The longer story
Data centers did not begin with generative AI.
Today's facilities grew out of decades of changes in how organizations use computers, connect them, and share their capacity. These milestones explain that evolution; they are not a claim that one company or date invented the entire industry.
- 1964Centralized computing
Compatible mainframes made expansion easier.
IBM introduced System/360 on April 7: a family of compatible mainframes that let organizations move to larger machines while retaining their programs. It illustrates the era of centralized business computing, not the first computer room or data center. IBM's institutional history
- 1969–1983Connected networks
Computing became a networked service.
ARPANET's first computer-to-computer signal connected UCLA and SRI on October 29, 1969, beginning a four-node research network. Its January 1983 transition to TCP/IP helped separate networks communicate. Connecting computers changed what remote users could access. DARPA's ARPANET history
- 1989–1993The World Wide Web
The Web made networked information easier to use.
Tim Berners-Lee proposed the Web at CERN in March 1989; the first server and browser were running by the end of 1990. CERN made the software freely available on April 30, 1993. The internet is the underlying network; the Web is a system that runs on it. CERN's history of the Web
- 2006–2011Rented cloud services
Organizations could rent shared capacity.
Amazon's March 2006 S3 launch made storage available through web interfaces—one commercial-cloud milestone, not the invention of cloud computing. NIST's 2011 definition describes on-demand access to shared resources that can be rapidly provisioned and released. Virtualization lets physical computing resources be divided and shared; it does not remove the hardware or its costs. The contemporary S3 announcement; NIST's cloud definition
- 2010–2018A historical efficiency lesson
More computation did not require proportional energy growth.
A 2020 Science study estimated that global compute instances grew about 550% during 2010–2018, while data center energy use rose from roughly 194 to 205 TWh, about 6%. More efficient hardware and operations, virtualization, and consolidation helped. These are historical modeled estimates, not a meter-by-meter census or a promise that efficiency will always keep pace with demand. Masanet and colleagues' original study
- The AI buildoutNew workloads, old infrastructure needs
AI adds another layer to the system.
Training builds a model; inference runs it for users. They can have different hardware and operating profiles. Conventional and AI-specialized servers coexist, and AI is not confined to one chip type. The question is how this added demand is supplied and governed alongside existing services. LBNL's 2024 server and workload analysis
Efficiency and shared infrastructure can change the footprint of computing substantially. Growing demand can also outpace those gains. Neither outcome is automatic: design, workload choices, power supply, and public rules matter.
Understand the system
One industry, many services and operating models.
Data centers support websites, email, media services, and organizational software as well as AI. Some organizations operate their own facilities. Colocation operators provide space and supporting infrastructure to tenants. Cloud providers sell access to computing, storage, and software. A cloud service is a delivery model; a data center is part of its physical infrastructure. JLARC's facility overview; NIST's service definition
The global picture
The IEA's April 2026 assessment estimates 485 TWh for all data centers in 2025 and projects about 950 TWh in 2030. The future figure depends on how demand and efficiency develop. IEA assessment
A separate U.S. outlook
LBNL's June 2026 update models a 2030 share of 9.5–15.3% of U.S. electricity, with an 11.8% reference case. Hardware shipments, utilization, cooling, and equipment life affect the result. Different geographic scopes cannot be averaged together. LBNL model
A national or global percentage cannot settle a local siting decision. A project's size, timing, connection point, and available resources determine the questions its host community must answer.
Physical limits
Annual energy and peak power answer different questions.
Two units to keep separate
- Energy • MWh or TWh
- Electricity used over a period, such as a year.
- Power • MW
- The rate of electricity use at a moment. Peak demand matters for equipment and grid capacity.
A facility needs electricity that can actually reach its site. Generation, transmission, substations, and connection schedules all matter. A clean-energy purchase or an announced power project does not by itself establish that capacity is available when the facility opens.
Annual renewable matching is an accounting approach: purchases cover an equivalent quantity over the year. It is different from clean electricity supplied in the same hours and relevant grid area as consumption. ACC recommends asking for both the accounting method and the actual supply plan, including added generation, storage, transmission, and the remaining fossil-powered hours.
Better chips and cooling can reduce energy per task. Total use can still grow when the volume or complexity of tasks rises faster than efficiency improves. The IEA treats efficiency, adoption, and task mix as separate demand drivers. IEA's demand assessment
Compare absolute electricity use and peak demand alongside efficiency ratios. Check reliability and emissions separately: shifting consumption to a cheaper hour does not necessarily shift it to a cleaner one.
The full physical footprint
Water and equipment need their own accounting.
Withdrawal is water taken from a source. Consumption is the portion not returned, such as water evaporated during cooling. A facility's direct water use and the indirect water associated with its electricity are different quantities. Their significance depends on the cooling design, power system, watershed, season, and competing needs. A 2026 PLOS Water essay examines these governance risks; it does not establish that every facility causes water shortages. Shah's water-security analysis
A closed loop at the chips still has to reject heat elsewhere. That final step may use evaporative cooling or dry heat rejection. “Liquid cooled” therefore does not mean “water-free.” Design and climate determine the tradeoffs. DOE/FEMP cooling guidance
Chips, servers, buildings, and replacement equipment also have manufacturing and end-of-life impacts. Reporting only operating electricity leaves those out. ITU-T L.1801 provides an AI environmental-assessment method; it is a technical recommendation, not a law that automatically binds every operator. ITU's February 2026 recommendation
Facility cooling water, electricity-related water, and full-lifecycle water are not interchangeable. Nor are operational emissions and embodied emissions. A lower footprint per task can coexist with a rising total footprint.
The substantive debate
Take the benefits and the concerns seriously.
Supporters can reasonably point to useful services, infrastructure investment, skilled work, and local revenue. Critics can reasonably ask about constrained resources, public costs, noise, siting, and whether promised benefits are proportionate. Neither side's strongest argument requires treating every facility as identical.
Useful services
The opportunity: Reliable digital services and specific tools that improve a real task.
The concern: A useful application does not establish that every workload is useful or that its benefits cancel out all impacts.
The test: Compare measured performance with a simpler alternative, including reliability, errors, and resources used.
Jobs and local revenue
The opportunity: Construction, skilled operating roles, and property-tax receipts.
The concern: Construction jobs are temporary; operating headcount is different. Incentives and public costs affect the fiscal balance.
The test: Separate construction from permanent jobs and gross tax receipts from incentives and public spending.
Power and water
The opportunity: Project-funded upgrades and well-designed resource use.
The concern: Demand can exceed deliverable capacity, or create costs and drought pressures others must bear.
The test: Require a site-specific capacity study, drought plan, and clear payer commitments.
Neighborhood fit
The opportunity: A suitable industrial site with effective mitigation and ongoing oversight.
The concern: Noise, generator activity, land use, and cumulative development can affect nearby residents.
The test: Use enforceable operating limits, public monitoring, and a way to resolve complaints.
Virginia's 2024 legislative review found that economic benefits arose mainly during construction, while facilities also contributed local tax revenue and raised infrastructure and siting questions. Its findings are specific to that region, not a universal economic template. JLARC's findings and recommendations
The wider AI debate also matters.
An efficient data center does not establish that an AI application is accurate, safe, secure, fair, or respectful of privacy. NIST's voluntary AI Risk Management Framework treats these as distinct parts of trustworthiness. They need evaluation in the context of the actual use, alongside environmental impacts. NIST's 2023 framework
ACC recommends making room for useful applications while checking their errors, who benefits, who carries the risks, and what alternatives exist. Neither an impressive demonstration nor a serious failure settles the value of every other AI application. The useful question is what a particular tool improves, under what conditions, and with what safeguards.
Verifiable claims deserve more weight than promises, whether they support or challenge a project. Missing evidence is a reason to investigate or set conditions, not proof of harmlessness or inevitable disaster.
Public participation • A practical U.S. guide
Protest, speak up, and stay involved.
Opposing a proposal is a legitimate position. Residents can ask for evidence, challenge unacceptable impacts, and seek a delay or denial as well as changes to a plan. The steps below help connect those concerns to decisions and responsibilities that can be identified.
Find the decision and the public record.
Start with the city or county planning office or clerk. Ask for the project name, application number, site plans, studies, staff report, decision-maker, and remaining approvals. Ask which decisions are still open to public input and when comments are due. Look for official agendas and notices, and request updates. Not every proposal requires the same hearing or approval route.
A land-use decision, an environmental permit, and a utility-rate case can involve different authorities. Ask where each concern belongs rather than assuming one board controls them all. EPA: identifying the decision process and points for input.
Examples of distinct channels: Loudoun County's data-center zoning process and Virginia's utility-case participation process. These are local examples, not nationwide instructions.
Meet officials. Speak at meetings. Submit comments.
Contact elected representatives through their published office channels and ask for a meeting or written response. Confirm the hearing agenda, speaker registration, time limit, and submission deadline. Ask early about interpretation, accessibility, or remote participation. A developer's open house may provide information without being the official hearing.
Prepare a short statement: identify the proposal, describe a specific concern, cite a record or direct observation, and ask for a clear decision or condition. Listen to the response and note unanswered questions. If you cannot attend, check whether written comments are accepted. Confirm that your submission reaches the official record; an email to an official or a social post may not do that automatically.
Save your submission and receipt. Check what will become public before including an address, medical details, or other personal information. Read EPA's guidance on specific, evidence-supported comments and the purposes and limits of public meetings.
Peaceful protest & community organizing
Make the concern visible and the request clear.
Peaceful assembly and petitioning government are protected by the First Amendment. The rules depend on the location and activity. Check official requirements for the proposed event, including any permits for a march, amplified sound, or use of a park. Review the ACLU's protest-rights guidance; contact local civil-liberties or legal help for a specific situation.
- Choose a clear, evidence-backed request: release a missing study, protect a water supply, change a plan, or deny an unsuitable proposal.
- Coordinate a peaceful, accessible gathering. Plan transport, water, a buddy system, and a way to leave safely. Keep emergency routes and entrances clear; do not enter private property without permission.
- Use verifiable statements on signs and leaflets. Respect other people's right to disagree, avoid harassment, and seek consent before sharing identifiable participant photos or contact details.
- Connect the event to the formal process. Petitions, rallies, and media coverage do not automatically become a submitted comment or decide a permit.
These are practical suggestions and general U.S. information, not legal advice or a guarantee of a particular outcome.
Ask for changes that can be checked.
Specific requests make it clearer what would need to change—or why a proposal remains unacceptable. These are options to evaluate, not claims that every authority can impose them or that every site can be made suitable.
Power and public costs
Ask for the peak load, a deliverable supply and connection schedule, and clear responsibility for upgrades or cancellation costs. An annual renewable-energy claim does not establish power availability at the site.
Water, noise, and air
Request site-specific water and drought plans, noise studies, and backup-generator emissions information. Ask for measurable limits, monitoring, complaint procedures, and consequences for missed conditions.
Scale, location, and timing
Ask whether a smaller project, different site, or phased opening addresses the problem. Seek delay or denial where essential capacity or protections are missing. Potential future benefits do not establish that a current proposal is acceptable.
Use the project decision checklist to connect each request to evidence, a responsible party, and an enforceable condition. The engineering and operating examples explain useful options and their limits.
Follow the decision beyond the meeting.
Check the published decision, conditions, minutes, and any responses. Ask what changed, what remains unresolved, and who will check compliance. If a project opens, compare public monitoring with its commitments and use the published complaint channels. If considering a formal appeal, promptly confirm the applicable process and deadlines with the responsible authority or qualified local counsel.
Hearing procedures, permit requirements, appeal rights, and access to records vary by jurisdiction and project. Always check the current official notice. This guide does not identify an active local case or promise that participation will change its outcome.
Disclosure and accountability
Compare evidence at the same level.
A model's footprint, a facility's annual consumption, and a company's global sustainability report answer different questions. Ask for the period, location, workload, measurement method, and lifecycle boundary. Label provider-reported measurements, independent checks, models, and future commitments separately.
- Absolute energy use and peak demand, alongside efficiency and utilization.
- Water withdrawal and consumption, source, seasonal limits, and cooling design.
- Operating and embodied emissions, with electricity-accounting assumptions.
- Construction and permanent jobs, incentives, public costs, and payer contracts.
- Performance against commitments, accessible reporting, and remedies for missed conditions.
A European reporting example
EU Regulation 2024/1364 covers installed IT power demand of at least 500 kW and requires annual energy and water indicators. Public data are aggregated at country and EU level; individual records remain confidential. This reporting rule does not itself impose a minimum efficiency standard. The regulation
An example of allocating risk
AEP Ohio's data-center tariff, effective July 2025, uses minimum-demand charges, financial security, contracts, and cancellation provisions to address stranded-cost risk. This is one utility's mechanism, not a national rule or a guarantee that every cost is eliminated. Current implementation details
From promises to practice
Useful applications and better infrastructure need separate tests.
Established engineering: start with avoidable waste.
DOE/FEMP guidance supports airflow management, matching IT capacity to demand, suitable temperature settings, climate-appropriate economizers, and direct liquid cooling. Heat reuse needs a nearby user, useful heat temperature, and backup heat rejection. These are practical engineering options; savings depend on the site. PUE measures supporting-infrastructure efficiency, not the usefulness of the computation or its total footprint. DOE/FEMP design guide
Flexible demand: useful options, with operating limits.
Workload scheduling, processor controls, facility adjustments, and storage can help reduce peak grid draw. Delay-tolerant jobs offer scheduling room; interactive services must respect response-time and reliability requirements. Storage shifts when electricity is drawn and has its own losses. These tools do not eliminate the need for bulk electricity, and new demonstrations should not be presented as sector-wide results. Generator substitution also needs separate emissions and permit review. LBNL's technical synthesis
Infrastructure + measurement
Reducing AI’s environmental impact
Two groups approach AI’s environmental demands from different angles: improving how data centers run, and developing tools to estimate computing impacts. Each needs to be assessed within its scope.
Elemental Impact
Elemental Impact is a nonprofit investor leading the Data Center Innovation Initiative. Its cooling priorities include HVAC and retrofits intended to reduce energy and water use while maintaining reliability. The initiative supports testing and validation of emerging technologies in data-center settings, including AI and other computing uses.
ACC has not independently verified completed DCII energy or water savings. Cooling choices also need to account for local water conditions and energy tradeoffs.
Explore Elemental’s initiative →CodeCarbon & EcoLogits
Built by the French nonprofit Code Carbon, these open-source tools help estimate computing impacts. CodeCarbon combines accessible hardware energy readings or estimates with electricity carbon-intensity data. EcoLogits models supported remote generative-AI API requests, including energy and water where supported.
EcoLogits does not directly meter provider servers. Its scope excludes training, networking and end-user devices; incomplete provider disclosures require assumptions. Its estimates cannot be treated as measured totals for every AI tool.
Elemental and CodeCarbon/EcoLogits information reviewed October 15, 2026.
Measurement can help identify opportunities; reported savings still need an explicit baseline and evidence. Supporting an organization does not demonstrate that our AI electricity, water or emissions have been offset.
Explore these organizations on ACC’s charities page →
Beginning October 15, 2026, we commit 1% of eligible earned revenue across our businesses and projects to work addressing AI’s environmental impacts. We’ve asked Elemental whether contributions can be directed specifically to its Facility Cooling & Water Efficiency work. The recipient remains pending while we await its reply. ACC donations and grants retain their designated purposes.
Read the commitment and calculation →State what changed, what stayed constant, which impacts were counted, and whether the result came from operation, a pilot, or a model. A climate-benefit claim needs evidence of the benefit and the added computing footprint; one cannot simply be assumed to offset the other.
ACC's recommendation • A project-by-project test
Set conditions that can be checked and enforced.
This checklist is ACC's practical recommendation, not a legal standard or a settled scientific verdict on any individual proposal.
Define the service and the alternatives.
What useful work will the project support? Can a smaller model, existing facility, or less resource-intensive method meet the same need?
Prove the capacity is deliverable.
Check the load, connection schedule, actual power supply, cooling design, water source, seasonal limits, and cumulative nearby demand.
Make the costs and benefits explicit.
Identify who pays for upgrades, delays, or cancellation. Separate temporary jobs, permanent roles, incentives, and net public revenue.
Attach commitments to accountability.
Specify resource limits, noise and generator conditions, reporting, independent verification where appropriate, and consequences for noncompliance.
Review performance after opening.
Compare actual use and benefits with the proposal. Correct failures and revisit expansion approvals rather than treating permission as the end of oversight.
When capacity, benefits, cost allocation, and enforceable protections are credible.
When a smaller scale, different location, or staged opening can resolve specific problems.
When essential capacity or protections are absent, or serious impacts cannot be acceptably addressed.
A workable direction
Support useful computing, reduce avoidable harm, and insist on evidence.
Field File 003 • Personal AI use
One person's text prompts can be small. “Irrelevant” is still too strong.
For a person who mainly drafts, summarizes, brainstorms, or asks questions in text, direct electricity use can be modest. That does not make every AI task small, and it does not make the combined infrastructure footprint disappear.
The proposed claim needed one important correction: the amount of data someone moves through the internet is not a reliable measure of environmental impact. Bytes describe traffic. Watt-hours describe electricity. Emissions and water depend on the power supply, location, cooling, hardware, provider, task, output length, and measurement boundary.
ACC evidence verdict
The idea is partly right, but the strongest wording outruns the evidence.
- Supported with caveats
- Routine text-based AI can use little electricity per prompt, and efficiency per task is improving quickly.
- Misleading as stated
- “Infinitesimal and irrelevant compared with everyday data use” compares unlike units and treats every AI task as if it had the same footprint.
- Not knowable from a username or bill
- A person's exact footprint cannot be calculated honestly without the provider, model, task mix, output length, location, and lifecycle boundary.
A transparent text-only scenario
This illustration applies Google's company-reported full-stack median of 0.24 watt-hours for one Gemini Apps text prompt in May 2025. The television comparisons use the same study's 100-watt reference. The arithmetic is transparent; the result is not a universal AI average.
Less electricity than about nine seconds of a 100-watt television in this Google measurement.
About three minutes of a 100-watt television. This is an ACC calculation from Google's per-prompt result.
About 14 minutes of a 100-watt television. It still describes text prompts only.
It does not describe every provider, ChatGPT or another named service, model training, user-device electricity, storage, image generation, video generation, deep reasoning, agents, or a full business workflow. Google's authors also show why published estimates can diverge when studies include different parts of the serving system.
Review the Google measurement →Bytes measure data volume. Watt-hours measure electricity. Grams of CO2-equivalent measure climate emissions. Liters or milliliters measure water. A responsible comparison keeps the task, provider, period, location, and lifecycle boundary visible.
Task type can change the answer
Simple text
Short drafting, summarizing, classification, and question-answering can be relatively low-energy when an efficient model fits the task.
Reasoning and agents
The IEA reports that some reasoning and agentic tasks can use hundreds or thousands of times more energy per query than simple text.
Images and video
The IEA highlights video generation among high-energy uses. Task complexity, output length, and the serving system matter more than prompt count alone.
One hundred short text prompts and one hundred long reasoning, agentic, image, or video jobs are not equivalent workloads. The International Energy Agency identifies efficiency gains, adoption, and the changing mix of tasks as the three central drivers of AI electricity demand.
Review the IEA findings →Why published numbers differ
Two disclosures can be useful without being directly comparable.
Google reported a median Gemini Apps text prompt using 0.24 Wh, 0.03 gCO2e, and 0.26 mL of water in May 2025. Mistral reported 1.14 gCO2e and 45 mL of water for a 400-token Le Chat response in its lifecycle study. The products, periods, power systems, and accounting boundaries differ, so this is not a provider ranking.
Review the Mistral lifecycle study →- GoogleProduction measurement for a provider-specific median text prompt; includes serving equipment, provisioned idle capacity, data center overhead, and reported emissions and cooling water.
- MistralProvider-specific 400-token response; includes upstream impacts such as server manufacturing and excludes the user's terminal.
The practical answer is proportion, not panic
- Use ordinary software when it already solves the problem.
- Default to text and a fit-for-purpose model for routine work.
- Ask for only the length and detail the job needs.
- Reuse strong outputs instead of regenerating them.
- Reserve long reasoning, agents, images, and video for work that benefits from them.
- Publish comparable energy, water, emissions, and lifecycle data.
- Make efficient modes and smaller fit-for-purpose models easy to choose.
- Use cleaner electricity and water-smart cooling without hiding tradeoffs.
- Prevent infrastructure costs and local impacts from being shifted onto communities.
- Report total growth alongside efficiency per task.
ACC conclusion
Use AI intentionally, not anxiously. Keep personal use in perspective and keep asking providers for the missing evidence.
Sources and review notes
Follow the evidence, including its limits.
Reviewed October 8, 2026. Publication dates and measurement periods remain visible because older evidence does not become a current measurement when a page is updated. Institutional histories explain milestones; modeled outlooks describe possible futures; provider disclosures retain their scope limits. ACC's recommendations are identified separately.
History and computing models
- IBM: System/360Institutional history; April 7, 1964 launch. An illustrative mainframe milestone.
- DARPA: ARPANETJuly 2020 history of the 1969 network and 1983 TCP/IP transition.
- CERN: A short history of the Web1989–1993 milestones; the Web and internet are distinct.
- AWS: Amazon S3 launch announcementMarch 14, 2006 contemporary company announcement.
- NIST SP 800-145: The definition of cloud computingSeptember 2011; service, deployment, and resource-pooling definitions.
- Masanet et al.: Recalibrating global data center energy-use estimatesScience, February 28, 2020; modeled estimates for 2010–2018.
Demand, resources, and engineering
- IEA: Key questions on energy and AIApril 16, 2026; global 2025 estimates and future outlook.
- LBNL: United States data center energy usage—2025 updateJune 2026; U.S. model with a 2030 uncertainty range.
- LBNL: 2024 United States data center energy usage reportDecember 2024; historical workload and server classifications, not the latest forecast.
- Shah: Data centers and local water securityPLOS Water, January 13, 2026; contextual governance essay.
- DOE/FEMP: Best practices for energy-efficient data center designJuly 2024; engineering guidance with design and location constraints.
- LBNL: Integrating AI data centers with the power gridMay 2026; technical synthesis, not a measured national savings estimate.
- Code Carbon: organization and toolsOfficial organization information; reviewed October 15, 2026.
- CodeCarbon methodologyHardware energy readings, estimation and emissions accounting; reviewed October 15, 2026.
- EcoLogits methodology and scopeModeled inference impacts and exclusions; reviewed October 15, 2026.
- Elemental Impact: Data Center Innovation InitiativeProgram description reviewed October 15, 2026; self-reported program scope and technology priorities, including cooling and water efficiency. ACC has not independently verified completed DCII project results or savings.
Communities, accountability, and an operating application
- JLARC: Data centers in VirginiaDecember 2024 archived legislative review; findings are region-specific.
- AEP Ohio: Data center tariffCurrent implementation page; tariff effective July 23, 2025. Territory-specific.
- EU Regulation 2024/1364Common reporting indicators, covered-facility threshold, and aggregate public-access limits.
- ITU-T L.1801Approved February 6, 2026; a technical environmental-assessment recommendation.
- ECMWF: IFS and AIFS v2 go liveMay 12, 2026 operational announcement; specific forecasting comparisons.
- NIST AI RMF 1.0: Artificial intelligence risk management frameworkJanuary 2023; voluntary guidance on contextual risks and trustworthiness. NIST's framework is undergoing revision.
Public participation and peaceful protest
- EPA: Public participation guide—process planningUpdated September 11, 2026; participation opportunities and decision-process guidance, not uniform local law.
- EPA: Public meetingsUpdated March 25, 2026; meeting purposes, inclusion, records, and limitations.
- EPA: Tips for providing effective commentsUpdated August 26, 2026; evidence, scope, and reasoning in EPA comments. The comment process is not a vote.
- Loudoun County: Data center standards and locationsOfficial zoning-project records and participation links; application-specific routes and grandfathering prevent universal assumptions about hearings.
- Virginia SCC: Participating in casesOfficial case numbers, deadlines, and comment channels; Virginia-specific, distinct from county zoning.
- National Archives: Bill of Rights transcriptionFirst Amendment; ratified December 15, 1791. Source page reviewed August 7, 2025. Not a complete guide to protest law.
- ACLU: Protesters' rightsCivil-liberties organization's general U.S. guidance, updated June 8, 2026; check local requirements and seek legal advice for specific circumstances.
Personal-use measurements
- Google: Measuring the environmental impact of delivering AI at Google scaleProvider-authored study; May 2025 median Gemini Apps text prompt, with explicit serving-system boundaries.
- Mistral AI: Model lifecycle analysisJuly 22, 2025 provider disclosure; a 400-token Le Chat response, not a directly comparable product ranking.
A short meeting statement or email to adapt
Replace the brackets with accurate details and use the official submission route. This is an example, not a legal filing; nothing is sent from this page.