There is no single kind of data center. They range from a closet of servers to a campus drawing more electricity than a city. This page explains how data centers evolved and what separates an AI training campus from the servers behind everyday computing.
Almost everything we do online runs through a data center: online shopping, banking, streaming, telehealth, weather models — and data centers have been around for a long time.
Traditional data centers grew out of 1960s mainframes and were built around central processing units, or CPUs, for general computing — things such as email, browsing, sending money. In the late 1990s and early 2000s, graphics processing units, or GPUs, emerged and by the mid-to-late 2000s were adopted for scientific computing and supercomputers. This was followed by deep learning around 2012 and the first purpose-built AI server (essentially a dense cluster of GPUs) around 2016.
During the 10 years between about 2006 and 2016, hyperscale data centers emerged. Hyperscale describes size. It tells you nothing about the hardware on the inside. It simply means a facility with at least 5,000 servers on at least 10,000 square feet. The largest facilities today draw hundreds of megawatts. (What is a megawatt?)
Between 2016 and 2024, AI as we know it today developed, and the first purpose-built AI data centers emerged around 2024/2025. Where traditional data centers are designed to run millions of separate applications for multiple customers, purpose-built AI data centers run one enormous job across millions of pieces of hardware.
The rack density of these facilities is about 5 to 10 times larger, making them more compute intensive. This means more power, heat, and cooling. In fact, between 2014 and 2023, total U.S. data center electricity use rose from about 60 TWh to 176 TWh (Berkeley Lab, 2024), with AI servers being the fastest growing piece. Berkeley Lab's 2026 update estimates it reached 192 TWh in 2024 (Berkeley Lab, 2026).
A data center is a building full of computer systems, servers, and network equipment that is used to store, process, and distribute digital information. Data centers contain rows of metal frames called racks, each holding servers that require a network connection so data can get in and out, electricity to run, and a way to get rid of heat.
Computers turn nearly all the electricity they draw into heat, and they need additional systems to remove that heat and keep the equipment cool. How a facility does this largely determines how much water it uses.
Cooling is why these buildings are sited where power and water are cheap and plentiful rather than where the customers are. Cooling systems trade water against electricity. Select a method to read more.
Fans push cooled air across the servers. The air is cooled by electric chillers or, in cool weather, by outside air. On its own it uses little water but more electricity. It works for conventional servers but struggles with the densest AI racks.
Water is evaporated, usually in cooling towers, to carry heat away. This is the most energy-efficient option and the most water-intensive: the evaporated water is gone, and a large facility can consume millions of gallons a day in hot weather. It is often added to air-cooled systems to help on hot days.
Liquid circulates through sealed pipes, often directly to the chips, and is reused rather than consumed. This is the standard approach for dense AI hardware.
The heat still has to leave the building, though, and the loop hands it off to either dry coolers (radiators and fans: little water, more electricity) or cooling towers (evaporation: water consumed). "Closed loop" describes the inside of the system; it does not, on its own, say how much water the facility consumes.
Direct water consumption refers to the water used during cooling; and the amount of water an individual facility uses for cooling depends on the type of cooling technology used. In 2023, data centers in the United States used an average of 97 million gallons of water per TWh of electricty consumed.
There is a tradeoff between water and energy for cooling. As discussed above, when a data center uses less water, it often requires more energy. While systems that use less water for cooling can save on direct water, the increased energy consumption shifts the water burden to the source of power generation.
Indirect water consumption refers to the water used by gas, coal, and nuclear power plants for cooling during power generation. The amount of indirect water consumed depends on the source of power generation; for example, coal uses more water than natural gas. In 2023, assuming a grid mix, data ceters in the U.S. used an average of 1.2 billion gallons per TWh of electricty consumed. (Berkeley Lab, 2024)
When a campus builds its own power plant, like the microgrids described below, that water is used on or near the site and both direct and indirect water consumption is compounded locally.
The first four categories describe who owns the building and what it is for, not the
hardware inside. Any of these facilities can hold conventional servers, GPU servers, or a mix.
The AI training and inference category defines a facility purpose-built for AI applications, and
Microgrid data center campuses describe a data center campus with onsight power generation.
Select a type to read more.
A company running its own servers for its own operations — a hospital system's records, a bank's transaction processing, a university's research computing. Often a room or a wing rather than a separate building, staffed by the organization's own IT department. This is what most data centers were until about 2010.
A landlord builds the shell, brings in power and cooling, and leases space to many tenants who install their own equipment. The operator may not disclose, or even control, what its tenants run. This matters locally: when a developer is a real estate company rather than a technology company, the question of who will actually occupy the building can stay open long after the land is bought.
The facilities behind cloud services — storage, streaming, business software, email. This is the category most people picture when they think of "the internet." Run by a handful of very large companies, built to a repeatable template, and increasingly clustered into campuses of several buildings sharing one power connection.
Small and distributed, placed close to the people using them so that data travels a short distance. A cabinet at a cell tower, a room in an office building. Sited for proximity rather than power, so they rarely raise the land, water and electricity questions the large campuses do.
Built around dense clusters of GPUs rather than general-purpose processors. Training means building a model, which runs one enormous computation across tens of thousands of chips that must stay in step with each other. Inference means running the finished model to answer requests. Training is what drives the extreme power density; inference is steadier and can be spread across more locations.
These differ in purpose and size. A conventional facility is designed to run millions of unrelated jobs for many customers; a training campus is designed to run one compute intensive job.
Most data centers in the U.S. are grid connected. However, as the energy demand of data centers contiues to grow, companies are looking for solutions in the form of on-site power generation. In West Virginia, these are called microgrids, and they were defined along with High Impact Data Centers under the HB 2014 legislation.
A microgrid is an energy production facility that produces energy specifically to power an onsight data center. Currently, WV has 3 proposed microgrids that would each produce power on the scale of several gigawatts.
| Measure | Conventional cloud or colocation | AI training campus |
|---|---|---|
| Power per rack | About 5 to 15 kilowatts | About 50 to 140 kilowatts |
| Site power draw | Tens of megawatts for a building | Hundreds of megawatts to gigawatts for a campus |
| Load variability | Fairly steady, with a gentle daily cycle following user demand | Large synchronized swings as tens of thousands of chips start and stop a job together |
| Cooling method | Air cooling, sometimes with evaporative assistance | Liquid carried directly to the chip, with heat then rejected through towers or dry coolers |
| Permanent jobs | Dozens to low hundreds per campus in both cases. Staffing does not scale with megawatts — a facility drawing ten times the power does not employ ten times the people. | |
| What drives siting | Closeness to users and to major network routes, so that data arrives quickly | Availability of power above all, then land, water, and how fast a site can be energized |
Figures are typical ranges, not limits, and vary considerably between facilities. Sources: Berkeley Lab, 2024; Berkeley Lab, 2026; [list].
Only one project in the state has described its hardware. Based on the limited information that has been disclosed, several WV projects will likely fit the definition of a mega-campus. These could be purpose-built AI facilities, like what is being built in Mason County. However, the kind of data center for most project has to be inferred from power density or cannot be determined at all based on available information.
Documented
The Monarch Project is the first state certified AI microgrid and is in the early stages of development. The plan includes an initial 1.35 GW of critical IT capacity and up to 8 GW or more at full buildout. The developer describes the project as an "AI factory." (What is a gigawatt?) ADD POPUP DEFINITION
Documented
Although no information about a tenant or the hardware has been released for the Bedington Project, it has been designated as West Virginia's first High Impact "Intelligence" Center. At full buildout, it will have a critical IT capacity of at least 600 MW on 1.9 million square feet of development. According to the press release, the Bedington Project will help "transform the Mountain State into a powerhouse for artificial intelligence and cloud computing." Capacity on that scale and with that footprint is consistent with AI-class deployment.
Not disclosed
No tenant has been announced and no hardware described. The developer is a real estate firm, which is consistent with a colocation model where the eventual operator decides what goes inside. Until an operator is named, the cooling design, the water demand, the noise profile and the load pattern all remain open questions. County officials have raised reclaimed water as a possible source (see cooling), but no plan has been published.
Not disclosed
No megawatt figure has been published for the QTS campus. Without one, no independent estimate of its energy or water use is possible.
Select a site to find out.
The (The developer says) the campus is cooled exclusively with closed-loop systems and will not draw from Mason County's drinking water supply. In its air permit application to the state DEP, the (developer's consultant described) water-cooled chiller plants serving the buildings, with air-cooled coolers rejecting the heat, both on closed loops. (A company representative) has compared its water use to that of a small hospital. Two points are worth keeping in view: "no drinking water" is not the same as no water, and the campus's own gas-fired power plant will have cooling needs of its own.
No cooling design has been disclosed. The likely approach depends on what the eventual tenant installs. Dense AI training hardware generally requires closed-loop liquid cooling to the chip. Cloud and conventional servers are more often air-cooled, frequently with evaporative assist. The state's announcement names artificial intelligence, cloud computing, advanced manufacturing, and supercomputing, so any of these remains possible.
The one water-related statement on record is from the Berkeley County Commission (present, who said) officials had identified "a significant opportunity to potentially utilize reclaimed water".
Reclaimed water is treated wastewater, and it is typically used in cooling towers, which suggests water-based cooling is at least being considered. Using it spares drinking water, but water evaporated in a cooling tower is not returned to local streams, where treated wastewater is otherwise discharged.
To earn its HIDC designation, the project had to submit water and mitigation strategies to the state; those plans have not been made public.