters,” arXiv preprint 2508.14318 (2025). Facility-scale power swing magnitudes and mitigation approaches. https://arxiv.org/html/2508.14318v1

[7] SemiAnalysis, “AI Training Load Fluctuations at Gigawatt-scale — Risk of Power Grid Blackout?” (2025). https://newsletter.semianalysis.com/p/ai-training-load-fluctuations-at-gigawatt-scale-risk-of-power-grid-blackout

[8] Norris, T., et al., “Rethinking Load Growth: Assessing the Potential for Integration of Large Flexible Loads in US Power Systems,” Nicholas Institute for Energy, Environment & Sustainability, Duke University (2025). 76 GW headroom at 0.25% annual curtailment (≈85 hours/year). Summary coverage: Utility Dive. https://www.utilitydive.com/news/us-grid-headroom-flexible-load-data-center-ai-ev-duke-report/739767/

[9] West Des Moines Water Works, monthly financial reports, 2022–2025 (monthly withdrawal data for the West Des Moines hyperscale facility and comparison of large-user demand patterns), as compiled and analyzed in [1].

[10] Prince William Water service-area analysis (measured daily data center water load factor of 10% in 2024; Northern Virginia weighted actual water load factor of 27–29%), as cited in [1].

[11] U.S. Environmental Protection Agency, Drinking Water Infrastructure Needs Survey and Assessment; Clean Watersheds Needs Survey; and 2024 Report to Congress on water affordability (national funding needs of $1.3T+ over 20 years; 12.1–19.2 million households lacking affordable water service), as cited in [1].

[12] DgtlInfra, “Data Center Water Usage: A Comprehensive Guide.” Cooling system mechanics, evaporation, blowdown, cycles of concentration, and on-site water storage for reliability. https://dgtlinfra.com/data-center-water-usage/

[13] Vantage Data Centers, “Cooling Without the Drain: How Closed-Loop Systems Cut Day-to-Day Water Use” (2026). https://blog.vantage-dc.com/2026/04/22/cooling-without-the-drain-how-closed-loop-systems-cut-day-to-day-water-use/

[14] Shehabi, A., et al., “2024 United States Data Center Energy Usage Report,” Lawrence Berkeley National Laboratory, LBNL-2001637 (2024). Utilization data gaps and transparency needs. https://eta.lbl.gov/publications/2024-lbnl-data-center-energy-usage-report

[15] Royal, C., “More states are considering putting a pause on data centers,” The Bond Buyer, July 8, 2026. State and local moratoriums; Moody’s and S&P analyst commentary; 120 canceled projects (S&P 451 Research); Texas and Virginia tax-incentive figures; Data Center Coalition statements. https://www.bondbuyer.com

[16] Institute for Energy Research, “Natural Gas is Preferred by Data Centers and Manufacturers, and More Capacity and Pipelines are Needed” (2026), citing BloombergNEF analysis of ~100 GW of planned on-site gas-fired capacity for U.S. data centers. https://www.instituteforenergyresearch.org/fossil-fuels/natural-gas-is-preferred-by-data-centers-and-manufacturers-and-more-capacity-and-pipelines-are-needed/

[17] American Oil & Gas Reporter, “Ramping Data Center Demand Prompts Improved Coordination Between Gas, Electric Sectors.” Firm versus interruptible transportation; winter coincidence of data center and heating loads. https://www.aogr.com/magazine/markets-analytics/ramping-data-center-demand-prompts-improved-coordination-between-gas-electric-sectors

[18] Utility Dive, “Behind-the-meter data center gas plants will raise US energy bills” (2026). Pipeline-capacity effects of behind-the-meter data center generation. https://www.utilitydive.com/news/data-centers-raise-energy-bills-not-for-reason-you-think/822205/

[19] American Farmland Trust / Farmland Information Center, “Cost of Community Services Studies” (fact sheet summarizing 150+ studies): commercial/industrial land typically requires $0.35–$0.65 of local services per $1.00 of revenue generated; residential requires $1.15–$1.50. https://farmlandinfo.org/publications/cost-of-community-services-studies/

[20] Institute for Energy Economics and Financial Analysis (IEEFA), “Projected data center growth spurs PJM capacity prices” (2025), and Office of the People’s Counsel (D.C.) / Synapse Energy Economics, “Drivers of PJM’s Capacity Market Price Surge” (2025). Capacity prices rose from $28.92/MW-day (2024/25) to $269.92 (2025/26) and $329.17 (2026/27); Monitoring Analytics attributes 63% of the 2025/26 increase (≈$9.3B) to data centers; residential impacts of $16–21/month in parts of the region. https://ieefa.org/resources/projected-data-center-growth-spurs-pjm-capacity-prices-factor-10

[21] Vendor and industry cost guides for large ground-level water storage (welded steel, bolted steel, and prestressed concrete tanks to 5+ million gallons), indicating roughly $2–4 per gallon installed at multi-million-gallon scale (site-specific). https://www.tank-depot.com/blog/water-tank-prices-how-much-does-a-water-tank-cost

[22] North Texas Municipal Water District, “Facts About Data Centers: Considerations for NTMWD Member Cities and Customers” (2026). Texas data center counts and regional distribution; water and electricity use figures; cooling-system water guidance; wastewater chemistry (biocides, corrosion inhibitors, phosphates, nitrites, heavy metals; high-TDS blowdown); guidance to engage District planning and pretreatment teams early; PUC/TWDB surveys and end-of-2026 report to the Legislative Budget Board and Governor. https://www.ntmwd.com

[23] Cook, M., “Thirsty Data and the Lone Star State: The Impact of Data Center Growth on Texas’ Water Supply,” Houston Advanced Research Center (January 2026). Texas data centers used ≈25 billion gallons in 2025; projected 29–161 billion gallons annually by 2030 (up to ≈2.7% of state water use); 400+ facilities; ≈9,500 MW today with ERCOT 2030 forecasts of ≈22,000–78,000 MW; notes ERCOT has a large-load interconnection process while no parallel statewide water-planning process exists. https://harcresearch.org/research/thirsty-data-and-the-lone-star-state-the-impact-of-data-center-growth-on-texas-water-supply/

[24] McLain, L., et al., “Data Center Fiscal Impact Model — Municipal Revenue & Cost Analysis Tool for Texas Cities,” CityBase.Net, Interactive Model v1.0 (2026). Twenty-year net fiscal impact from the municipal perspective: property and sales tax, utility revenues and costs, power demand, employment, municipal service costs, capacity audit, scenarios, and case studies. https://datacenterfiscalimpact.netlify.app/