Methods and sources

How Water Ledger's numbers are made

Model version and data date load below.

This is Michigan's full methods page. The model is the same in every state; how other states are built, and a methods page for each, are at methods for every state.

What it estimates

For a data center of a given size, cooling design and Michigan climate region, Water Ledger estimates three amounts of water on an average day, on the hottest days (a 1-in-10-year peak day) and over a year:

• Taken from the water system (what the utility has to supply).
• Evaporated (it leaves as vapor and does not come back).
• Wastewater to the sewer (mostly cooling-tower blowdown and staff use).

Water used at power plants to make the electricity is shown separately and never added to these on-site numbers.

It also shows the place before and after: public water, the sewer plant, runoff, sediment, phosphorus, road salt, generator fuel and electricity, today and with the campus.

How

1. Weather. Hourly temperature and humidity for 1995–2024 at eight Michigan locations, one per climate region (ERA5 reanalysis) ; wet-bulb temperature from Stull's formula .
2. Heat. The campus's electricity becomes heat that must be removed: computers, power losses, lighting, fans, pumps and chillers. Outside air, brought in to ventilate the building, adds or takes away a little more. Heat through the water or coil is all of that except the heat rejection fans, whose heat goes straight to the exhaust air (model 1.5). An economizer carries as much as the weather allows and chillers carry the rest .
3. Evaporation. Evaporative cooling removes heat by evaporating water (about 1,050 BTU per pound). The share of heat removed by evaporation rises with air temperature; we use the relation from Rutberg et al. (2011) as given in the U.S. Geological Survey's method for cooling towers .
4. Blowdown and drift. A small amount escapes as droplets (drift), a share of the water passing through the wet part of the cooler . Towers also drain water to control minerals. Minerals in must equal minerals out, so blowdown + drift = evaporation ÷ (cycles of concentration − 1) . The model samples 3 to 7 cycles. FEMP calls two to four typical and six or more possible, so the range sits above its typical one, and no figure for large campuses is published [A]. At the middle of each range, 2 to 4 cycles would take about 20% more water than 3 to 7 (1.5 times the evaporation against 1.25), all of it as blowdown.
5. Lost to evaporation and wastewater. Every gallon taken is either lost to evaporation (cooling, drift, leaks, humidifiers) or becomes wastewater, to the sewer or a stream (blowdown and nearly all sanitary use). Wastewater also carries the condensate from outside air, which was never taken from a water system. One rounding rule keeps them adding up: water taken to two significant figures, wastewater to the same place, each at the model's median, and water lost what is left, including the condensate. So the middle figure for water lost can sit a few percent above the model's own median for it (about 3% for cooling towers), since medians do not add. The likely ranges are the model's own. Numbers are shown to two significant figures (26,000; 6,200; 1.9 million), the same on every page and in the review, so the rounded figures may not add up exactly. Multiples are shown to one decimal below 10 (7.4 times) and two significant figures above.
6. Water that is not for cooling. Every design also takes some water that is not for cooling, in three parts (model 1.3). At 300 MW, staff water and maintenance together are about 6,900 gallons a day (likely 4,200 to 9,500), about 640 of it lost. Staff water: restrooms, kitchens and cleaning for the people on site, which Microsoft says is all its zero-water design keeps water for . It follows the number of people, not the megawatts: people on site = 20 to 60 plus 0.1 to 0.5 per MW of computing capacity, times 10 to 30 gallons per person a day, so about 130 people and 2,300 gallons a day at 300 MW (middle of the runs). Gallons per person come from EPA's onsite wastewater manual, about 9 a calendar day for office and industrial workers , Michigan's septic design criteria, 35 per factory worker per shift , and one applicant's about 30 . People per MW follow announced campuses . Virginia's auditors put a typical building at about 50 full-time workers once open, about half of them contractors , and Michigan's sales tax exemption needs at least 30 new jobs . About 10% of staff water is lost (drinking, cleaning that dries; sampled 5 to 15%) and the rest goes down the drain . Maintenance and cleaning: washing equipment and floors, janitorial use, fire-system tests and loop flushing, up to 2,000 gallons a day for the campus plus 2 to 20 per MW, about 4,500 at 300 MW; 3 to 15% of it lost . No site publishes metered figures, so the range is bracketed by what applicants have projected [estimate]: Hermantown, Minnesota's loop maintenance at the low end, Crusoe's 200 MW building at the high end, checked against Saline Township and Microsoft's letter to Tyrone, Georgia. This is the largest uncertainty in this water. Loop top-up: a closed loop is filled once and loses a little to small leaks, typically under 5% of its volume a year ; the page samples 1 to 5% of a loop holding 2,000 to 12,000 gallons per MW of computing [estimate, from applicants' fills], all of it counted as lost. That is tens to a few hundred gallons a day. Outside air: the designs here cool with liquid loops, chillers or towers and recirculate the air inside the data halls, so outside air comes in only to pressurize and ventilate (model 1.5). The flow is sampled from 100 to 1,500 cfm per MW of computing capacity, an estimate: ASHRAE 62.1's 0.06 cfm per square foot for computer rooms at about 5,000 square feet per MW gives 300, and no published flow for large campuses was found. Above an indoor dew point of 15 °C a coil condenses the extra moisture; that condensate goes with the wastewater. It is clean water and would not go to a drainfield, so with a septic system the flow and Michigan's septic rules are worked out without it, and it is assumed to go to a site drain or stormwater [A]. Below −9 °C, humidifiers may add moisture, using water that is lost; how many campuses do this is not published, so the share is sampled from none to all . A design that cools with outside air directly would use far more air and is not modeled. Landscape watering is not counted, since it depends on the site and climate; ask how the site will be landscaped. Before model 1.2, this was one "base use" of 0.001 to 0.008 gallons per kWh of computing, 10 to 50% of it lost, which grew with the megawatts and counted humidification; at 300 MW it was about 26,000 gallons a day, four to seven times what Michigan applicants have projected (see the checks below).
7. Cooling designs differ in when they evaporate: dry systems not at all; dry coolers with adiabatic pre-cooling evaporate water into the air ahead of the coil above a switch temperature (20 to 30 °C, an assumption: makers set it for each unit and publish none ), calculated from how far the air is above its wet-bulb temperature, pad effectiveness of 70 to 80% , a 9 to 13 K air temperature rise from makers' selection sheets , and fans slowing in wet mode, with bleed of 0 to 50% of evaporation by pad type ; hybrids in warm weather; cooling towers all year. These follow the system types in the national studies . The model runs the pads at full effectiveness above the switch temperature. A controller that wet the pads only enough to hold the air at the switch temperature would evaporate less: on central parameters, 12 to 41% less on average (West Michigan 41, Columbus 33, Ashburn 26, Austin 13, Houston 12) and 2 to 8% less on the peak day. Actual controls are not published, so the estimate is the higher of the two [A].
8. Uncertainty. Real designs and operations vary, so every estimate is run 200 times across the published parameter ranges below. The calculator shows the middle result in bold and the middle 80% as the likely range.

Mass and energy balances

Energy, per kW of IT load, each hour

Water, per kWh of IT load, each hour

The equations

• Energy: facility power + outside air heat = heat through the water or coil + heat rejection fans
• Heat through the water or coil: H = q + air handler fans + outside air heat + chillers + pumps, with q = 1/UPS + 1/(1 − distribution loss) − 1 + lighting per kW of IT
• Chillers: load × (1 − economizer share) × (w / COP water-cooled + (1 − w) / COP air-cooled)
• Tower fans: design fan power × (air needed / design air)³, from 10% of design air to all of it, where air needed goes with the tower's water over the enthalpy the air can pick up. Design is the 78 °F rating wet-bulb [A]. Dry cooler fans follow the same law.
• Outside air heat: air flow × (enthalpy outside − enthalpy indoors). A dehumidifying coil carries the latent part.
• Evaporation: H × w × evaporative ratio / latent heat. Assist evaporates at pads instead, from the air's own energy balance: H × depression / (switch point + air rise − air temperature + depression) / latent heat, where depression is pad effectiveness × (air temperature − wet-bulb).
• Drift: H × w / (8.34 × range) × drift share
• Blowdown: evaporation / (cycles − 1) − drift, from the dissolved solids balance: makeup × solids = (blowdown + drift) × cycles × solids
• Water: withdrawn + condensate = consumed + returned. The middle estimate takes withdrawn and returned at their medians, and consumed is the remainder.
• Lake-source: flow × density × specific heat × rise = heat = electricity, since the lake water is colder than the loops and needs no chillers or heat rejection fans [A]. Withdrawn equals returned at the pipe. The added heat then drives evaporation from the lake surface, so that balance is the lake's.
• Power plants, off site: generation = campus electricity / (1 − line loss). Plant water is generation × water per MWh generated, reported separately and never added to the campus's own water.

Terms

IT load, IT capacity	The power the computers draw, and the most they can draw. Utilization is load over capacity.
Facility power, PUE	All the electricity the campus uses. Power usage effectiveness (PUE) is facility power over IT load.
Sensible and latent heat	Sensible heat changes the temperature of air or water. Latent heat evaporates or condenses water without changing its temperature, about 1,050 BTU per pound.
Enthalpy	The heat air carries, sensible and latent together.
Economizer	Cooling without running a chiller, when outside conditions are cold enough.
Chiller, COP	A machine that moves heat from cool water to warmer surroundings using electricity. Its coefficient of performance (COP) is heat moved per unit of electricity.
Wet share (w)	The share of the heat going through evaporation: 1 for towers, 0 for closed loop and assist coils, and rising with air temperature for hybrid.
Evaporative ratio	The share of a tower's heat that leaves by evaporation, the rest warming the air. It rises with air temperature (Rutberg et al. 2011, as in USGS SIR 2013-5188).
Range	How much a tower cools its water, °F.
Drift	Droplets carried out of a tower in the air.
Blowdown	Water drained from a tower to keep dissolved minerals from building up. Pads drain a similar bleed.
Cycles of concentration	How many times more concentrated the tower's water is than the water coming in.
Loop top-up	Water replacing small leaks from a closed loop.
Outside air	Air brought in to pressurize and ventilate the building. It is a small flow, 100–1,500 cfm per MW of IT capacity [A], not used for cooling.
Dew point, condensate	Air cooled to its dew point starts to condense. Above the 15 °C indoor limit, a coil removes the extra moisture as condensate, water that comes from the air and not from the water system.
Humidifier	Adds moisture when the air's dew point is below −9 °C. How many campuses do this is not published, so the share is sampled from none to all [A].
Withdrawn, consumed, returned	The model's names for water taken; water evaporated or lost as droplets and leaks; and wastewater, sent to a sewer, septic system or stream.
Line loss	Electricity lost between power plants and the campus, about 5% (EIA, U.S. average 2018–2022). The national figure is used in every state, and a campus on the transmission system may lose less [A].

Parameters

Results by region (a 100 MW campus)

Middle result with the likely range (10th–90th percentile), for 100 MW of computing capacity including staff water at that size. Staff water grows more slowly than the computers, so a 300 MW campus takes less than three times these amounts with closed loop. "Water per kWh" is per kilowatt-hour of computing energy, for the model's cooling and loop water without staff water: evaporated, and taken (the basis national studies report).

How the estimates compare with published and reported numbers

Most operators do not publish a site's size, so few site-by-site checks are possible. We are not aware of a published study that checks a site-level data center water model against metered use at named sites; we show what we can, with its limits.

Water at power plants

Electricity, and the water used to make it

• Electricity by cooling design. Worked out hour by hour from local weather, in the same heat balance as the cooling water (model 1.6), after Lei and Masanet's hourly model : facility power = computers + power supply and distribution losses + lighting + air handler fans + chillers + pumps + heat rejection fans. An economizer carries as much of the load as the weather allows, and chillers carry the rest. Their efficiency is a share of the theoretical (Carnot) limit, set against code minimums for large chillers. Heat through the water or coil, the source of the evaporation, is that power and outside air's heat, less the heat rejection fans. Each hybrid chiller rejects heat through the same split as everything else: to evaporating water in the wet share, to air in the rest. Power losses and lighting are heat in the room's air, whatever cools the computers. Dry cooler and tower fans follow the fan laws, with a 10% floor. A tower's fans move the air the hour's heat needs, sized at the standard 78 °F rating wet-bulb [A] and capped by what air can carry, so they slow in cool weather. Dry cooler selection data are proprietary, so their approach temperature and air pressure are assumptions [A], tested in the sensitivity analysis. Liquid-cooled computers use a warmer water loop (ASHRAE classes W32 to W45), with 10 to 30% of their heat still removed by air on a cooler loop; the share of a campus that is liquid-cooled is not published, so all of 0 to 100% is sampled [A]. Lake-source cooling has its own electricity from its heat factor (below). Checks: cooling towers fall within or close to Lei and Masanet's published range for each climate zone tested, and closed loop against towers is about 1 to 2% more in Ohio and 4 to 6% in Texas with air-cooled computers, and 7 to 8% in Phoenix and Las Vegas, against 6 to 10% in a published Nevada comparison (a preprint). With liquid-cooled computers, closed loop comes out below towers in cool climates, as in LBNL's simulations . No peer-reviewed comparison of that case by climate was found, so it is a modeled result.
• Homes comparison. EIA's 2025 average for Michigan homes, 634.9 kWh a month across 4,538,056 residential customers , which also gives the campus's share of Michigan homes (residential customers include seasonal homes; Michigan has about 4.0 million occupied households). Homes' worth of electricity is rounded to the nearest thousand, so the few-percent differences between cooling designs stay visible.
• Average and peak. The size setting is the computers' full-load power (IT capacity). On average they run below it: utilization is sampled from 50 to 80%: EPRI measured 75% of nameplate at a large hyperscale site and 57% at co-location sites , and the IEA puts hyperscale at 53 to 56% . The peak is every computer running in the hottest hour of a typical decade (the 90th percentile of each year's hottest hour), when cooling works hardest, an estimate, since no public source gives a data center's hot-day draw. The utility has to be able to supply that peak. A utility's reported peak below its average load cannot be right, so it counts as missing and that comparison is left out.
• Compared with the local utility. The utilities serving a place come from the Michigan Public Service Commission's service-area map , with the share of the place each territory covers (territories overlap in places, so shares can add to more than 100%). Each utility's load comes from its EIA-861 report : its Michigan retail sales over the year divided by 8,760 hours (an average MW, counting customers who buy power from another supplier over its wires) and its summer peak demand where it reports one (multistate utilities and small utilities that file EIA's short form report no Michigan peak). The campus's average is compared with the utility's average and its peak with the utility's summer peak. Where more than one utility serves a place, which one would serve the campus depends on the site. Michigan's total is all retail sales in the state over 8,760 hours (about 100 million MWh in 2025, about 11,500 MW). The comparison is with the utility's customers, not with where its power is made: Michigan's utilities buy and sell through the MISO market. For closed loop, the page also gives the extra peak against cooling towers for the same computers: full capacity times the difference in the top of their PUE ranges (1.24 against 1.16).
• Water at power plants by source. Water consumed (evaporated) per MWh. Michigan's current mix 1.01 L/kWh (0.74 to 1.38), about 268 gallons per MWh: its plants' 2024 reports to EIA over the state's generation, with published factors where a report is implausible or missing, updated with 2022 reports and Siddik et al.'s regional value on the same basis (the same method in every state; the published regional values are not averaged in directly, since they count the same plant reports and treat them differently). Gas combined-cycle plants with cooling towers 205 (130–300); Michigan's coal plants 250 (100–317), since the large ones (Monroe, Belle River, Campbell) are cooled with lake or river water passing through once; Michigan's nuclear plants 396 (251–540), weighted by 2024 generation at Cook, cooled with Lake Michigan water passing through once, and Fermi 2, with cooling towers; wind 0 and utility solar about 1 (0–5), about 0.2 for Michigan's mix of the two ; the MISO average excluding hydropower 1.57 L/kWh, from Siddik et al. for 2022, which runs above plant data for the reasons given in the methods log . Withdrawals at power plants are much larger, about 59 L/kWh for Michigan's mix (38 to 90), about 98% of it returned warmer, so totals compare water evaporated only. Power lost on transmission and distribution lines (about 5%, an estimate) is not added, so power-plant water is slightly low.

Lake-source cooling and where wastewater goes

• Lake-source (once-through) cooling. Offered only with lake surface water. The flow follows from the heat load and the temperature rise: flow = heat ÷ (density × specific heat × rise). Heat is the computing load times utilization and a heat factor of 1.05 to 1.17: computers, power losses, lighting, air handler fans and pumps, with no chillers or heat rejection fans because the lake water is colder than the loops [estimate]. All of that electricity leaves in the water, so the heat factor is also the PUE. The rise is 6–10 °C, an estimate: Cornell's lake source cooling loop returns water about 9 °C warmer than it draws it . Nearly all of the water goes back to the lake.
• Extra evaporation from the lake. The returned heat makes some lake water evaporate. Following the USGS method for once-through power plants , the share of added heat lost to evaporation is the change in evaporation divided by the change in evaporation, conduction and radiation at the natural water temperature, using Ward's wind function at 5–15 mph (an estimate). Natural surface temperatures are NOAA's daily long-term averages for each lake ; Lake Erie stands in for Lake St. Clair. The result is about 40–50% of the heat, compared with roughly 75–85% in a cooling tower.
• Rules shown. Federal intake standards for new facilities with a design intake over 2 million gallons a day that use at least 25% of it for cooling ; state discharge (NPDES) permits for discharges to rivers and lakes .
• Where wastewater goes. Public sewer, a septic system on the site, a direct discharge to a river or lake, or treatment and reuse on site. The amounts are the same for each, except reuse. With full reuse, the water taken falls to the amount evaporated, by mass balance; the leftover salts leave as hauled brine or solids, which the page does not estimate.

Rain, runoff and water quality

A screening estimate of what the campus does to runoff and to the streams it reaches, before and after. The same calculation runs in Water Ledger's Python models (waterledger_models.runoff), and every parameter comes from one file, so the page and any report agree.

• Runoff. EPA Simple Method split by surface: annual runoff = rain × 0.9 × Rv, with Rv 0.95 on roofs and paving and 0.05 on fields, woods or lawn . Rain is the county's 1991–2020 average from NOAA . The campus is assumed to be half roofs and paving (40% of that roofs), an estimate the detailed calculator lets you change.
• Pollution in the runoff. Load = 0.226 × runoff (inches) × concentration (mg/L) × acres. Roof and parking-lot concentrations of sediment, phosphorus and nitrogen are from USGS measurements in Marquette, Michigan , with ranges from the Minnesota Stormwater Manual, the National Stormwater Quality Database and EPA's STEPL .
• Farm fields today. Fields lose most of their sediment by erosion in large storms, which runoff concentrations miss. Farmland sediment is therefore an estimate of eroded soil reaching the field edge, 0.25–1.0 tons per acre per year (U.S. cropland sheet-and-rill erosion averaged 2.67 tons per acre in 2017, and only part of it leaves the field) , with phosphorus on that soil at 0.03–0.06% by weight (an estimate). Nitrate lost through tile drains is not counted, so farmland nitrogen is low.
• Controls. Michigan's post-construction standard (EGLE MS4 permits and county drain commissioner rules): treat the first inch of runoff or 90% of storms with at least 80% sediment removal, and hold rate and volume to pre-development up to the 2-year, 24-hour storm . It applies inside regulated stormwater areas, and elsewhere only where the township or county adopts it. The data center numbers assume these standard runoff controls, modeled as wet ponds: sediment 80%, phosphorus 51% and nitrogen 26% removed from the 75–95% of runoff captured . The detailed calculator also shows infiltration, which soaks in about half the captured runoff and treats the rest (sediment 95%, phosphorus 70%, nitrogen 51%) .
• Construction. Construction sites erode at 10 to 20 times the rate of farmland ; the page uses 30 to 200 tons per acre per year before erosion controls, the range EPA reports from Wisconsin . The pages compare this with the sediment that leaves the land today. For farm fields that is the soil reaching the field edge (above), a fraction of what erodes on the field, so the shown construction figure is roughly 30 to 800 times today's rather than 10 to 20 times; the pages give both numbers rather than a multiple. Permits: Part 91 soil erosion and the construction storm water permit-by-rule .
• Receiving waters. The water's path follows USGS watershed (HUC12) links downstream . Listings are EGLE's 2026 Integrated Report in EPA ATTAINS; "cleanup plan" means a TMDL is in place (category 4A) . Listings cover all assessed waters in each watershed, so a listed lake can sit off the water's path. About 10% of Michigan's assessment units carry older watershed codes (the Huron River basin was recoded, for example); those are placed in today's watersheds by their own mapped location. Mercury, PCBs and PFAS in fish and E. coli are shown in the detailed calculator but not counted as runoff problems. Trout streams are Michigan DNR's designated trout streams, which are coldwater waters ; runoff may not warm them more than 2 °F , and summer runoff from pavement averages about 77 °F . The DNR layer lists only streams classified by type, so it can miss some designated waters.
• Sewer plants. All 394 Michigan wastewater plants in EPA's 2022 Clean Watersheds Needs Survey with a design flow and location, including lagoons and plants that discharge to groundwater under state permits ; last year's average flow, receiving water and combined-sewer status come from EPA ECHO where a federal discharge permit matches . ECHO flows that don't match the plant's size (often unit errors) are left blank. The page shows the place's own plant if one is in the same city or township, otherwise the nearest plant of half a million gallons a day or more; the utility decides which plant would actually serve a site.
• Cooling water to the sewer. Blowdown carries the makeup water's dissolved solids concentrated 3 to 7 times , and phosphate inhibitors can add up to about 6.5 mg/L of phosphorus , against the usual 1 mg/L plant limit . How plants permit data centers: .
• Road salt and fuel. Salt: 0.75 to 3 lb per 1,000 sq ft per event on parking and drives over 20 to 60 events a winter (an estimate; other states' pages scale it to each climate division's snow and ice days); chloride limit . Diesel: generator capacity of 1.2 times the computing load in 2.5 MW units (estimates), 66.8 gallons per MWh at full load , and 12 to 72 hours stored (12 hours is the Uptime Institute minimum ; reported campuses hold 1.8 to 3.3 million gallons ). Spill rules: .

Before and after

• Before is the place today from public data: public water used in the county (EGLE 2025) ; the sewer plant's flow last year (EPA ECHO) ; and the runoff, sediment and phosphorus the land sheds today, from the rain model above. Where nothing like it exists today (water evaporated on site, salt, diesel, the campus's electricity), before is "none".
• After is the scenario's middle estimate. Sediment and phosphorus assume Michigan's standard runoff controls. The construction row compares bare ground during the building years with the land today.
• Change. The simple view says More, Less or Same; Same only when the rounded numbers shown match. The same rule holds everywhere on the pages: a hottest day, runoff or power-plant comparison says "the same" only when the two numbers shown are equal, and otherwise says more or less. The detailed calculator shows each estimate with its likely range (the middle 80% of model runs for water and electricity, the listed low and high values for the runoff, salt and fuel estimates) and the change as a number.
• The same calculation opens Northline's pre-submission Water Review for a place, and an automated check confirms the page and the review give the same rows.

Starting assumptions and the notes under each setting

Each setting starts from what is known, and a short note says which: Known (published for a proposal), Your place (from public data for the place) or Assumed. Every setting can be changed.

• Cooling. A proposal's published design. Without one: closed loop, because 6 of the 9 Michigan proposals in the tracker have said closed loop. Where a proposal is undecided, hybrid as the middle case (Ypsilanti); where a reported water figure implies evaporative cooling, cooling towers (Augusta).
• Water source. The public system if EPA maps one in the place. A published source wins (Lowell's proposal buys City of Lowell water). Where EPA maps none [A], the page finds the nearest sizable system: one serving more than 3,300 people (EPA's medium size and up) that covers at least a quarter of a city or township, measured center to center between the places, since we have no map of water mains. If it is within about 10 km (the same reach as for the sewer), a closed-loop campus starts on that system, extending its pipes, usually together with the sewer; cooling towers and hybrid start on new wells, which avoid a utility's peak-day limits, with the nearby system named as the alternative. Beyond that reach, new wells. Well water with the public sewer is a common combination; the sewer card then asks how the wastewater will be metered, since there is no utility water meter. The source follows the cooling design until someone chooses one. The note explains the choice: a closed-loop campus needs little water, so it would most likely buy from the local system; a campus drills its own wells or builds a lake intake mainly when it needs a lot, as with evaporative cooling on hot days. A system is called small when it has fewer than 5,000 connections or the campus's hottest day would equal more than a quarter of its customers. Places within about 5 km of a Great Lake are told they could draw lake water directly.
• Wastewater. A proposal's published plan (Saline Township: an on-site plant discharging to the Saline River; Lowell Township: the City of Lowell sewer). Otherwise the public sewer [A] when the place has its own sewer plant; when EPA maps a public water system in it and a large sewer plant (0.5 million gallons a day or more) is within about 40 km; or when no system is mapped but a large plant is within about 10 km of the place's center, which a campus would reach by a sewer extension or an agreement. We have no map of sewer service areas, and EPA's water system boundaries often stop at city limits, so a missing system is weak evidence that a township has no sewer; data centers also tend to locate where utilities can reach. For evaporative cooling the plant must also have spare design flow (its design less last year's average) for the campus's average wastewater . Otherwise, closed loop's mostly sanitary flow goes to a septic system on the site (large tanks and a drainfield). Michigan's rules sort these by flow: under 6,000 gallons a day the county health department approves the design; 6,000 to 10,000 a day is permitted by rule, approved by the health department, metered and reported to EGLE each year; larger systems need an EGLE groundwater discharge permit . Blowdown from evaporative cooling, which a drainfield can't take, goes instead to the campus's own treatment plant discharging to the stream it sits on, under a state discharge permit. The note names the plant, says whether it is the local one or the nearest large one, and gives its design flow from the Clean Watersheds Needs Survey. It also says what the wastewater is for the chosen cooling: mostly sanitary flow for closed loop, salty blowdown for evaporative cooling, and for lake cooling the cooling water returns to the lake under its own permit. The destination follows the place, cooling design and campus size until someone chooses one.
• Size and land. A proposal's published size and acres. Otherwise a township starts at 300 MW on 200 acres of farmland (most recent data center sites have been farmland), and a city at 100 MW on 50 acres already built on, closer to the Southfield proposal (100 MW on about 12 acres). Half the site ends up as roofs and paving.
• Changing places. Choosing a new place starts every setting over from that place: a place with a tracked proposal loads the proposal's published or stated settings; any other place gets the defaults above, its own climate region, and Michigan's current power mix. Wastewater follows the rule above.
• Pictures. For water, each house is 1, 100 or 1,000 homes, whichever keeps the count readable (the key says which; 300 gallons a day per home) ; for electricity, each house is 10,000 homes at Michigan's average use; each golf flag is one typical Michigan course (the median of courses reporting to EGLE) . Icons are drawn smaller as the number grows, and a partly filled icon shows a fraction.

Local facts for your place

Every Michigan city and township (1,541 places with land area in the Census list) has a profile built from public data only:

• Where it is. Municipal and county boundaries from Census TIGERweb . Each county is assigned to the nearest of eight climate regions by distance from its center to the region's weather location.
• Public water. Community water systems whose service areas overlap the municipality, with the population and connections each reports, from EPA's Community Water System Service Area Boundaries . The share of the municipality each covers is measured by sampling points inside the municipal boundary; overlaps under 5% are dropped. EPA's boundaries are a mix of state-reported and modeled areas, so edges can be off. Where several are mapped, the campus is compared with one covering at least half of the place (the largest share if more than one), otherwise the largest by population among those covering at least 5%, otherwise the largest by population [A]. Where that system has under 100 connections and covers under half of the place [A], no percent is shown, and the page says a campus would need a larger supply.
• Watersheds. USGS Watershed Boundary Dataset subwatersheds (HUC12) covering the municipality, with shares measured the same way, and the river basin (HUC8) at its center . In sentences, USGS names are read naturally ("Headwaters Little Rabbit River" becomes "the headwaters of the Little Rabbit River"); tables keep the USGS names and codes.
• Great Lakes shore. The distance from the municipal boundary to the nearest Great Lake or Lake St. Clair, using Natural Earth's generalized 1:50m shoreline . Places within about 5 km get the lake surface water option. Withdrawal rules come from Michigan's Part 327 and the Great Lakes Compact review level as reported by WGVU .
• Groundwater. Household wells logged in EGLE's Wellogic records, by township: how many draw from glacial sand and gravel (drift) and how many from bedrock, with their average depth and static water level . These are averages of drillers' records, a guide to the local aquifer rather than a site study. New withdrawals over 100,000 gallons a day are screened by Michigan's Water Withdrawal Assessment Tool for their effect on nearby streams' summer flow .
• Design storms. NOAA Atlas 14 24-hour rainfall depths at the county center ; site runoff uses the EPA Simple Method with your inputs .
• Other water uses. EGLE's 2025 facility-level water use data, grouped by county into plain categories and averaged over the year . Golf courses are irrigation entries whose names mark them as golf courses or country clubs; the "typical golf course" is the median of those reporting more than zero (about 43,500 gallons a day averaged over the year; the middle half 20,000 to 77,000). These are withdrawals from a course's own wells, ponds, lakes and streams, which is the right comparison for what a data center draws from a water system or wells. Almost every course waters from its own wells or ponds, since city water would cost far more; any that use city water are counted in the city's total instead. Courses water mostly from May to September, so the summer-day figure (roughly 100,000 gallons) is the yearly total spread over 153 days, an estimate. Golf's own surveys put the North Central and Northeast among the lowest-water regions in the country (median 36.8 acre-feet a year for a Northeast 18-hole course in 2013) . One Washtenaw County farm irrigation entry (37.3 billion gallons, about 102 million gallons a day, many times the next-largest farm in the state) is left out as a likely reporting error and noted on that county's chart.
• Not counted. The calculator covers the campus's water, electricity and runoff on the site. It does not count new water or sewer connections, power lines, access roads or other construction beyond the site.
• Not shown, because no statewide public dataset has it: water plant capacity, current peak demand and wholesale contract limits. These come from the utility. Water Ledger does not guess them. Sewer plant design and recent flows are shown from EPA data (see above).

Limitations

• Actual designs, setpoints and operations are rarely published; when an operator publishes a figure, Water Ledger shows it first.
• Cooling towers are shown as an upper bound. Utility records from Northern Virginia show that most "water-cooled" data centers use far less than a pure tower would, because they run economizers and hybrids .
• Because computing load is steady, physics alone gives tower peaks only about 1.3 times the average day; utility data suggest 2–2.5 for tower-heavy portfolios . Read tower peaks as conservative. The peak day is at each run's average utilization, 50 to 80% [A], since campuses seldom run every computer flat out all day. As an upper case, with every computer at capacity all day, the peak day is about 1.5 times that: about 3.6 million gallons instead of 2.4 for 300 MW of cooling towers in West Michigan (withdrawn_peak_full_capacity in the package).
• One-time fills of closed loops (reported at about 0.05–1.5 million gallons per loop) are not included in daily averages.
• Reanalysis weather slightly smooths local extremes. Regions are broad: a place is given the weather of its region's station, which can be up to about 100 miles away. A site-specific review would use the nearest station.
• County water-use figures cover large registered users only . Public water is counted in the county where the utility draws it from a lake, river or wells: water a utility buys from a system in another county (for example, Great Lakes Water Authority water piped from Detroit) is not in its own county's total, so the campus's share can read high there.

About the data

Water Ledger is a free public resource from Northline Hydrology. It uses public data only; nothing from client work goes into it. It does not rate projects as good or bad; it shows what they could mean for water so people can decide with numbers instead of guesses.

Sources

What has changed in the numbers, and when: changelog. See something wrong? Suggest a correction.