Walk a watt from the grid to the chip and back out as heat, explore U.S. data-center geography, compare cooling and water systems, hear the sources of community noise, and investigate what public records still cannot tell us.
An AI data center is not a silent warehouse full of interchangeable computers. It is a power system, a high-speed network, a heat-removal machine, a water customer or air-cooled heat source, a construction project, a regulated industrial site, and—only after all of that—a place where a model can run.
A conceptual cutaway, not a specific campus. Colored lines organize the illustration; they are not traced electricity, water, or fiber. Actual equipment, ownership, workload, heat rejection, and water paths remain facility-level facts to verify.
Tap an AI app and the answer appears weightless. Somewhere else, electricity crosses a substation, thousands of chips switch, pumps and fans move almost all of that electrical energy away as heat, and a utility, water authority, grid operator, landowner, developer, cloud, model company, and local government each hold a different piece of the story.
This is the Infrastructure layer of the AMICE stack: Applications, Models, Infrastructure, Chips, and Energy. The research is frozen at August 13, 2026. Every interactive claim keeps its date, geography, denominator, status, and evidence type.
That discipline matters here. A map marker is not necessarily a facility. A campus plan is not operating load. A utility interconnection ceiling is not IT capacity. A company fleet total is not an AI workload allocation. “Liquid cooled” does not tell us whether the site evaporates water. “Powered by” does not tell us which generator served a particular hour.
Mapped feature
≠ facility
A point, building outline, and campus polygon may describe the same place.
MW
needs a basis
Critical IT, gross service, interconnection, generation, and planned buildout are different.
PUE
facility ÷ IT
Efficiency ratio—not total demand, useful work, carbon, or water.
Water
withdrawal ≠ consumption
Returned water, evaporation, source, watershed, and season all matter.
Megawatts can mean critical IT load, total facility service, an interconnection request, onsite generation, or a future buildout. GPU count can describe one building, a campus, a distributed cluster, or an operator's global portfolio. PUE divides total facility energy by IT energy; it says nothing by itself about total demand, carbon, water, embodied hardware, or useful work. Water use can mean withdrawal, discharge, or consumption.
The page therefore refuses to turn every disclosure into a single leaderboard. The honest unit is a dated claim with a boundary.
The lineage begins before the phrase “data center.” The publicly unveiled ENIAC occupied about 1,000 square feet, used roughly 18,000 vacuum tubes, and weighed about 30 tons in 1946; it was an ancestor, not a modern facility. NIST's SEAC then became the first fully operational stored-program electronic computer in the United States in 1950. The physical plant and the computer occupied nearly the same world. IBM launched the compatible System/360 family on April 7, 1964, letting institutions expand centralized computing without replacing their entire software environment at every step. Both the ENIAC history and Google's per-prompt water methodology below are included in the downloadable source ledger—not only linked in the prose.
The decisive hyperscale shift was not merely putting more servers in a larger building. In 2003, Google described search clusters with more than 15,000 commodity PCs and a filesystem designed around routine component failure. AWS put storage and compute behind on-demand APIs in 2006. In 2009, Google described the data center itself as one warehouse-scale computer: servers, network, software, power, and cooling designed together. Facebook then published clean-sheet server, rack, power, and facility ideas through Open Compute in 2011.
AI changed the density and the scale of the supporting systems. It did not invent centralized computing, industrial cooling, or large electricity demand.
From computer room to AI campus
Focus or select a dated event. The timeline separates documented milestones from company retrospectives and modeled national estimates.
2009 · hyperscale
The warehouse becomes the computer
Schematic only: cabinet count is illustrative and does not encode historical machine quantities.
What changed
Google described servers, network, software, power, and cooling as one warehouse-scale machine.
Why a consumer should care
The useful unit was no longer a server rack; it was a whole facility co-designed around one service fleet.
Evidence boundary
Novel as a whole-system engineering frame, not the invention of centralized computing or cooling.
Google treated ordinary component failure as a software assumption, stripped servers to the functions its fleet needed, built distributed storage and networking around enormous clusters, and optimized power and cooling at the scale of the whole service. The useful computer became the coordinated warehouse rather than the most reliable individual box.
That is a more precise claim than saying Google invented the data center. Mainframe rooms, colocation facilities, supercomputers, telecommunications exchanges, and industrial cooling all have longer histories. Google's published contribution was the unusually complete co-design of cheap machines, fault-tolerant software, network, power, cooling, and operations at internet scale.
There is also a common cooling myth to remove. Google did not start “active cooling” in the last few years. Its 2012 “Hot Huts” disclosure showed sealed hot aisles, forced air, pumps, and facility water cooling. In 2018, DeepMind described autonomous controls adjusting cooling setpoints every five minutes from thousands of sensors. The widely repeated 40% result referred to cooling energy, not total facility electricity.
The AI-era change was that liquid moved closer to the heat source. Google says TPU v3, deployed in 2018, was its first direct-liquid-cooled TPU generation. By 2025, it was describing accelerators above 1,000 watts, machine-learning racks expected to exceed 500 kilowatts before 2030, and power and cooling designs capable of supporting a 1 megawatt rack. Those are company roadmap and capability statements—not the density of an average rack.
Electricity enters through utility lines and substations, passes transformers, switchgear, batteries or UPS systems, and power distribution, then reaches CPUs, accelerators, memory, storage, and network equipment. Inside the computing boundary, almost all of it ends as heat.
That heat must cross several interfaces: the silicon junction, thermal material, heat spreader, heat sink or cold plate, server or rack loop, facility loop, and finally an outdoor sink. Small amounts of energy can leave as light, signals, or stored energy, so “almost all” is more defensible than a boundary-free 100%.
An exploded heat journey
Select a stage, then build a cooling system with two independent decisions: how heat leaves the rack and how it finally leaves the site.
Server
Accelerator junction
Transistors switch, memory moves data, and almost all input electricity becomes heat inside the facility boundary.
Water
Still no requirement for water at the silicon itself.
Sound
The chip is not the loud part; its supporting fans and pumps are.
Boundary
Small amounts of energy leave as signals or stored energy, so the page says “almost all,” not a boundary-free 100%.
Decision 1 · capture at the rack
Decision 2 · reject beyond the facility
The heat finishes in outdoor air without routine evaporation.
Water: Low direct cooling-water consumption; upstream electricity and chip-fab water remain.
Sound: Outdoor fans are continuous or load-varying sound sources.
Tradeoff: Hot design days can require more coil area, fan energy, or supplemental refrigeration.
Liquid cooling describes heat capture and transport—not the final heat sink.
The cooling vocabulary becomes much clearer when it is split into two independent decisions.
First, how does heat leave the rack? Fans can force room air through heat sinks. A rear-door exchanger can intercept hot exhaust. Cold plates can carry heat directly from processor packages into a sealed technology loop. Immersion can place electronics in dielectric fluid.
Second, how does heat leave the site? A dry cooler can blow outdoor air over a closed coil. A mechanical chiller can move heat across a difficult temperature difference. A cooling tower can evaporate water. A river, lake, or sea can absorb heat through an indirect exchanger. A district-heating or industrial customer can use some of it when temperature, distance, timing, and economics align.
These choices can be mixed. A liquid-cooled rack can finish at a dry cooler with little routine onsite evaporation. An air-cooled server can hand heat to a water-cooled chiller and evaporative tower. “Air cooled” does not mean passive, and “liquid cooled” does not mean water consuming.
“Passive cooling” usually means natural convection or building features that need little mechanical work. It is not a realistic complete heat path for today's dense AI racks. An airside economizer still needs controlled airflow, filtration, humidity management, and usually fans. A CRAC unit uses a refrigeration cycle at or near the computer room; a CRAH unit blows room air across a chilled-water coil supplied by a central plant. Both are air cooling at the server and active cooling at the facility.
The Department of Energy's data-center guide describes airside and waterside economizers, containment, cold plates, immersion, chillers, and liquid loops. “Free cooling” means compressor work is reduced or avoided under suitable conditions; fans, pumps, filtration, controls, and the final heat sink remain.
They do not all need to be.
Large-model training repeatedly synchronizes work across accelerators. Those machines exchange parameters, activations, checkpoints, and failures through extremely fast local networks. A large campus can share a substation, cooling plant, storage system, security boundary, operations team, and low-latency fabric. Dense racks reduce floor count, but they increase the intensity of power delivery and heat removal around each rack.
Inference creates a different geography. Smaller sites near population centers can reduce latency, keep regulated data inside a jurisdiction, improve failure isolation, and sometimes find nearby heat customers. Distribution also duplicates spare capacity and operations, consumes expensive land, and turns a local fabric into slower and less predictable wide-area links. The workload determines whether concentration or distribution is valuable.
Density calculator
These are transparent design arithmetic—not a forecast for a named facility. Change actual average IT load, rack density, and PUE.
1,000
illustrative racks
120
facility MW
20
overhead MW
One logical machine
Dense racks reduce floor count; they do not remove the systems around them.
Substations + redundancy
Low-latency fabric
Cooling plant + loops
Storage + checkpointing
Security + failure domains
Fiber + utility corridors
Staged buildings
Operations + repair
The calculator is arithmetic, not a forecast. Actual average IT load divided by average rack density produces an illustrative rack count. Multiplying IT load by PUE estimates facility load. None of those inputs can be safely inferred from a press release that announces land, utility service, or eventual campus capacity.
The chips are not usually what a neighbor hears. Sound comes from the equipment that feeds them and removes their heat.
Server fans, computer-room air handlers, rooftop units, cooling towers, and dry coolers move extraordinary volumes of air. Fan turbulence creates broadband sound while blades, motors, and bearings can create repeating tones. Chillers, compressors, and pumps add mechanical tones and structure-borne vibration. Transformers can produce a persistent low-frequency hum. Generator tests add combustion, exhaust, radiator-fan, and mechanical sound; an outage can make many units operate together.
The acoustic footprint
This is a source explorer, not a synthetic sound-level prediction. A defensible measurement needs distance, weather, operating state, background, spectrum, and protocol.
Bars indicate source activity and spectral visibility only. They do not encode decibels.
Selected source
Transformers
Electromagnetic forces vibrate the core and enclosure at persistent frequencies.
Mitigation to verify
≥15 sites
Virginia’s JLARC found resident-regarded problematic noise at at least 15 operational sites—about 10% of the state’s then-known sites. It is not a national prevalence estimate.
29%
Of 131 sites across eight reviewed Virginia localities were within 200 feet of residential zoning. That is zoning distance, not building-to-home distance.
dBA ≠ whole story
Low-frequency and tonal sound can remain noticeable at night even when an aggregate A-weighted limit is met.
A single A-weighted decibel number cannot describe every experience. Tonal or low-frequency sound may remain noticeable at night as background levels fall, and A-weighting can understate part of that spectrum. A defensible study needs source operating state, simultaneous load, distance, barriers, topography, weather, background, duration, and frequency bands. Source sound power is not the same as sound pressure at a home.
Virginia's 2024 legislative audit found resident-regarded problematic noise at at least 15 operating sites, roughly 10% of the state's then-known sites. That is a useful Virginia finding—not a national prevalence estimate. Better controls include larger slower fans, variable speed, quieter transformer designs, acoustic enclosures and louvers, vibration isolation, setbacks, barriers, equipment orientation, daytime staggered generator tests, and baseline plus post-opening spectral monitoring.
Evaporation can reject a great deal of heat with less compressor energy than all-dry cooling under favorable conditions. That trade is why water appears in many facilities. The DOE cooling schematic follows heat from racks to room air, chilled water, a chiller, a condenser loop, and a cooling tower.
Three water ledgers must stay separate:
In a cooling tower, makeup replaces evaporation, blowdown, and a small amount of drift. The evaporated molecules have left the liquid system. Meanwhile, calcium, magnesium, chloride, silica, and other dissolved material remain and concentrate. Without controlled blowdown, treatment, and monitoring, scale blocks heat transfer, salts corrode equipment, and biological growth creates operational and health risks.
The water ledger
Model site WUE separately from a normalized cooling-tower chemistry loop. The first calculator uses actual average IT load, not a campus announcement or interconnection ceiling.
115.7M gal/year
438,000,000 liters = MW × 8,760 h × 1,000 kWh/MWh × WUE. This is scenario arithmetic, not a named-campus prediction.
Balance ignores small drift for clarity: makeup = evaporation + blowdown. Higher cycles can reduce blowdown, but chemistry sets a safe ceiling.
modeled national · 2023
≈17.4 billion US gal
U.S. direct consumption
All U.S. data centers, not AI alone; modeled national estimate.
modeled national · 2023
≈211 billion US gal
U.S. grid-associated consumption
Electricity-generation water consumption modeled from balancing-area generation; excludes facility PPAs and behind-the-meter generation.
company fleet · 2025
13.562 billion US gal
Google data-center withdrawal
Alphabet-owned and fully leased data centers under its operational-control boundary; excludes seawater returned to sea, is not AI-only, and includes estimates where actual data were unavailable.
company fleet · 2025
3.039 billion US gal
Google data-center discharge
Same owned/fully leased reporting boundary; excludes seawater returned to sea and includes estimated discharge where actual data were unavailable.
company fleet · 2025
10.523 billion US gal
Google data-center consumption
Withdrawal minus discharge for the same reporting boundary, excluding seawater returned to sea; not a Gemini allocation.
workload subset · May 2025
0.26 mL
Median Gemini Apps text prompt
Company-estimated median for Gemini Apps text serving under a stated boundary; excludes model training, external networks, end-user devices, and data storage and is not a universal prompt constant.
Filtration can remove suspended particles but not every dissolved ion. Softening, chemistry, ion exchange, and reverse osmosis can raise safe cycles of concentration. Reverse osmosis produces a reusable permeate and a smaller, saltier reject stream. Zero-liquid-discharge systems add energy, chemicals, maintenance, and solids disposal. Condensing evaporated exhaust moisture requires another colder sink and more energy, undoing much of the reason the tower was used.
The DOE cooling-tower guidance says typical systems operate around two to four cycles of concentration and that higher cycles can reduce makeup and blowdown where local chemistry allows. That is not permission to apply one target everywhere.
Lawrence Berkeley National Laboratory modeled approximately 17.4 billion U.S. gallons of direct water consumption by all U.S. data centers in 2023. It also modeled roughly 211 billion gallons associated with electricity generation. The 2024 LBNL report is a national scenario model—not an AI-only total, a complete facility-meter census, or a number that can be assigned to one provider.
Alphabet's limited-assurance statement reports that Google's global data-center fleet withdrew 13.562 billion gallons, discharged 3.039 billion gallons, and consumed 10.523 billion gallons in 2025. The reporting methodology allows estimates where direct records were unavailable. These are fleet totals, not a Gemini allocation.
Google separately estimated that a median Gemini Apps text prompt in May 2025 consumed 0.26 milliliters of water under its stated serving boundary. The company study excludes model training, external networking, end-user devices, and data storage. It is useful evidence about one measured service population—not a universal “water per prompt” constant.
Water also has a place and a season. A modest annual share can still collide with a small utility's peak day or a drought. Reclaimed municipal water can reduce demand for potable supply, but it still needs treatment and produces discharge or residuals. Dry cooling can reduce direct site consumption while raising electricity use, shifting some water and emissions into the power system.
Cold outdoor air expands the hours when a facility can reduce compressor work. It does not remove the thermal interfaces between a high-power chip and the environment, and it does not answer the other siting questions: grid capacity, fiber, latency, data sovereignty, workforce, construction logistics, smoke, humidity, ice, drought, hazards, permits, and redundancy. A site must still survive its hot design day.
Training jobs can often tolerate more geographic distance. Latency-sensitive inference, content delivery, and regulated data often cannot. Cold weather is an engineering input, not a complete siting strategy.
Ocean water is possible too. Google's Hamina site reuses a former paper mill and existing tunnel to exchange heat indirectly with the Bay of Finland. Seawater does not flow through its servers. Microsoft's Project Natick operated a sealed research module on the North Sea floor from 2018 to 2020. Natick was an experiment, not a commercial fleet design.
Siting is a multi-variable problem
Switch the workload lens, then inspect a site archetype. The interface does not produce a fake universal score because power, fiber, latency, water, hazards, sovereignty, workforce, and permits change by project.
Cold, remote region
Location-flexible training or batch work when power, fiber, and jurisdiction also work.
What it helps
What comes with it
Marine cooling brings chloride corrosion, biofouling, intake screening, pumps, tunnels, heat exchangers, thermal discharge, coastal storms, and permitting. Water withdrawal can be high even when consumptive evaporation is low. The EPA's cooling-water intake work documents the risk of impinging larger organisms on screens and entraining eggs and larvae. Hamina is compelling partly because unusual industrial infrastructure was already there; it is not a template that every coastline can copy.
The broad atlas below freezes public OpenStreetMap features carrying documented data-center tags. It includes points, building outlines, and campus or site geometries. Those objects can overlap: one point, three buildings, and a campus polygon may all describe the same real place. Tags can be incomplete, stale, inconsistent, or absent.
That is why the map counts mapped features, not facilities. A construction tag does not prove completion. The lack of a construction tag does not prove operation. Operator and owner tags are preserved separately. A community record does not establish tenant, workload, chips, power, cooling, water, permit status, or capacity.
A broad map with a narrow claim
The frozen layer includes OSM points, buildings, and campus/site outlines matching documented data-center tags. They overlap and are incomplete, so the number is never called a facility count.
The empty fields are a result, not a defect. The public can see much of the industrial geography while remaining unable to answer what is operating inside it.
For a narrower set of AI campuses, the next atlas uses company, regulator, utility, and government disclosures. Each dossier keeps the landowner, developer, operator, cloud, workload customer, utility relationship, hardware, power, water, cooling, status, and unknown fields distinct. A company can occupy several roles without owning the whole stack. Google may control a model, accelerator design, network, cloud, and much of a campus while depending on a foundry, memory makers, utility, grid, and builders. xAI may control Grok and Colossus while depending on NVIDIA hardware, electric and water utilities, pipelines, and the manufacturing chain below the accelerator.
Facility geography
Public evidence, not telemetry. Grid region is not delivered electricity, a PPA or annual matching claim is not an hourly facility mix, and cooling design is not measured water consumption.
18
disclosed sites
8
countries
Abilene, Texas, United States
Parts of the campus are live and training and serving frontier AI systems while the larger build continues.
7
explicit unknowns
The live site runs on Oracle Cloud Infrastructure; this does not resolve every campus operating entity.
OpenAI identifies Crusoe as an Abilene campus partner in its cooling and community description.
Lancium describes Stargate I as its flagship Abilene Clean Campus project.
OpenAI reports running its own early training, inference, and frontier-research workloads.
Oracle began delivering NVIDIA GB200 racks in June 2025.
Lancium says the Abilene campus has an ERCOT-approved 1.2 GW interconnect.
Boundary: Interconnection capacity is neither measured demand nor evidence that 1.2 GW of compute is operating.
OpenAI early training, inference, and frontier research
OpenAI does not name a model or publish the fraction of workload assigned here.
NVIDIA GB200
Racks were delivered, but installed rack and GPU counts were not disclosed.
Grid / utility
Utility undisclosed · ERCOT
ERCOT approved an interconnect. The retail utility, delivered generation, and hourly fuel mix remain undisclosed; Lancium also describes on-site gas generation.
Water
Closed-loop cooling designed to recirculate an initial fill rather than use ongoing evaporative cooling water.
This is a design and company claim, not a meter reading; domestic, construction, and supply-chain water are not quantified.
Facility evidence
Reporter lead desk
These are unresolved, source-specific questions—not allegations. They point to the permits, utility dockets, customer allocations, and cooling disclosures that would materially improve the public map.
What cooling system and metered water use apply?
What powers the supercomputer, at what approved capacity, and under which operator?
Which GPU model and quantity are installed now?
Has the 50/50 owning venture closed and construction begun?
Which NVIDIA generation is contracted and how many GPUs are installed?
Who supplies power, and what metered water use will result?
Has the 200 MW first phase entered service?
What power, chip, and water contracts apply?
Which entity owns, builds, and operates the campus?
How much of the exploratory GPU offtake is assigned to Cobalt Park?
What is built, contracted, or commissioned?
Who provides power and cooling water?
What permits have been issued and for how many megawatts?
What water source and cooling design received approval?
Which chip generation is contracted?
What generation and storage projects have permits and financing?
What is the site-specific capacity and construction schedule?
What water source and cooling system are planned?
What capacity and in-service dates were approved in the dedicated utility rate?
What hardware has been ordered for each phase?
What water withdrawal and wastewater volumes are permitted?
What phase capacity and tariff were approved?
What water design and operating limits apply?
Which generation assets and batteries are incremental to this load?
What is the permitted phase size and construction schedule?
Which chips and models will be allocated?
Who supplies electricity and water?
Has any phase been commissioned since the groundbreaking?
What capacity, chips, and named workloads are planned?
Which utility and water systems will serve the site?
Has the campus completed construction or begun operation?
What legal entity owns the leased facility and what capacity was built?
Which chips, models, power contracts, and water systems apply?
What are permitted and measured water volumes and the delivered hourly electricity mix?
When will the first compute phase be commissioned, and at what capacity?
Which chips and Llama workloads will run there?
Has any phase passed commissioning?
How is the combined 1.5 GW split between Lordstown and Milam?
Who supplies power and water?
What are annual potable and non-potable water withdrawals measured at the site?
What compute capacity and chip count are actually energized today?
Which utility meters the campus, and what is the hourly delivered power mix?
What are metered potable and recycled-water withdrawals?
What is actual hourly load by grid, on-site generation, training, and inference?
Which named Grok models used the cluster and when?
What are the facility power rating, electric supplier, and hourly matching methodology?
What water withdrawals and consumption accompany the cooling and heat-recovery systems?
Which AI models and chip generations, if any, are allocated to Hamina?
How much water is withdrawn versus consumed, by cooling mode and season?
Which chip generations and AI workloads are physically allocated to Quilicura?
Who supplies electricity and what is the residual hourly generation mix?
How many Trainium2 chips and Claude workloads are in St. Joseph County?
What are this site’s power draw, utility, and annual cooling-water consumption?
Which other facilities form Rainier?
There is no defensible single list unless the denominator is fixed.
Meta describes Hyperion as an eventual 5 gigawatt compute buildout. Vantage describes Frontier as a planned 1.4 gigawatt critical IT load. Abilene's 1.2 gigawatts describes full-campus total power capacity, with only part of the campus live at the cited date. xAI reports more than 220,000 installed GPUs at Colossus 1. AWS describes nearly 500,000 Trainium2 chips across multiple U.S. data centers in Project Rainier.
Those facts are significant and incompatible. A planned campus, current draw, critical IT load, total service, installed accelerator count, distributed cluster, lease, and hardware description should never be silently sorted in one column.
“Largest” has no single denominator
Filter to compare a compatible disclosure basis. The unfiltered view is an evidence inventory, not a sorted leaderboard.
campus compute buildout
Richland Parish, Louisiana · Meta
5 GW
Eventual full-campus compute capacity
Boundary: Planned buildout—not present electricity draw or operating IT load.
critical it load
Shackelford County, Texas · Vantage Data Centers
1.4 GW
Ten-building campus critical IT load
Boundary: A full-build design disclosure; first delivery was targeted for the second half of 2026.
total power capacity
Abilene, Texas · Crusoe / Oracle Cloud
1.2 GW
Eight-building full-campus total power capacity
Boundary: The site was operating by this date; 1.2 GW is full-build total power capacity, not current IT load.
installed accelerators
Memphis, Tennessee · xAI
>220,000 GPUs
H100, H200, and GB200 GPUs disclosed at Colossus 1
Boundary: Do not combine with a separate campus or roadmap-to-one-million figure.
distributed cluster accelerators
Multiple U.S. data centers · Amazon Web Services
Nearly 500,000 Trainium2
Distributed U.S. cluster serving Anthropic
Boundary: Not a single campus, one map pin, or a site-level MW disclosure.
portfolio hardware description
Mount Pleasant, Wisconsin · Microsoft
“Hundreds of thousands” of GPUs
Campus architecture; first facility live and second planned for 2028
Boundary: No exact GPU count or site MW is public.
lease buildout
Milam County, Texas · SB Energy / OpenAI
1.2 GW
Initial leased buildout
Boundary: Contracted/buildout capacity—not measured electricity use.
campus compute buildout
Saline Township, Michigan · Oracle / Related Digital
1 GW
Full-campus plan
Boundary: Groundbreaking and design capacity do not establish operation.
campus compute buildout
Lebanon, Indiana · Meta
1 GW
Designed capacity once operational
Boundary: Construction-stage capacity, not current load.
campus compute buildout
New Albany region, Ohio · Meta
1 GW
Multi-building and adjacent-capacity cluster
Boundary: Cluster scope—not a single building’s current MW.
campus compute buildout
El Paso, Texas · Meta / BlackRock
1 GW
Compute-capacity plan expected to begin online in 2028
Boundary: Future capacity; Meta is the initial sole occupant, not necessarily the property’s only owner.
total power capacity
Childress, Texas · Crusoe / Lancium
1 GW
Full-scale grid-connected design
Boundary: Announced with construction expected later in 2026; no operating load.
Opposition is not one argument. A project can concentrate continuous sound, peak-day water demand, transmission lines, backup-engine emissions, industrial buildings, road traffic, tax incentives, and utility investment while creating far fewer permanent jobs than construction jobs. Benefits and burdens can coexist.
Scale also moves risk between parties. If a utility builds generation or transmission for a forecast load that arrives late, never arrives, or leaves early, who pays? If a site is dry cooled, what happens to its electricity demand during a heat wave? If it uses reclaimed water, where does blowdown go? If a generator permit allows thousands of hours, how many hours actually occur? If a sound study models the first phase, what happens when all phases run simultaneously?
No single permit answers all of those questions. Local government controls land use, setbacks, roads, buildings, and often noise. Water and sewer authorities control service and discharge. Air agencies permit engines or turbines. Utility commissions decide tariffs, contracts, and cost recovery. Utilities, regional grid organizations, and FERC address interconnection and transmission under different jurisdictions.
The permit stack
Land, electricity, air, water, wastewater, wetlands, roads, taxes, and construction can move on different calendars. Select an authority to see what it can—and cannot—decide.
Selected authority
Can decide
Does not prove or control
It usually does not dispatch the grid or set interstate wholesale transmission rules.
Current example
Effective July 1, 2026, Virginia requires a sound-profile assessment for a new ≥100 MW high-energy-use facility seeking rezoning, special-exception, or special-use approval; the statute has an exemption for certain previously approved sites and smaller expansions.
Community questions and targeted tools
Each mitigation fixes a boundary—not “the data center.”
Noise
Constant low-frequency or tonal sound, especially at night.
Useful response
Baseline and post-opening spectral measurement, tonal penalties, setbacks, barriers, equipment schedules, and a public complaint log.
Limit: Acoustic controls do not solve power, water, or land impacts.
Water
Peak-day and drought demand can matter locally even when statewide share looks small.
Useful response
Monthly and peak metering by source, reclaimed-water infrastructure, drought modes, dry/hybrid cooling, and public blowdown routes.
Limit: Dry cooling can increase electricity and capital; reclaimed water still needs treatment.
Power bills
Utility assets can become stranded if forecast load never arrives or leaves early.
Useful response
Long contracts, minimum bills, collateral, upfront contributions, staged ramps, curtailment, and exit charges.
Limit: These reduce risk; they do not guarantee zero regional price or transmission effects.
Air
Backup engines, bridge generation, and onsite turbines have different operating profiles.
Useful response
Equipment inventory, enforceable hour/fuel limits, aftertreatment, batteries where duration works, and actual-emissions reporting.
Limit: A permit proves an allowed envelope, not actual hours or measured compliance.
Land + jobs
Large industrial buildings, lines, roads, tax incentives, and temporary construction can reshape a place.
Useful response
Setbacks, corridors, wetlands review, enforceable hiring/community agreements, actual job reporting, and incentive clawbacks.
Limit: Benefits and impacts can coexist; one does not cancel the other.
Virginia’s unusually detailed 2024 audit · useful, not nationally universal
2.1B gal
Estimated 2023 data-center water use; just over one-third reclaimed. Less than 0.5% statewide, but 0.2%–21% of non-reclaimed demand at six reviewed utilities.
≈50 jobs
Typical full-time operations for a 250,000-square-foot building; roughly half contractors. Construction can peak near 1,500 for 12–18 months.
$14–$37/mo
Modeled 2040 generation/transmission effect for a typical Dominion residential customer—not an observed bill increase and not one campus.
$928M
Estimated FY2023 Virginia sales/use-tax revenue forgone under the data-center exemption; forgone revenue is not the same as a net fiscal loss.
The practical response is equally layered: monthly and peak water reporting; drought operating plans; dry or hybrid modes where appropriate; baseline and post-opening spectral noise tests; setbacks, barriers, and tonal limits; enforceable generator hours and emissions reporting; staged grid commitments; minimum bills, collateral, exit charges, and curtailment terms; wetlands and road review; public job reporting; and incentive clawbacks.
Each measure addresses one failure mode. None turns “the data center” into a single yes-or-no policy object.
You do not need to operate a chiller plant to understand the implications.
The strongest leads sit in the joins public sources do not make.
Which “liquid-cooled” campuses end at evaporative towers? How much site water is metered? What are monthly withdrawal, discharge, source type, and peak-day demand? Where do reverse-osmosis reject and cooling-tower blowdown go? Did the acoustic model include every phase at simultaneous load and examine one-third-octave bands? How much announced power is queued, contracted, energized, and actually operating? Who pays for stranded utility infrastructure? Which party is the property owner, developer, operator, tenant, cloud, workload customer, and utility counterparty? Is a GPU count tied to this campus or merely the company's portfolio?
Reporter lead desk
A mapped building is not a verified campus, a permit is not measured operation, and a company fleet metric is not an AI workload allocation. Every claim keeps its geography, date, denominator, evidence type, and public unknowns. Research frozen Aug 13, 2026.
Which liquid-cooled projects ultimately reject heat through an evaporative tower?
What are monthly withdrawal, discharge, source type, and peak-day water demand?
How much water reporting is metered versus estimated?
What happens to reverse-osmosis reject and cooling-tower blowdown?
Does the acoustic study cover every phase at simultaneous peak load and nighttime weather?
Are tonal and low-frequency bands measured, or only an aggregate A-weighted number?
How much announced MW is queued, contracted, energized, and actually operating?
Who pays if a utility builds assets for load that never arrives?
Which party is landowner, developer, operator, tenant, utility counterparty, and workload customer?
Is hardware disclosed for this facility or only somewhere in the operator’s portfolio?
Are jobs construction, permanent operations, contractor, indirect, committed, or realized?
Can heat-reuse claims be matched to a nearby year-round customer and signed offtake?
Dated source ledger
46 of 46 sources
The source ledger and evidence bundle are downloadable because this page should create reporting, not merely summarize it. The stable method is to keep physical location, legal organization, commercial role, facility status, measurement boundary, date, and source separate until a document actually connects them.