Permission — complete text edition

The previous volumes describe physical possibility: materials, patterned silicon, packages, racks, networks, buildings, cooling, and power. Permission asks a different question—under which financial, legal, political, and social conditions can those systems be built and operated? Each institution sees a different object: an investment, a load request, a land use, an emissions source, a tax base, a security concern, or a neighbor.

The IEA estimates that global data-center electricity consumption rose seventeen percent in 2025, with AI-focused facilities growing faster. Rapid growth compresses decisions that normally unfold on different clocks, forcing model developers, utilities, equipment suppliers, regulators, builders, and communities to confront uncertainty at the same time.Claim

The IEA estimates global data-center investment reached roughly half a trillion dollars in 2024. That aggregate is not a count of completed AI campuses, but it conveys the capital intensity of the wider buildout. Investors still need project-level evidence: customer demand, delivery rights, utility arrangements, construction budgets, operating costs, technology assumptions, and risk allocation.Claim

Public authorities apply other tests. Planning bodies examine land use; environmental programs examine air, water, habitat, and cleanup; utilities and commissions examine service and cost allocation; security agencies examine technology and data flows; communities examine local burdens, credibility, and the distribution of benefits. One press release cannot satisfy all of them.

Build one decision map that links every institutional claim to an owner, authority, evidence item, status date, dependency, affected group, and remedy. A favorable land vote cannot substitute for a utility agreement; a customer announcement cannot substitute for finance; an air permit cannot settle water, noise, tax, or labor questions. The map should also record who can withdraw, appeal, modify, enforce, or renew each permission. That makes disagreement legible and prevents a conditional milestone from being narrated as blanket approval for the full campus life.

Training concentrates large pools of accelerators around synchronized jobs that can run for extended periods, checkpoint, restart, and sometimes shift in time or location. Inference serves recurring requests whose latency, availability, privacy, and geographic requirements can favor distributed deployment. Fine-tuning, evaluation, data processing, and synthetic-data generation sit between those simple categories.

PyTorch documents data, fully sharded data, tensor, and pipeline parallelism as complementary approaches to distributed training. Each approach changes how compute, memory, and communication are consumed. A demand forecast that counts accelerators without the workload's parallel structure can misstate network needs, useful utilization, and the amount of capacity a job can actually employ.Claim

Utilization can be lost to data preparation, communication, scheduler gaps, failures, checkpointing, software incompatibility, thermal limits, power constraints, or simply insufficient demand. A high device-activity metric may still hide poor application throughput. Operators need workload-level measures tied to service quality and completed work, not only equipment occupancy.

The IEA estimated that data centers used about 415 TWh globally in 2024. That system-wide estimate should not be allocated to AI or a particular model without measured boundaries. For a project, model economics must connect capital cost, energy, cooling, networking, staff, maintenance, financing, depreciation, and useful output under explicit utilization and hardware-life assumptions.Claim

Utilization is not one number. Device activity, reserved capacity, scheduler occupancy, application throughput, and billable service can move differently. Underwrite them as a time series by workload class, including ramp, maintenance, failures, seasonality, customer concentration, and capacity held for service quality. Stress both excess demand and weak demand: congestion can degrade revenue just as idle equipment can. The audit should show which party bears take-or-pay commitments, curtailment, migration cost, performance penalties, hardware obsolescence, and the gap between technical output and customer value.

A hyperscaler may develop and operate its own campus, lease powered capacity, or place hardware in a colocation facility. A developer may finance land and infrastructure against customer commitments; utilities may build dedicated or shared upgrades under approved terms; equipment may be purchased, leased, or supplied through service arrangements. Each structure locates completion, demand, technology, and residual-value risk differently.

Global data-center investment reached an estimated half trillion dollars in 2024, according to the IEA. Aggregate spending does not reveal whether a specific project has firm customers, committed equity, closed debt, ordered equipment, utility milestones, or sufficient contingency. Finance diligence must work at that lower level.Claim

Procurement commitments can carry as much schedule risk as financing. DOE reports that distribution-transformer lead times grew from roughly three to six months in 2019 to twelve to thirty months in 2023. That national indicator is not a quote for every project or transformer class, but it demonstrates why equipment slots, specifications, deposits, substitutions, and delivery evidence must be underwritten explicitly.Claim

The four disclosures below describe infrastructure-adjacent spending by different companies, but their accounting definitions and fiscal periods are not harmonized. They are useful as evidence of corporate investment scale, not a project ledger or an AI-only comparison. Read each filing for the included assets, cash versus noncash treatment, leases, depreciation, geographic scope, and period. Never infer completed data-center capacity, utilization, revenue, or financing from expenditure alone; reconcile announcements to balance-sheet additions, cash flow, commitments, and placed-in-service status.

The IEA base case projects worldwide data-center electricity demand to about 945 TWh in 2030. It is a scenario-dependent projection, not a project revenue guarantee. Deployment, efficiency, utilization, power constraints, and market conditions can move the result. A developer should not treat a macro forecast as a substitute for site-specific demand evidence.Claim

A coherent model aligns delivered megawatts, rack deployment, utilization ramp, customer start dates, tariffs, fuel or energy contracts, maintenance, replacement cycles, and financing terms. Sensitivity analysis should reveal which assumptions control viability and whether downside protections merely transfer cost to customers, utilities, suppliers, or the public.

Announced means an intention has been made public. Land-controlled adds a site right. Permitted means specified approvals are in hand, not that every condition has been satisfied. Utility-approved records a defined service decision. Under construction describes active works. Energized means electrical service is physically available under stated conditions. Operating means systems are commissioned and serving workloads.

These states can overlap by phase: one building may operate while another is only land-controlled. Reporting should name capacity at each stage, the date verified, key conditions, expected transition, and evidence owner. This prevents pipelines from being totaled as if every proposed megawatt will arrive on the same schedule.

LBNL's 2030 reference case for U.S. data-center electricity use includes wide compounded-uncertainty bounds rather than a single certain future. The IEA global outlook is also scenario-dependent. Together they show why planners should stress deployment, efficiency, workload, and infrastructure assumptions rather than extrapolate one growth line.ClaimClaim

Stranding risk includes an unavailable interconnection, changed permit conditions, construction inflation, obsolete rack geometry, inadequate cooling, network limitations, customer concentration, model-efficiency shifts, export restrictions, or a local relationship that collapses. Modular phasing and reusable utility infrastructure can preserve options, but only when the alternate use is technically and contractually realistic.

The U.S. CHIPS and Science Act provides 52.7 billion dollars over five years, including manufacturing incentives and research-and-development funding. NIST's awards catalog, updated August 2026, distinguishes proposed and final project records. Final direct-funding amounts are maximum, milestone-based awards—not evidence of cash disbursed, completed construction, qualified output, or private investment realized.ClaimClaim

Industrial-policy programs can seek domestic capacity, supply resilience, technology leadership, national security, workforce development, regional growth, or research infrastructure. A recipient must translate those objectives into project scope and compliance systems. Analysts should distinguish authorization, appropriation, award announcement, executed agreement, disbursement, milestone achievement, and operating output.

The European Chips Act sets a goal of doubling the EU share of global semiconductor production to twenty percent by 2030. The figure is a policy target, not a forecast. The 2023 act remains in force while Chips Act 2.0, presented in June 2026, remains a proposal; analysis must not treat the proposal as enacted replacement law.ClaimClaim

Comparing industrial-policy programs requires more than headline currency conversion. Programs may support factories, equipment, research, tax credits, loans, infrastructure, skills, or strategic coordination, and they may measure success differently. Public value should be evaluated against additional capacity, capability, resilience, spillovers, labor outcomes, environmental performance, and the opportunity cost of support.

Evaluate incentives against a documented counterfactual: what investment, site, timing, and public revenue would likely occur without the support? Separate statutory maximum, awarded amount, claimed amount, cash paid, and verified performance. Benefits should be measured net of public infrastructure, administration, financing, displacement, and environmental or service costs rather than quoted only as gross investment. Durable agreements publish definitions, beneficiary ownership, audit access, change-of-control treatment, clawback calculation, appeal process, and enforcement history. Confidentiality should not erase the public's ability to test value received.

The U.S. December 2024 package added controls covering categories of semiconductor-manufacturing equipment, software tools, and high-bandwidth memory, together with Entity List additions. The counts describe that dated package, not the entire current regime. The consolidated EAR contains transaction-specific exceptions, while a January 2026 BIS announcement describes case-by-case review for H200, MI325X, and similarly performing chips under stated conditions. Current text and transaction facts must be checked before action.ClaimClaimClaim

Compliance begins with product and party data: technical classification, performance thresholds, destination, ownership, consignee, end user, intended use, service access, and transaction parties. Engineering changes can alter classification; corporate changes can alter party risk. A static country label is therefore not a complete control process.

Semiconductor design, materials, equipment, fabrication, and assembly remain distributed across specialized regional clusters. That structure creates multiple control points and makes unilateral changes propagate through firms and countries that perform different stages of one product route.Claim

Firms may create market-specific products, relocate support functions, qualify new suppliers, add screening, or reconsider joint research and cloud access. Governments must weigh security objectives against enforceability, allied coordination, domestic industry effects, substitution, and the time needed to build alternate capacity. Evidence should separate the rule's stated objective, legal coverage, observed compliance behavior, and measured industrial outcome.

A sovereignty requirement may concern where data is stored, where it is processed, who can administer systems, which law can compel access, how encryption keys are controlled, or whether hardware and software remain supportable during geopolitical stress. Those are different requirements and should be written separately before an architecture is chosen.

A locally sited facility can still depend on foreign chips, firmware, cloud control planes, network operations, spare parts, and corporate decisions. Conversely, cross-border services can incorporate segmentation, local key control, auditable access, and contractual safeguards. The useful question is which authority and dependency must be controlled for the specified risk—not whether a map pin sits inside a border.

The European Chips Act's production-share target expresses a strategic objective to strengthen regional capability; U.S. export controls express a different objective around technology access and national security. Neither policy claim, by itself, establishes end-to-end autonomy across design, equipment, materials, fabrication, packaging, cloud operations, and energy.ClaimClaim

A resilience plan should identify the service that must continue, the disruption it must survive, the acceptable degradation, and the recovery period. It can then combine trusted suppliers, inventories, diversified regions, interoperable designs, local skills, controlled access, legal agreements, and tested continuity procedures. Autonomy without a bounded threat model becomes an unlimited and unmeasurable objective.

Ireland's regulator reports the national electricity share of data centers as an observed five percent in 2015 and twenty-two percent in 2024, with thirty-one percent for 2034 presented as a projection. This is a country-level electricity-demand share, not an AI-only measure, installed-capacity ratio, or forecast for every grid. The sequence shows why sovereignty and infrastructure policy converge: connection rules must address system adequacy, location, flexibility, and decarbonization while preserving transparent assumptions and a route to revise forecasts.

FERC regulates specified interstate electricity matters and wholesale market structures; NERC develops and enforces reliability standards under federal oversight; regional operators plan and operate organized systems in much of the country; state commissions oversee utilities and retail terms under state law; local bodies govern land use and construction. Authority changes with the asset and transaction.

A private power contract cannot displace a required interconnection study, utility tariff, environmental permit, zoning decision, or reliability obligation. Project maps should show the legal authority, technical decision, cost decision, schedule, public process, and appeal or compliance path for each institution involved.

FERC Order No. 2023 reforms generator-interconnection processes through cluster studies, project-readiness requirements, and transmission-provider delay penalties. The rule is effective, but FERC's chair said in June 2026 that transmission-provider compliance was not fully complete. It therefore cannot be read as a universally implemented tariff or project schedule.ClaimClaim

The NRC regulates the design, siting, construction, and operation of new commercial nuclear facilities, with formal opportunities for public involvement. A data center considering nuclear supply must distinguish buying energy from the grid, contracting with an existing plant, colocating with generation, and developing a new facility; each creates different technical and legal questions.Claim

Congressional Research Service finds that individual data centers typically fall under state and local siting, while the federal nexus varies by project. FERC, NRC, the Army Corps, and delegated air and water programs may apply depending on location and configuration. The report is a jurisdictional map, not a permit determination for any site.Claim

A July 2025 executive order directs federal agencies to examine categorical exclusions, FAST-41 coordination, brownfield guidance, and federal-site opportunities for data-center infrastructure. These are review and coordination mechanisms, not permits; they do not predetermine outcomes or erase substantive project requirements.Claim

The stack can include zoning, subdivision, site plan, building, fire, stormwater, road access, wetlands, water withdrawal, wastewater, air emissions, cultural resources, threatened species, transmission, generation, and utility approvals. Some apply to the campus; others attach to a substation, pipeline, backup system, or off-site upgrade.

EPA's 2026 redevelopment guidance says candidate sites should fit site conditions, reach power and fiber, comply with regulation and cleanup controls, and incorporate planning and community input. The guidance concerns redevelopment screening and does not confer approval, but it demonstrates how physical suitability, legal compatibility, and engagement intersect.Claim

A permit matrix should connect each decision to the design basis it relies on: acreage, building envelope, generators, fuel, cooling, water, discharge, traffic, noise, lighting, construction phases, and operating hours. Teams should record submission, completeness, review, conditions, appeal, expiration, modification triggers, and compliance reporting separately.

A federal environmental-review process may arise when a project has the necessary federal nexus; state and local review laws can create separate requirements. The applicable process depends on governmental action and project facts. Public participation may include notices, document review, hearings, written comments, consultations, responses, and opportunities to challenge decisions.

Participation is most useful when materials are understandable, alternatives are genuine, translation and access needs are addressed, and the project explains what feedback changed. A record of comments is not the same as agreement, and a sparsely attended meeting is not evidence that affected people lack concerns.

CRS emphasizes that the federal permitting nexus varies with project configuration and location. A complete review map therefore follows associated infrastructure and governmental decisions rather than assuming every effect belongs to one lead campus approval.Claim

EPA's redevelopment guidance highlights site conditions, regulation, cleanup controls, infrastructure access, planning, and community input. Those screens can expose cumulative questions early: whether a new load shifts generation, whether water withdrawals interact with existing users, whether backup engines affect local air, and whether a brownfield remedy is compatible with construction.Claim

A load study examines location, magnitude, ramp, hourly profile, power quality, redundancy, restoration expectations, and interactions with transmission and distribution. The result may require substations, lines, transformers, protection, generation, or operating agreements. A reservation in a queue is not the same as completed network capability.

LBNL's 2030 U.S. reference case is 649 TWh with broad compounded-uncertainty bounds. It is a national scenario, not a utility forecast, but it shows why planners need transparent, regularly refreshed load pipelines. Duplicate requests, speculative phases, efficiency changes, and cancellations can otherwise drive both overbuild and delayed preparation.Claim

Utilities and commissions can consider line-extension policies, contributions in aid of construction, minimum-demand terms, contract duration, exit fees, security, special tariffs, riders, interruptibility, and treatment of shared upgrades. The governing options vary by jurisdiction and utility structure. Transparent proceedings should distinguish project-specific facilities from network investments that serve broader users.

FERC's generator-interconnection reform concerns generation queues, not retail data-center tariffs, but it illustrates the importance of readiness requirements, cluster planning, and accountability for delay. Large-load policy needs its own defined rules so speculative requests do not block credible projects and stranded upgrades do not silently migrate to other customers.Claim

Virginia's legislative analysts modeled a wide range of future system outcomes and estimated that by 2040 residential customers could pay between fourteen and thirty-seven dollars more per month under the examined data-center demand cases. These are modeled statewide monthly effects, not current bills, a tariff decision, or a prediction for every household. The useful audit question is which load, generation, transmission, financing, contract, and cost-allocation assumptions drive the range—and which enforceable customer protections remain if projects arrive late or consume less than forecast.

EPA's redevelopment guidance links candidate-site fit with power and fiber access, regulatory compliance, cleanup controls, planning, and community input. CRS likewise shows that jurisdiction follows project configuration. Together they argue for an operating register that ties each condition to a responsible owner, monitoring method, reporting date, and corrective path.ClaimClaim

St. Louis's April 2026 Armory approval provides a project-level example: the city published conditions for cooling and water, noise and generator operation, PUE, heat-rejection and reuse reporting, certified e-waste handling, decommissioning, long-duration job floors, and financial or occupancy remedies. These are auditable obligations and future commitments—not evidence that the projected facility performance, jobs, tax receipts, or community benefits have already been delivered.Claim

Some obligations are physical—installing noise attenuation or stormwater controls. Others are procedural—maintaining records, testing equipment, notifying agencies, or opening complaint channels. Conditions also need change control: new generators, cooling modes, fuel, load, building phases, or operating schedules may trigger review rather than fit automatically under the original approval.

LBNL estimated average U.S. data-center site water usage effectiveness at just over 0.36 liters per kilowatt-hour through 2023. The estimate is a modeled national average; individual facilities vary materially. It should not be applied to a proposed site without a compatible WUE definition, load profile, cooling mode, and climate assumption.Claim

A local utility record answers a different question. Prince William County reported that 25 operating data-center buildings averaged about 18,000 gallons per building on an average day and about 88,000 gallons on a maximum day. Together they represented about 1.5 percent of Service Authority average-day demand and 3.9 percent of maximum-day demand. Those historical building and system boundaries should remain attached; they are not a campus design allowance or a national benchmark.Claim

Local review also asks questions a ratio cannot answer: source reliability during drought, competing users, seasonal peaks, treatment energy, wastewater destination, chemical handling, emergency operation, and whether water-saving modes increase electricity or other impacts. A project should disclose both annual totals and relevant peak conditions under defined operating scenarios.

Potential benefits include construction work, permanent operations roles, supplier demand, tax revenue, infrastructure investment, training, and reuse of difficult sites. Virginia JLARC estimated 74,000 supported jobs, 5.5 billion dollars of labor income, and 9.1 billion dollars of GDP annually, while finding that most economic benefit came from construction and that data-center revenue ranged from under one percent to thirty-one percent of total local revenue across five mature local markets. These are statewide and local-market estimates, not promises for a new project.ClaimClaim

Potential burdens include electricity-cost exposure, land conversion, water demand, emissions, noise, traffic, housing pressure, visual change, construction disruption, and opportunity costs. JLARC modeled generation- and transmission-related costs of fourteen to thirty-seven dollars per month in real 2040 dollars for a typical Dominion residential customer under its scenarios. It also found one-third of Virginia data centers near residential areas, documented low-frequency noise challenges at some sites, and recommended explicit local authority to require proposed water-use estimates. None of those findings is a universal facility forecast.ClaimClaimClaim

Temporary and permanent jobs should be separated, as should gross tax revenue and net fiscal effect after incentives or public infrastructure. St. Louis's Armory framework illustrates the distinction: it specifies escalating twenty-year job minimums, projected-tax shortfall damages, operating conditions, and enforcement routes. The existence of those terms is evidence of an accountable commitment structure; employment, revenue, environmental performance, and payments remain future results that must be reported separately.Claim

In Pew's January 2026 U.S. survey, thirty-nine percent said data centers were mostly bad for the environment, thirty-eight percent mostly bad for home energy costs, and thirty percent mostly bad for nearby quality of life. In the same survey, twenty-five percent called them mostly good for local jobs and twenty-three percent mostly good for tax revenue.ClaimClaim

Those estimates describe a national probability-based survey at one moment. They do not predict sentiment around a specific project. Public-hearing comments are valuable stakeholder positions but self-selected; they are not representative polling. Strong reporting keeps each evidence type separate and publishes question wording, field dates, population, method, and project context.

A project can affect treaty resources, cultural landscapes, archaeological sites, water, habitat, access, or community infrastructure beyond a parcel boundary. Federal, state, or local actions may create distinct consultation and review processes. Developers should not substitute an open house or stakeholder mailing for government-to-government responsibilities carried by the appropriate authority.

CRS explains that the federal nexus varies by data-center project, while EPA's siting guidance places community input alongside site conditions and regulatory compatibility. Those sources orient the inquiry; they do not determine which Tribal Nations are affected or what consultation is legally sufficient for a particular location.ClaimClaim

Community engagement should begin while alternatives remain open: site layout, construction routes, water systems, noise controls, workforce programs, tax agreements, emergency response, and monitoring. Participants need enough technical and financial information to understand tradeoffs, independent expertise where appropriate, and a clear record of what can still change.

Legal approval and social acceptance can diverge. A permit may authorize construction while trust continues to erode; opposition may also persist despite mitigations and benefits. A credible process does not promise unanimity. It documents stakeholder positions, explains decisions, assigns commitments, funds monitoring, maintains a grievance path, and revisits agreements as operations or ownership change.

Do not use consultation, engagement, agreement, and consent as synonyms. Their legal meaning depends on jurisdiction, rights, authority, and context. A community-benefit or impact agreement should name the parties and representation, define obligations and eligible beneficiaries, fund administration, publish deliverables where lawful, protect culturally sensitive information, and provide monitoring, dispute resolution, remedies, succession, and termination rules. The audit also asks whether participation was early enough to influence siting and alternatives, whether power imbalances were addressed, and whether saying no or withholding agreement had a defined consequence.

UNITAR estimates that the world generated sixty-two million tonnes of electronic waste in 2022, with only 22.3 percent documented as formally collected and recycled. This is a global all-electronics estimate, not a data-center or AI subtotal. It supplies context for why asset disposition requires traceable evidence.Claim

Refresh planning should begin with asset identity, data sensitivity, condition, remaining workload value, repairability, energy performance, warranty, spares, and software support. EPA's federal-equipment hierarchy places reuse and redeployment of functional electronics ahead of recycling, with repair, refurbishment, component harvesting, and secure data handling used where practicable. The guidance is not a measured private-sector reuse rate, but it establishes the operational sequence this field guide applies.Claim

When recycling is appropriate, contracts should control custody, worker and environmental practices, downstream processors, recovered fractions, and residual disposal.

UNITAR projects global electronic waste reaching eighty-two million tonnes in 2030 under current trends. The projection is not AI-specific, and policy, reuse, repair, collection, and recycling can change the outcome. Its relevance is the scale of the wider system into which rapid hardware refresh will feed.Claim

Decommissioning should address customer and worker transition, data destruction, equipment removal, batteries and fuels, refrigerants and fluids, utility disconnection, permit closure, contaminated materials, demolition or adaptive reuse, financial security, and long-term monitoring. Commitments should survive ownership changes. The intelligence buildout is complete only when its physical liabilities and public promises remain traceable from announcement through final disposition.