If you have been following the debate over the facility proposed for Textile Road, you have probably noticed that both sides use the phrase “data center” to mean very different things. That ambiguity is doing a lot of work in this conversation, so it is worth slowing down and defining terms.
This page is not an argument. It is the vocabulary you need to evaluate the arguments — including ours. Every factual claim below is footnoted to a primary source.
The basic definition
A data center is a building purpose-built to house computers and keep them running. Not desktop computers: racks of servers, storage arrays, and network equipment, packed into rows, operating continuously.
Everything else in the building exists to serve those machines. Electrical distribution to power them. Cooling to remove the heat they produce. Backup power for when the grid fails. Fire suppression, physical security, and network connectivity. A data center is, in a real sense, a large machine that happens to have a roof.
The computers do the work. The building keeps them alive.
They are not all the same size — and size is the whole conversation
This is the single most important thing to understand, because “data center” covers an enormous range. The word alone tells you almost nothing.
Facilities are usually described by their electrical capacity, measured in megawatts. The industry uses a widely cited threshold for the largest category: a facility is generally called hyperscale if it houses at least 5,000 servers, spans at least 10,000 square feet of floor space, and offers at least 40 megawatts of capacity. Anything smaller is conventionally described as an enterprise data center.1 At the other end of the range, a small data center typically occupies 5,000 to 20,000 square feet and hosts 500 to 2,000 servers.2
Two facilities can both be accurately called data centers while one uses as much electricity as a small office and the other uses as much as a mid-sized city. Facilities larger than 100 MW are a recent and still relatively uncommon development, and the Uptime Institute identifies their growth as the reason water use is becoming a contentious local issue in a way it was not before.6
This is why any claim comparing a proposed facility to “commercial data centers” needs a follow-up question: which ones? A comparison to the largest AI training campuses in the world makes a 100-megawatt facility sound small. A comparison to the actual population of data centers in existence makes the same facility very large indeed.
Understanding the power numbers
Two units get used almost interchangeably in public discussion, and they mean different things.
Megawatts (MW) measure power — the rate of electricity use at any given moment. This is the number that determines how much grid capacity a facility requires, what size substation it needs, and how much generating capacity a utility must have available to serve it.
Megawatt-hours (MWh) measure energy — power multiplied by time. This is what shows up on a bill.
The distinction matters for scale. A 100 MW facility running continuously consumes about 2,400 MWh per day. The average American household purchases about 10,791 kilowatt-hours of electricity per year,3 or roughly 30 kWh per day — so that facility’s daily consumption is comparable to that of roughly 81,000 households. Ypsilanti Township contains fewer than a quarter that many.
There is a third number worth knowing: power usage effectiveness, or PUE, which is total facility power divided by the power drawn by the computing equipment itself. A PUE of 1.0 would mean every watt goes to computing and none to cooling or electrical losses. The Uptime Institute’s 2025 global survey put the weighted average at 1.54, essentially unchanged for six consecutive years, with facilities of 20 MW and above averaging 1.44.6 This means a stated megawatt figure can describe either the computing load or the total facility draw, and those differ substantially. It is worth asking which one is being quoted.
For context on the sector as a whole: the Department of Energy’s Lawrence Berkeley National Laboratory found that U.S. data centers consumed about 4.4% of the nation’s electricity in 2023, with total usage rising from 58 terawatt-hours in 2014 to 176 TWh in 2023.4 Berkeley Lab’s 2025 update projects data centers could account for 11.8% of total U.S. electricity by 2030, within a range of 9.5% to 15.3%.5
When you see MW and MWh confused in a news article or a policy paper — and it happens often — treat the rest of the numbers with caution.
Two consequences follow from the fact that data centers run at close to constant load.
They are a firm load, unless designed otherwise. A facility that must run at full capacity around the clock forces a utility to have that capacity available at all times, including during the hottest afternoon of the summer when the grid is most stressed. Meeting peak demand is expensive.
Their infrastructure is site-specific, and someone pays for it. Large facilities typically require new substations and sometimes new transmission. A 2025 policy brief from the University of Michigan’s Ford School of Public Policy, prepared with the Michigan Environmental Justice Coalition, describes the mechanism directly: as demand surges, utilities commonly pass the costs of infrastructure upgrades and increased energy procurement on to residents and small businesses through higher rates.7 Whether that happens is a decision made by state utility regulators, and it is one of the most consequential decisions in any data center siting process.
Cooling, and why water comes up
Essentially all the electricity a data center consumes becomes heat. A 100 MW facility is, thermodynamically, a 100 MW heater. That heat has to go somewhere.
How it is removed depends on how densely the computing equipment is packed — and the density of an AI or high-performance computing facility is genuinely unusual. In the Uptime Institute’s 2025 survey of more than 800 operators, the average of typical rack densities was just under 9 kilowatts, and more than 80% of operators reported that their facility has no racks above 30 kW. Racks above 100 kW exist but remain rare, and Uptime notes that high-performance computing and AI training hardware is concentrated in relatively few sites.6 Those high densities are what require liquid cooling rather than air.
Air cooling uses fans and mechanical refrigeration. It consumes more electricity and little or no water. Water-based cooling — chilled water loops, evaporative cooling, direct-to-chip and immersion liquid cooling — moves heat more efficiently but uses water, sometimes a great deal of it.
The Ford School brief states the trade-off plainly: with current technology, a data center must prioritize either water efficiency or energy efficiency, because existing systems cannot optimize both at once. It reports that about 22% of data center facilities use water-based cooling, and that most facilities using water consume over 10 million gallons per year.7
When water use comes up, two distinctions matter.
Withdrawal is water taken from a source. Consumptive use is water that does not return to that source — mostly evaporated. A facility can truthfully say it does not withdraw from a river while still consuming very large volumes of municipal water.
Closed-loop systems circulate the same coolant continuously without ongoing withdrawal or discharge. This is genuinely different from evaporative cooling, and it is a meaningful commitment when a facility makes it.
So a promise not to draw from the river and a promise not to consume significant water are two different promises. So are a closed-loop system and a municipal water supply. Ask which one is being made. It is worth noting that industry-wide, only about 47% of operators track water usage data at all.6
Backup power
Data centers are built to keep running when the grid does not. That usually means two layers: batteries or flywheels that carry the load for the seconds to minutes it takes generators to start, and diesel or alternative-fuel generators that carry it from there.
How much backup a facility installs depends on what it does, and this is a real technical distinction rather than a marketing one.
Commercial facilities hosting customer workloads under contractual uptime guarantees generally back up close to their full load. Downtime costs them money and breaches agreements. Uptime’s outage data shows why operators take this seriously: power problems remain the leading cause of impactful outages, and one in five significant outages now costs more than $1 million.6
Research and high-performance computing facilities often do not. If the grid fails, a scheduler can save the state of running calculations, shut the compute floor down in an orderly way, and restart the work later. Only storage systems, networking, and critical cooling need to stay powered. Backup capacity well below full load is a normal and defensible design choice for this kind of facility.
That design has a consequence worth noting: a facility that can shut most of itself down on command is, by definition, capable of reducing its demand on the grid — which is a different and much less burdensome thing than a load that must run at full power no matter what.
On fuel: backup generators are combustion engines regardless of what they burn. Hydrotreated vegetable oil (HVO), also called renewable diesel, is a drop-in replacement for conventional diesel that requires no engine modification. Its environmental benefit is real but specific, and worth understanding precisely.
The headline reduction figure is a lifecycle number. Caterpillar reports the carbon intensity of HVO at between 15% and 78% of fossil diesel, depending on feedstock and pathway, using the Department of Energy’s GREET model.8 Cummins puts well-to-wheels emissions at roughly 70% lower than diesel after accounting for processing, transport and distribution.9 The reduction comes from the carbon being biogenic — recently captured from the atmosphere by the feedstock — not from the exhaust being cleaner in the ordinary sense.
On the pollutants that determine local air quality, the picture is narrower. Particulate matter drops meaningfully: Cummins measured 30% to 60% lower PM on HVO than on ultra-low-sulfur diesel in generator set testing.10 Nitrogen oxides largely do not. Cummins states that emissions of criteria pollutants such as NOx are comparable to those of diesel,9 and Caterpillar’s generator set testing found NOx equivalent to diesel at loads above 35% and below 15% of rated power.8 NOx is a precursor to ground-level ozone, and it is the emission most relevant to nearby residents.
A large reduction in lifecycle carbon is a genuine environmental benefit. It is not an answer to a question about what neighbors will breathe.
Also worth knowing: emergency generators do not only run during emergencies. Under EPA rules, emergency stationary engines may operate without time limit during actual emergencies, and for up to 100 hours per calendar year for maintenance checks and readiness testing, of which up to 50 hours may be used for certain non-emergency purposes.1112 Monthly load-bank testing is standard industry practice. That is a scheduled, predictable, recurring emissions event, not an emergency-only one.
How many people work in one?
Fewer than most people expect, though the honest answer is a range rather than a single number.
Construction employs a lot of people for a couple of years — electricians, pipefitters, operating engineers, concrete and steel trades — at roughly 0.7 to 2.0 workers per megawatt during the build.13 Those are real jobs, generally union, generally regional. That phase is genuine economic activity.
Operations are different, and much thinner. The Ford School brief puts it directly: data centers operate as infrastructure projects rather than job-creating businesses, and once built require relatively few employees because the facilities primarily house computers. It adds that the local jobs they do create tend to be low-wage, term-limited and non-technical — security, maintenance, custodial — and are frequently filled by contractors rather than employees.7
Published staffing estimates put operations at roughly 0.15 to 0.35 full-time equivalents per megawatt, with the most automated campuses above 100 MW running on as few as 20 to 30 permanent staff per 100 MW.13 The Brookings Institution summarizes the pattern: data centers are among the least labor-intensive structures in the economy, and large projects often promise only dozens to a few hundred permanent workers while the associated construction jobs are temporary.14
The right question for any specific project is not which end of that range applies in general, but what the operator will commit to in writing for this building.
One point causes confusion in nearly every siting debate: the researchers, engineers, and analysts who use a computing facility do not work inside it. They submit jobs to a scheduler from wherever they happen to be — a campus office, another state, another country. When employment figures for a proposed facility include people who use the computers, it is worth asking how many of those positions will actually be located at the site.
Questions worth asking about any proposed facility
If a data center is proposed near you, these are the questions that produce useful answers.
- What is the maximum electrical capacity in megawatts, and is that the IT load or the total facility draw? Ask separately for the announced figure, the substation design rating, and the total interconnection capacity requested from the utility. These are frequently not the same number.
- Who pays for the electrical infrastructure? Substations, transmission, and generation capacity are expensive. Are those costs assigned to the facility or spread across all ratepayers?
- What is the projected consumptive water use, in gallons per day? And what is the source — river, groundwater, or municipal supply?
- Is the cooling system closed-loop? If not, what happens to the water? And will water use be publicly reported?
- What is the backup generation capacity, how much fuel is stored on site, and what is the testing schedule?
- How many permanent positions will be physically located at the site? Not enabled by it — located at it.
- What does the facility pay in local taxes? Some are exempt from equipment taxes; publicly owned facilities may pay nothing at all.
- Can the facility reduce its load on demand? And if so, is it contractually committed to doing so during grid emergencies?
Every one of these has a specific, factual answer. A project that will not provide them has told you something.
Sources
- Data Center Dynamics, “What is a hyperscale data center?” datacenterdynamics.com
- IBM, “What is a hyperscale data center?” ibm.com
- U.S. Energy Information Administration, “How much electricity does an American home use?” (2022: 10,791 kWh per residential customer). eia.gov
- Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report, prepared for the U.S. Department of Energy under the Energy Act of 2020. full report (PDF) · DOE announcement
- Lawrence Berkeley National Laboratory, United States Data Center Energy Usage Report: 2025 Update. eta.lbl.gov
- Uptime Institute, Global Data Center Survey 2025 (Keynote Report 180, July 2025); survey of more than 800 data center owners and operators. PUE and rack density figures at pp. 6–10; water metric reporting at p. 11; outage cause and cost data at pp. 17–20. full report (PDF)
- Terry Nguyen and Ben Green, What Happens When Data Centers Come to Town?, Science, Technology, and Public Policy Program, Gerald R. Ford School of Public Policy, University of Michigan, in partnership with the Michigan Environmental Justice Coalition, July 2025. Cooling trade-off and water use at pp. 4–5; employment at p. 7; utility rate impacts at p. 8. stpp.fordschool.umich.edu (PDF)
- Caterpillar, “3500 and C175 Series Generator Sets Frequently Asked Questions” (renewable liquid fuels). cat.com
- Cummins, “Comparing emission reductions across alternative fuels.” cummins.com
- Cummins, “Generator Set Performance on HVO Fuel: QSK95 Test Report Summary.” cummins.com
- U.S. Environmental Protection Agency, “Understanding the Stationary Engines Rules.” epa.gov
- 40 CFR Part 63, Subpart ZZZZ, §63.6640(f) — National Emissions Standards for Hazardous Air Pollutants for Stationary Reciprocating Internal Combustion Engines. ecfr.gov
- Hamm Institute for American Energy, Data Center Employment Forecast Analysis, November 2025. hamminstitute.org (PDF)
- Bahar and Wright, “New evidence on data center employment effects,” Brookings Institution, May 2026. brookings.edu
Page last updated August 2026. If you find an error on this page, please tell us — we will correct it and note the correction.