Enterprise
A business or institution runs computing for its own operations. This can range from a server room to a substantial dedicated facility.
Behind the digital world is a physical one. Explore the machines, resources, money and decisions shaping the next generation of data centers.
Different places, years and measures, deliberately labeled. The global figure and U.S. forecast come from separate models. Select any number to inspect its source.
25 chapters · follow the sequence or choose a subject
No matching chapters. Try a broader term.
Use the chapter rail or Back / Forward. Left and right arrow keys work when a control is not focused. Source badges open records without losing your place. Read all shows one continuous document. Core text, images and tools work offline; external sources and media require internet access.
Every digital service rests on buildings, machines, people and a place.
A data center is the physical home of computing: servers process information, storage keeps it, networks move it, and electrical and cooling systems keep the equipment running. Email, banking, hospitals, video, scientific computing and artificial intelligence all depend on this infrastructure. The term does not identify a single building size, business model or resource footprint.
An AI facility is designed or equipped for AI work. It may be a dedicated campus, a rented hall inside a colocation building, or an upgraded part of an existing cloud region. A single operator can run conventional cloud services and AI in the same broader campus. A national total for all data centers should never silently become a total for AI alone.
A business or institution runs computing for its own operations. This can range from a server room to a substantial dedicated facility.
A landlord supplies space, power, cooling and connectivity. Tenants install or operate the computers. The visible developer may not be the final user.
A large cloud or internet platform operates at enormous scale, across many campuses and regions. A “region” is not necessarily one building.
Accelerator-rich computing for training and serving models. Density and workload differ across generations and applications.
Cryptocurrency mining is another category of computational load. Its hardware, economics and ability to pause are different. A “large-load” queue may include mining, manufacturing and other projects as well as AI; adding all requests together does not reveal AI demand.
Once the building types are clear, the physical systems inside them are easier to follow.
Continue: Inside the machine. →Follow the electricity in, the information through, and the heat out.
A data hall is one part of a much larger operating system. The fence may enclose substations, transformers, power electronics, backup batteries, generators, cooling equipment and water-treatment systems. Fiber routes and new transmission lines can extend far beyond the property.
NVIDIA’s GB200 NVL72 provides a concrete example: 72 Blackwell GPUs and 36 Grace CPUs connected in a liquid-cooled rack-scale system. Other architectures use different chips, layouts, networking and cooling; this named system makes rack-scale computing concrete.
Higher power density concentrates more electricity and heat into a small floor area. A building that has spare square footage may still lack electrical capacity, structural capacity or the right piping. Retrofitting a conventional data hall can therefore be a substantial engineering project.
The maximum or contracted capability of equipment or a connection. It can include reserves and future phases.
What the facility consumes at a particular time. It changes with occupancy, workload, temperature and operating choices.
The purpose of those machines helps explain their density, location and operating patterns.
Continue: What are we building it to do? →Training, inference and ordinary cloud computing have different needs.
The model’s parameters are adjusted using data. Large training runs coordinate many accelerators. Research, testing, retraining and fine-tuning make this an ongoing development process, not one permanent, once-only event.
A trained model produces an answer, prediction, image, recommendation or other output. Work ranges from a short text request to long reasoning, video and multi-step agents.
These workloads overlap with ordinary cloud services, enterprise software, advertising, scientific simulation and storage. A campus can support several categories. AI may be the reason for expansion without being the only workload on the premises.
The IEA’s 2026 assessment finds large improvements in energy per AI task, alongside a shift toward more demanding applications. A single short query cannot stand in for all AI use.
A smaller energy cost per task can be offset by more users, more tasks per user, longer outputs or more complicated work. Conversely, improved models and hardware can reduce resource use for a fixed workload. The eventual total depends on all of those moving parts.
| Workload feature | Why it matters to a community |
|---|---|
| Response time | Some services benefit from proximity to users. Some batch work can run farther away. |
| Scheduling flexibility | Work that can move in time may respond to power or water constraints. Contracted reliability can limit this. |
| Communication intensity | Large training jobs need fast links within a cluster; a distant spare computer is not always a substitute. |
| Utilization | Idle equipment still draws power. Occupancy, computational activity and electrical load factor are different measures. |
A practical question for the developer: what work will run here, what flexibility is contractually available, and what public performance reporting will show it?
In its April 2026 outlook, the IEA reports that capital expenditure by five major technology companies exceeded $400 billion in 2025 and was expected to rise by about 75% in 2026. This is company-wide capital expenditure, not a measured total for U.S. AI data-center buildings. It includes equipment and other assets, and expected spending is different from completed investment.
The same report estimates global data-center electricity rose 17% in 2025, while electricity at AI-focused facilities grew 50%. These modeled growth rates describe different populations. They help explain the construction interest without establishing that every proposed campus will be needed or financed.
Continue: The people behind the buildout. →The industry is an ecosystem, not one company and one permit.
Chip designers, fabricators, cloud providers, AI labs, colocation landlords, utilities, lenders and contractors perform different jobs. A cloud customer may rent computing rather than own a building. A developer’s announcement can include a tenant’s equipment investment and spending over many phases. Read the ownership and financing structure before adding headline dollar amounts.
| Example | Role in the infrastructure |
|---|---|
| NVIDIA / AMD | Design processors, accelerators and related computing platforms; suppliers rather than the operator of every building using their chips. |
| Amazon Web Services | Provides cloud services and its Trainium family of custom AI accelerators. |
| Google Cloud | Provides cloud services and custom Tensor Processing Units (TPUs). |
| Microsoft Azure | Provides cloud infrastructure and AI services to customers. |
| Meta / Entergy Louisiana | The Richland Parish example links a large computing customer to a utility’s planned generation and transmission buildout. |
The companies are examples of different functions, not a market-share ranking. An organization can occupy several positions in this chain.
| Participant | What to understand |
|---|---|
| Operator / tenant | Who consumes the electricity and runs the equipment? Who is contractually responsible if the tenant leaves? |
| Developer / landlord | Who owns the land and buildings? Are approvals transferable? |
| Utility / grid operator | Who studies capacity, schedules connections and plans shared infrastructure? |
| Water supplier | Who controls treatment, storage, pipes, supply rights and drought restrictions? |
| Local government | Who controls zoning, site plans, setbacks, noise and local fiscal agreements? |
| State / federal bodies | Which commissions and agencies oversee utility rates, environmental permits and other relevant requirements? |
Land-use approval, grid service and water allocation follow separate decision tracks. The governance chapter returns to who controls each one.
That chain of participants comes together at a particular parcel of land.
Continue: Why here? →A proposed site is the intersection of land, power, water, fiber and time.
Data centers cluster because existing infrastructure has value: fiber networks, reliable electrical service, suppliers, specialized workers and nearby customers can attract the next project. Virginia’s legislative audit identifies these factors in Northern Virginia’s growth.
A parcel with a credible connection schedule can be more useful than a cheaper parcel waiting for major transmission.
Latency, redundancy and fiber access shape where different workloads can operate.
Large contiguous parcels, suitable terrain, compatible zoning and predictable approval processes can shorten development.
Outdoor conditions and dependable supplies affect cooling options, cost and operating flexibility.
A campus can require off-site power lines, substations, generation, pipelines, roads or water infrastructure. Some facilities may share those assets; others may be dedicated. The host municipality and the people affected by new infrastructure are not necessarily the same population.
A realistic map should show the campus, nearby homes, schools and sensitive sites, the source watershed, the utility service area and any planned off-site works. Exact security-sensitive layouts are not needed to understand public impacts.
For a local proposal, compare the selected parcel with feasible alternatives. Ask which constraints genuinely rule alternatives out, which are matters of cost or timing, and which impacts can be reduced through site design.
To understand why a site can become difficult to serve, first put its proposed size into units that can be compared.
Continue: How big is big? →Power is a rate. Energy is what accumulates over time.
A megawatt (MW) measures power, like the rate of flow through a pipe. A megawatt-hour (MWh) measures energy, like the amount that flowed over an hour. A gigawatt is 1,000 MW; a terawatt-hour is one million MWh. A proposed 1 GW campus does not automatically consume 8.76 TWh a year: that assumes full draw every hour.
Berkeley Lab’s June 2026 update puts U.S. data-center electricity at 649 TWh in 2030 in its reference case, or 11.8% of forecast national electricity use. Its compounded uncertainty range is 521–843 TWh (9.5–15.3%). These are model scenarios for all data centers, not AI-only demand or a probability interval.
The IEA’s April 2026 update estimates 485 TWh globally in 2025 and projects roughly 950 TWh in 2030. These numbers come from a different model and geography. Do not splice the two organizations’ projections into a single harmonized forecast.
| Boundary | Example of a mistake |
|---|---|
| IT versus facility power | Comparing a server-only rating with a whole-campus grid connection. |
| Average versus peak | Calling an annual household grid-purchase equivalent a count of homes that could be served during a peak emergency. |
| Operating versus proposed | Counting future campus phases, speculative queue requests and live buildings as though all were consuming power now. |
Use the Scenario Lab to try the arithmetic with explicit assumptions. The household comparison uses a labeled historical EIA baseline, which you can change.
The national total sets the scale; the connection process determines what can actually arrive at one address.
Continue: From announcement to connection. →Electricity must arrive at the right place, at the right time, reliably.
Enough annual energy somewhere in the region is not the same as enough deliverable capacity at this site. Generation, transmission, substations and distribution must work together. Planning also needs reserves for failures and extreme conditions. Adding a major load changes more than annual sales.
Can the system supply the added energy and dependable capacity when needed? New plants and fuel arrangements take time.
Can power move through the relevant network without overloading lines or violating operating limits?
Are transformers, substations, switchgear, studies and construction ready for the promised energization date?
How quickly can load change? What happens during faults, heat waves, equipment failures and emergency curtailment?
In June 2026, FERC issued orders to six regional grid operators under its jurisdiction to justify or reform large-load rules, including cost allocation, flexible service and co-location. That action does not create one uniform tariff for the entire United States.
A developer may propose an on-site gas plant, batteries, a renewable contract or co-location with an existing generator. Each arrangement has different physical and financial consequences. A site that expects backup from the public grid still needs its dependence studied, even if much of its routine supply comes from elsewhere.
Batteries can bridge interruptions and manage short variations; their duration, usable capacity, control permissions and recharge needs determine what service they can provide. An existing generator’s output committed to a new campus may also change what is available to other customers.
A technically feasible connection still leaves a financial question: who funds it, and who carries the risk?
Continue: Who pays for the new demand? →The electricity bill depends on the system and the contract.
New load can spread existing fixed costs over more sales. It can also require expensive new infrastructure or push a tight supply market toward higher prices. The result depends on timing, spare capacity, procurement, tariffs and what happens if promised demand never appears.
Berkeley Lab’s 2026 price review describes places where load growth accompanied lower inflation-adjusted prices. A separate 2026 working paper estimates modest average rate reductions from data centers during 2015–2024. Neither guarantees the outcome of the next campus.
PJM’s independent market monitor attributed $29.4 billion of capacity charges across four auctions to data-center load in its July 2026 report. That is a modeled counterfactual involving existing and forecast demand, not a household-bill increase of the same percentage.
| Cost | What is being paid for |
|---|---|
| Energy | Electricity actually produced and delivered, usually measured in MWh. Fuel, generation mix and the hour of use matter. |
| Capacity, where that market exists | A commitment that resources will be available in a future delivery period. PJM’s auction figures concern this account; not all U.S. regions use PJM’s market design. |
| Networks and service | Transmission, substations, connection facilities and other infrastructure. The tariff and service agreement determine how costs are recovered. |
| Contract feature | What it addresses |
|---|---|
| Minimum payment / term | Reserved infrastructure may need payment even if the campus uses less than expected. |
| Collateral / credit support | A deposit, guarantee or other security may reduce the chance that others inherit an unpaid obligation. |
| Exit and cancellation provisions | Who pays the remaining costs if development stops, shrinks or relocates? |
| Dedicated cost allocation | Which connection assets and shared upgrades are assigned to the large customer? |
| Milestones and phased service | Infrastructure spending can follow demonstrated construction and contracted demand. |
Ohio offers a concrete example: its consumer advocate describes a July 2025 AEP tariff requiring qualifying new data centers to pay for at least 85% of contracted capacity for up to 12 years, and notes a later industry appeal. The docket, rather than a national headline, governs the legal details.
Power is only one resource account. Cooling connects electricity demand to a second system: water.
Continue: Two cooling questions. One crucial distinction. →How does heat leave the chip? How does it leave the site?
A liquid-cooled chip transfers heat into a circulating fluid. That fluid then transfers heat to another system or to the outdoors. The final stage may use dry coolers, chillers, cooling towers, or a combination. A sealed loop beside the chip can coexist with an evaporative cooling tower outside.
| Heat-rejection approach | Water considerations | Energy / site considerations |
|---|---|---|
| Evaporative cooling | Some water evaporates; additional makeup replaces evaporation and blowdown. | Can reduce cooling electricity in suitable conditions. Performance depends on weather and design. |
| Dry cooling / air-cooled chillers | Can greatly reduce operational cooling-water consumption. Other site water use remains. | May require more electrical power or larger equipment in hot weather. |
| Hybrid systems | Water use can vary by season and operating mode. | Can shift the energy–water tradeoff; peak operating plans matter. |
Microsoft announced a new design in December 2024 intended to avoid evaporating water for cooling. This is a specific company design claim, not proof that every existing Microsoft facility uses it or that the entire supply chain consumes no water.
Ask for the complete cooling diagram, the heat-rejection method at design-day conditions, the source of makeup water, and any expected changes between phases. “Liquid cooling,” “closed loop” and “water efficient” alone do not answer those questions.
PUE divides total facility energy by IT energy over the same period. At PUE 1.25, every 100 units used by IT require 125 units at the facility: 25 for overhead. Overhead is 25% of IT energy but 20% of the whole-facility total. A lower PUE reduces overhead for a fixed IT workload; it does not measure how much useful computing is delivered.
Following the heat explains the machinery. Following the water explains which resource is being counted.
Continue: Follow the water. →Withdrawal, consumption, discharge and indirect use are different accounts.
| Term | What it means | Why a reader should care |
|---|---|---|
| Withdrawal | Water taken from a source or delivered by a supplier. | Treatment and pipe capacity must accommodate the delivered flow. |
| Consumption | The portion withdrawn and not returned to the immediate water environment, often because it evaporates. | This is a different measure from the volume entering the site. |
| Discharge | Water leaving for treatment or another permitted destination. | Temperature, dissolved solids, treatment chemicals and timing affect what the receiving system must manage. |
| Indirect water | Water associated with electricity generation and other upstream activities. | It may occur in another watershed and cannot be treated as demand on the host town’s water pipe. |
Berkeley Lab estimated U.S. data centers consumed about 66 billion liters directly in 2023, approximately 17.4 billion U.S. gallons. Its estimate of electricity-related indirect consumption was nearly 800 billion liters, about 211 billion gallons. Both are historical estimates for all U.S. data centers, not current AI-only meter readings.
The indirect estimate uses regional grid mixes and does not incorporate facility-specific power-purchase agreements or behind-the-meter generation. Those boundaries matter. A contract is not automatically a physical change in power-plant water consumption; a grid-average estimate is also not a facility meter reading.
Which river, reservoir, aquifer or reclaimed-water system supplies it? Where does wastewater return?
Is this an annual average, a maximum day, a maximum hour, or a future permitted amount?
Is it drinking-quality water, treated wastewater or another source? What treatment is needed?
Does the figure include only the site, electricity generation, manufacturing, or a full life cycle?
CRS’s July 2026 FAQ explains that most operators obtain water from public systems and describes state and local water allocation and federal discharge requirements. It also makes clear why water quantity and water quality must be assessed separately.
Google’s 2026 environmental report separates withdrawal, discharge and consumption for 2025. The selected locations below show why the columns should remain separate. These are company-reported annual quantities for data-center locations, not isolated AI workloads, peak delivery requirements or a representative national sample.
| Google location / 2025 | Withdrawal, million gal | Discharge, million gal | Consumption, million gal |
|---|---|---|---|
| Ashburn, Virginia | 76.5 | 23.3 | 53.2 |
| The Dalles, Oregon | 654.3 | 185.3 | 469.0 |
| Berkeley County, South Carolina | 988.5 | 86.8 | 901.7 |
The arithmetic at The Dalles is 654.3 − 185.3 = 469.0 million gallons. Dividing annual withdrawal by 365 would yield about 1.79 million gallons per day on average; that calculated average is not the maximum-day demand. Facility size, cooling, source and operations differ, so these totals alone do not rank efficiency.
Consumed water is not destroyed: evaporation moves it into the atmosphere, where it remains in the wider water cycle. The local concern is that it is no longer available at the same source, place and time. In Google’s table, withdrawals are potable unless a row states otherwise; the three selected rows do not identify an alternative source.
Annual accounting is the starting point. The local system also has to meet its busiest day.
Continue: The hottest day is a different question. →Annual totals do not reveal a community’s available water capacity.
A town can have adequate annual supply and still face a shortfall in treatment, storage or pipe capacity during a hot, dry week. Existing homes and businesses, agriculture, environmental flows and future growth may all depend on the same system. The relevant comparison is not simply the project’s gallons divided by all water in the state.
A March 2026 research preprint by Han and colleagues focuses on this capacity problem. Its projected national need for additional water-system capacity changes substantially with growth and water-efficiency assumptions. Its central practical contribution here is the distinction between annual consumption and peak withdrawal.
Which restrictions apply to this customer? Are industrial commitments interruptible, and who has priority during a shortage?
What happens if several approved projects reach full operation together, alongside planned housing and industry?
Can the treatment plant supply the needed quality and peak flow? What existing downstream uses depend on that water?
Who funds new pipes, treatment and storage, and who pays if the expected customer never arrives?
A replenishment project can have value, but its place, season, additionality and durability need scrutiny. A watershed restoration project elsewhere does not by itself create the extra local flow needed at 4 p.m. during a heat wave. Google’s environmental report is an example of corporate water accounting that should be read alongside location-specific data.
The Scenario Lab includes a simple peak-capacity balance. It is an arithmetic illustration; it does not model hydrology, water rights, emergency reserves or the reliability of a real supply.
Water supply is one part of the environmental picture; air, materials and discharges have their own records.
Continue: Beyond the water meter. →Air, carbon, materials and waste follow different pathways.
Electricity-related emissions depend on the power system and time of use. Backup engines and on-site generation can create local air emissions. Chip fabrication, concrete, steel and equipment manufacturing add other impacts. Counting one boundary does not describe the entire life cycle.
| Impact | Useful evidence |
|---|---|
| Local air quality | Engine type and count, fuel, hours of use, permit limits, controls and cumulative dispersion modeling where relevant. |
| Electricity emissions | Location-based grid emissions and market-based procurement accounting, with each method identified. |
| Construction and materials | Embodied emissions, land clearing, concrete and steel, equipment production and transport. |
| Water quality and drainage | Wastewater acceptance conditions, blowdown management, stormwater plans and construction controls. |
| Equipment turnover | Reuse, resale, secure data destruction, repair and documented end-of-life handling. |
EPA explains that state and local agencies issue most air permits for data centers under approved Clean Air Act programs. Emergency use, routine testing and regular generation have distinct regulatory implications. A permitted worst-case amount is not the same as actual annual emissions.
A long-term power contract can support new supply. Annual renewable matching does not establish carbon-free electricity in every hour or prove that a site can operate off-grid. IEA’s supply analysis explicitly separates the physical generation mix from corporate contractual procurement.
Waste-heat recovery is another site-specific option. Its value depends on useful temperatures, a nearby heat customer, matching demand, pipes and economics. A cooling system can reject substantial heat while having no practical customer for it.
For 2024, IEA’s 2025 assessment estimated the global physical data-center electricity mix at roughly 30% coal, 27% renewables, 26% natural gas and 15% nuclear, with rounding and other sources accounting for the remainder. The shares vary by region. These historical estimates are not the contractual energy claims of individual companies.
The same assessment projected renewables would meet nearly half of the additional electricity demand through 2030, with gas and coal together supplying more than 40% of the increase. This is the published 2025 scenario, not an inventory of new plants that are certain to be built. A low-carbon supply agreement can coexist with a grid that still needs fossil generation in some hours.
Continue: A large investment is not a local balance sheet. →Separate private spending, tax collections, incentives and public obligations.
A campus announcement may include buildings, servers, utility works, financing and years of planned expansion. Much of the equipment may be manufactured elsewhere. That investment can generate local business and tax revenue, but the headline amount does not tell a town what remains after exemptions, public costs and contractual commitments.
| Account | Keep it separate |
|---|---|
| Private capital spending | Buildings, machines and other assets purchased by companies; timing and geographic scope matter. |
| Local revenue | Real-estate tax, equipment tax, fees and negotiated payments actually retained locally. |
| State incentives | Sales-tax exemptions, credits or other forgone revenue; these may affect a different government budget. |
| Public spending and risk | Roads, utilities, emergency services, debt, administration and potential stranded assets. |
| Community benefits | Grants or negotiated commitments; distinguish binding terms from voluntary programs. |
JLARC compared a facility with $150 million in taxable computer equipment under different Virginia local tax rates and depreciation schedules. Five-year equipment-tax collections ranged from about $0.4 million to $10.8 million. Real-estate taxes were excluded from that comparison. This is a controlled fiscal illustration using the report’s tax rules, not a forecast for an arbitrary new campus.
Local officials in the audit described uses of the revenue: lower real-estate tax rates in Loudoun and Prince William, an affordable-housing trust fund in Henrico, reserve funds, and new schools in Mecklenburg. These are specific examples of public benefits; their size and availability depend on the local tax base and budget choices.
Virginia’s 2024 audit found substantial variation in local receipts and in the treatment of computer equipment. Its examples show why identical equipment investment can produce very different revenue under different tax rates and depreciation schedules. That is a lesson about fiscal structure, not a universal estimate of profit per campus.
Equipment value changes over time. Replacements can renew the base, while abatements, depreciation or slower investment can reduce it. A community relying heavily on a few large taxpayers may want reserves and a downside forecast. The project’s promised growth is not the same as a diversified local economy.
The useful comparison is with realistic alternatives for that land and infrastructure, including leaving them available for future uses. A national total of “economic impact” cannot replace this local accounting.
Public revenue and private investment are different from employment. The next chapter follows the workers.
Continue: The jobs arrive in different waves. →Construction employment, permanent staffing and wider economic effects are not interchangeable.
Construction can employ many trades: electricians, pipefitters, equipment operators, concrete crews, engineers and commissioning teams. A multi-phase campus can sustain activity for years. Once operating, the staffing mix changes to facility technicians, network and hardware workers, security, management and contracted services.
Count workers over time, local hiring and job-years.
Count ongoing positions, contract roles and local wages.
Virginia’s audit reports a representative 250,000-square-foot facility with approximately 50 full-time workers, roughly half contracted, and around 1,500 workers at peak construction. These are contextual examples from interviews, not staffing rules for every AI campus.
| Ask for | Because |
|---|---|
| Peak headcount and job-years | 1,500 people at the busiest point is different from 1,500 jobs maintained for ten years. |
| Local residence and hiring | A worker lodging nearby during construction is not necessarily a permanent addition to the local labor force. |
| Wages, benefits and contractor roles | An average across engineers, guards and contractors can hide substantial differences. |
| Training and access | Apprenticeships, community-college programs and clear qualifications determine who can realistically benefit. |
| Direct versus supported employment | Supplier and induced jobs are model estimates across an economy, not all people employed on the campus. |
A project can have meaningful value with limited permanent staffing if its revenue, training and supplier opportunities suit local priorities. It can also disappoint expectations if temporary work is marketed as long-term employment. The presentation supplies the distinctions rather than ranking which outcome a community should prefer.
Employment describes one relationship with a campus. Living next to it creates another.
Continue: Life on the other side of the fence. →Daily experience is measured in sound, light, traffic, views and trust.
Fans, chillers, pumps and generator testing can have different noise patterns. Nighttime sound, tones and low-frequency components deserve attention alongside average levels.
Truck routes, dust, road wear, work hours and temporary housing demand may matter for years during phased development.
Building height, screens, vegetation, setbacks and new power lines affect views and nearby land uses.
Fire access, high-voltage systems, fuels and battery equipment require a site-specific response plan and coordination.
Virginia’s audit distinguishes noise that violates existing ordinances from persistent sound that some neighbors still find disruptive. Compliance and lived experience do not always align. The evidence does not support treating every facility as equally noisy, or assuming that one neighbor’s experience predicts every site.
An understandable condition identifies the metric, measurement location, operating conditions, reporting schedule, responsible authority and remedy. A promise to be a “good neighbor” leaves those details open. A plan should address the combined sound of all campus phases and relevant nearby facilities, including simultaneous testing where allowed.
For residents, useful records include preconstruction sound measurements, predicted worst-case operating sound, lighting plans, public complaint channels and the process for verifying a complaint. Health concerns should be heard without turning anecdotes into a diagnosis or assuming a medical outcome from a facility label.
Public trust also depends on what can be learned before approval. Commercial confidentiality may protect specific business details, but communities can still seek meaningful disclosure about resource demand, public obligations and enforceable operating conditions.
Clear operating conditions matter most when someone has authority to enforce them.
Continue: One project. Several decision rooms. →A zoning hearing, a utility tariff and an environmental permit answer different questions.
| Decision track | Typical questions |
|---|---|
| Land-use planning | Is this use allowed? What setbacks, height, landscaping, noise and phase conditions apply? |
| Electric service and system planning | What load can connect, when, under what studies, service conditions and upgrade requirements? |
| Utility rates and contracts | Who pays, for how long, with what guarantees and exit provisions? |
| Water availability and supply agreement | What source, peak delivery, drought priority, infrastructure and payment terms are available? |
| Environmental permits | What air emissions, withdrawals, discharges, stormwater or wetland impacts require approval? |
| Fiscal and development agreements | Which incentives and obligations are binding, transferable and enforceable? |
Authorities vary across states and between municipal, cooperative and investor-owned utilities. Some towns control zoning but not the electric tariff. Some water arrangements depend on state law or regional suppliers. Neither “the federal government approved it” nor “the county approved it” describes the entire process.
Readers need enough detail to understand the proposal without needing operational security diagrams or private customer data. A public fact sheet can identify owner and operator, phases, maximum and expected power, normal and peak water, sources, discharge plans, fiscal obligations and complaint contacts.
NIST’s July 2026 initial public draft on AI data-center security examines threats and architectural differences from high-performance computing. Its draft status matters; a draft is not a final mandatory standard. Security does not settle whether a particular local development agreement is well structured.
The following examples show how those decisions differ across real places.
Continue: Four places. Four different questions. →Case studies illuminate mechanisms; they do not settle every community’s decision.
The legislative audit documents a large established industry, major local tax receipts in some jurisdictions and substantial projected power needs. The lesson is to examine fiscal success and infrastructure pressure together; maturity does not remove tradeoffs.
AEP’s tariff example addresses the risk of building for demand that does not arrive. The consumer advocate reports minimum capacity payments and a multi-year term, and notes an industry appeal. The transferable idea is financial commitment; the exact legal terms are local.
Texas SB 6 addresses large-load interconnection and reliability. Separately, the governor’s September 21, 2026 directive announced a permit halt tied to audits. These are dated policy developments, not proof that every proposed facility is harmful or that all permitting steps stopped.
Entergy’s March 2026 Meta agreement describes seven planned gas plants totaling more than 5,200 MW, transmission and batteries, plus projected customer savings. These are utility projections and proposed investments, not completed assets or verified future savings.
For each case, keep the same columns: project stage; capacity boundary; water source and peak demand; public investment; revenue retained; ongoing jobs; operating conditions; and the entity that carries downside risk. A proposed campus and a mature county-wide industry should not occupy the same row without clear labels.
A nearby example may be informative but still differ in weather, generation mix, water rights, tax law or utility ownership. That is why a local record is more useful than assuming a national average applies.
Continue: A global buildout, with local constraints. →Computing crosses borders. Electricity, water and permits stay rooted in places.
The IEA’s 2025 regional analysis places the United States and China at the center of projected demand growth, with additional growth in Europe and other markets. National totals conceal regional clusters and different power systems. The updated 2026 global outlook should be used for the current aggregate trajectory.
Large cloud markets and AI investment meet a patchwork of utilities, state commissions, local land-use rules and water systems.
A major computing market with a different generation mix and infrastructure-planning system. U.S. tax and tariff examples should not be transferred directly.
Power availability, grid access, carbon accounting and potential heat-reuse customers can influence site choices.
New capacity may be driven by domestic services, data residency, connectivity and access to energy. Each region requires its own evidence.
Governments and businesses can value domestic computing for research, industry, defense, resilience and control over sensitive data. Those possible national benefits do not tell a host town how much water a particular facility will use or what the local tax agreement delivers.
There are also dependencies outside the fence: semiconductor manufacturing, advanced packaging, memory, power equipment, fuel and financing. A building can be domestic while essential components come from a global supply chain. IEA’s 2026 report discusses the bottlenecks connecting computing growth to these physical systems.
These different settings also help explain why one forecast cannot describe every future project.
Continue: What changes the outlook? →Demand, technology, finance and infrastructure can move in different directions.
Forecasts are conditional stories built from assumptions. The number of accelerators shipped, their useful life, idle power, workloads, occupancy, cooling and electricity constraints all affect demand. Berkeley Lab’s update makes several of these sensitivities explicit instead of presenting one number as inevitable.
| If this changes… | Possible consequence, all else equal |
|---|---|
| More efficient software and chips | Less energy for the same task; lower costs can also encourage additional use. |
| More reasoning, video and agents | A different task mix can raise energy requirements even as simple text becomes more efficient. |
| Slower connection or equipment delivery | A proposed campus may enter service later or at a smaller first phase. |
| Tighter financing or weaker revenue | Projects can be delayed, reduced, sold or cancelled; contract protections become important. |
| More flexible operation | Some work may move in time or place, potentially reducing stress during constrained periods. |
| Different cooling / electricity supply | Direct water, indirect water and emissions can change in different directions. |
A useful project timeline distinguishes land control, approval, service agreement, financing, construction, first energization, occupied capacity and later expansions. Announced megawatts or investment are not equivalent to committed demand or operating capacity.
On-site gas, batteries, geothermal, nuclear and other supply arrangements each have development schedules and constraints. A memorandum or purchase agreement is evidence of intent, not proof that a technology will be online by the campus’s required date.
Continue: Questions behind the headline. →Replace a sweeping claim with the measurement that would make it meaningful.
There is no universal per-question quantity. Model, task, answer length, location, time and accounting boundary change the result. Google’s 2025 Gemini study reports a median text prompt of 0.26 mL of direct water; academic modeling includes other boundaries. Neither is a measurement of every chatbot interaction.
Ask which loop. A sealed chip loop can transfer heat to an evaporative outdoor system. A dry heat-rejection design can avoid most operational cooling consumption, while bathrooms, commissioning, electricity generation and manufacturing remain outside that narrow claim.
Specify capacity, average annual draw and the comparison city. A household-energy equivalent excludes shops, schools and industry. A rated connection is also not the same as actual consumption. Use the calculator to see the assumptions.
Both mechanisms exist: spreading fixed costs over more sales, or adding expensive infrastructure and tightening supply. Historical averages and a constrained regional market can point in different directions. The service agreement and local supply conditions are essential.
Annual contractual matching differs from hourly physical supply. Ask about the project’s location-based footprint, the contract, whether new generation is added and what supplies the site when contracted resources are unavailable.
Separate peak construction workers, job-years, ongoing employees, contractors and modeled supplier jobs. Ask for local hiring and wages rather than a single combined headline.
Replenishment can support water stewardship, but timing, watershed, additionality and delivery capacity matter. It does not by itself guarantee available municipal supply during a local peak.
A national comparison can show scale but cannot establish local capacity or priority. Use equivalent metrics, the same year and watershed, and distinguish withdrawal from consumption. Both existing uses and new demands belong in the local balance.
Efficiency reduces resource use for a given task. Overall demand also depends on adoption, task complexity and pricing. Forecasts should show several assumptions rather than choosing one direction as certain.
Local benefit depends on retained revenue, costs, jobs, supplier access and the alternatives available. Private spending, state tax incentives and local receipts are different accounts.
Change the assumptions. See the arithmetic. Keep the limits visible.
The energy tool is a steady annual accounting illustration. It does not predict a real load profile, marginal generation, emissions, shortages, retail bills or AI throughput. Entered average load is an electrical fraction, not GPU utilization. A constant PUE and WUE simplify real weather and operating variation.
The peak-water tool compares three entered daily volumes. A positive balance is not a water-availability approval: reserves, outages, water rights, sustainable yield, treatment quality, maximum-hour flow and other users’ commitments still matter. No calculator score here says to approve or reject a project.
The household baseline is the EIA’s 2022 average annual utility purchases of 10,791 kWh per residential customer; it excludes self-generated electricity, such as rooftop solar, and is labeled by year and editable. Water uses U.S. gallons (3.785411784 liters per gallon). No private data is sent or saved.
Continue: Read the proposal, not just the rendering. →A field guide for residents, officials and anyone trying to understand a campus.
Start with a short public project record. Record the date of each document and distinguish the first building from the entire planned campus. Mark an item unanswered rather than replacing it with a national average.
| Document | What it can reveal |
|---|---|
| Site plan and land-use application | Phases, layout, setbacks, access, outdoor equipment and proposed operating conditions. |
| Utility study, tariff and service agreement | Capacity, energization milestones, upgrades, reserved-demand payments and exit risk. |
| Water agreement and supply analysis | Source, annual and peak volumes, delivery capacity, drought priority and infrastructure costs. |
| Air permit and wastewater conditions | Equipment, emission limits, operating assumptions and discharge obligations. |
| Fiscal analysis and development agreement | Tax treatment, public costs, enforceable benefits and downside exposure. |
| Operations and monitoring plan | Public reporting, testing schedules, complaints, corrective actions and responsible contacts. |
This is an original reading guide built from the distinctions in the source record. It is not a legal checklist for every jurisdiction or a substitute for the actual permit and agreement language.
Continue: See the infrastructure. Hear the explanation. →Selected media and primary documents, with context attached.
These tours make the machinery tangible. Google shows an operator-selected commercial facility; Frontier shows scientific supercomputing. Read them alongside the resource records to connect what you see with what a particular campus needs.
Inspect the 2026 reference case and sensitivity assumptions.
Read the CRS explanation of sources, discharges and authority.
Explore Virginia’s economic and local-impact evidence.
See what capacity-market attribution measures and what it does not.
The words that make numbers comparable.
Search the source library. Inspect the boundary. Follow the original record.
This presentation was developed from a supplied 25-page explainer and website package, then reorganized and researched independently. It preserves useful questions while adding newer reports, clearer accounting boundaries, community decision tools and a richer chapter structure. It is intended for public education, not an endorsement or rejection of data-center development.
The research cutoff is October 21, 2026. Dates on sources and numbers remain visible. “Historical estimate” means an estimate of a past period; “scenario” means a conditional future result; “company disclosure” identifies information reported by an interested operator or supplier; “illustrative” means inputs selected to explain a calculation. A source record describes what was checked; it does not imply every page of every report was audited.
Primary research, government documents, market-monitor records and original company disclosures receive priority. Company plans remain plans. A working paper or initial public draft is labeled as such. Where a public mirror was needed, it is identified. All chapter citations open annotated source records; external documents and videos require internet access.
This is a broad U.S.-focused guide with international context, not an inventory of every facility or a site-specific engineering, environmental or fiscal study. Public data often lack comparable facility-level detail. The 2023 national water estimates remain labeled historical because a newer electricity forecast is not a new water measurement. The presentation does not claim a universal percentage of U.S. water consumption, a universal permanent-job count, or a universal resource cost per AI prompt.
The review package includes the presentation, source register, numeric data and methods, a change log against the supplied project, and integration notes. The HTML itself contains its images, fonts and scripts so the core presentation works offline. External links and videos need a connection. A final designed PDF will be produced after review.
Return to the chapter guide ↑