Plant — complete text edition

A data-center campus joins a global technology supply chain to one parcel of land. At that boundary, abstract compute demand meets local transmission, distribution equipment, fiber routes, geology, weather, roads, water systems, construction labor, emergency services, zoning, taxation, and neighbors. No component can be evaluated independently because the cheapest land may have the longest grid schedule, while a power-rich site may lack fiber diversity or an acceptable cooling path.

Site selection is therefore systems engineering under uncertainty. Teams must identify hard constraints, model enabling work, assign schedule and cost risk, and distinguish capacity that is merely discussed from capacity that can be permitted, built, commissioned, and energized. A campus becomes useful only when all critical paths arrive together. Early diligence exists to expose the dependency that would otherwise appear after capital and schedule are committed.

Developers compare land control, utility studies, substation and transmission work, fiber construction, grading, drainage, foundations, environmental controls, water and sewer arrangements, and access improvements. Each can require third-party action outside the developer's direct schedule. Tax incentives may improve project economics, but they do not repair unsuitable soils or create firm electrical capacity on demand.

EPA's redevelopment guidance illustrates the breadth of a responsible screen: candidate sites should fit their physical condition, reach power and fiber, comply with regulation and cleanup controls, and include planning and community input. The guidance addresses redevelopment contexts rather than every data-center project, yet its framing is useful. Suitability is not a single score. It is an evidence trail showing how the site can host the proposed facility without hiding unresolved obligations.

A comparable site model uses the same demand case, reliability target, climate period, and project state for every candidate. Record who controls each enabling asset, what evidence supports availability, when it must arrive, and which dependency sits outside the developer's authority. Then test common-mode failures: one corridor may carry grid power, fuel, fiber, and water access despite separate contracts. The audit should end with residual risks, priced mitigations, decision owners, and a dated trigger for reopening the site choice.

Setbacks, easements, wetlands, drainage corridors, flood zones, slopes, protected resources, utility rights-of-way, and road access remove or condition areas that appear available on a map. Geotechnical investigation determines whether foundations, yards, tanks, and heavy equipment need soil improvement or special design. Campus planning must also reserve future blocks, construction staging, security buffers, stormwater controls, and safe maintenance access.

Hazard analysis extends across flood, wildfire, seismic activity, wind, heat, cold, smoke, dust, and adjacent industrial risk. The response is not to claim that every hazard can be eliminated. It is to define the exposure, design protections, operational thresholds, recovery plan, and residual risk. Climate assumptions should cover the expected facility life and equipment ranges rather than rely only on historical annual averages.

This aerial image depicts a government data-center campus. It is not an OpenAI project, a hyperscale AI training campus, or evidence of the density and equipment described elsewhere in this volume. It is used to reveal site relationships that close photographs hide: buildings sit within roads, perimeter space, mechanical areas, drainage, utility access, and a larger land context.

An AI facility may arrange those elements differently and can demand much greater power or cooling infrastructure. The planning lesson still holds. Site capacity must include everything required to operate and expand safely, not only the white-space floor area. A credible master plan marks the boundaries between current work, reserved expansion, third-party infrastructure, and land that cannot be developed.

A campus needs high-capacity routes to users, other data centers, cloud regions, and carrier networks. Teams inspect long-haul reach, carrier availability, entrance diversity, conduit ownership, splice locations, rights-of-way, and the route inside the property. Logical diversity from two providers is weak if both lease the same buried segment or enter through one room. Construction and repair times for new fiber belong in the project schedule.

The campus network also has lifecycle needs. Additional halls consume strands and pathway space; optical generations change equipment; maintenance requires access without exposing both routes; and a civil project outside the fence can threaten multiple services. Route records should identify shared-risk groups from carrier points of presence to internal distribution. Connectivity is not fully delivered until paths are built, tested, documented, and accepted under failure.

Transformers, switchgear, generators, cooling equipment, steel, electrical rooms, modular assemblies, and racks can demand special transport and lifting plans. Bridge limits, road geometry, delivery windows, laydown space, weather, and security affect when equipment can reach its foundation. Phased campuses need routes that allow later construction without compromising live operations or emergency access.

Labor is another infrastructure layer. Civil, structural, electrical, mechanical, controls, fiber, safety, commissioning, and operations skills peak at different times. Local availability, training, travel, housing, and competing projects can constrain the sequence. A schedule that assumes every trade appears exactly when needed is not a schedule; it is an unpriced dependency. Workforce planning should begin with design and continue through permanent operations.

In the United States, data-center siting is commonly governed through state and local land-use authority, building and fire codes, and utility processes. Federal involvement varies. Wetlands or waterways, federal land or funding, air emissions, water discharges, transmission, pipelines, or a nuclear project can create different federal nexuses. CRS emphasizes that the actual mix is project- and location-specific.

A permit matrix should name the agency, legal trigger, application, studies, dependencies, public process, decision standard, appeal risk, conditions, and renewal obligation. It should also distinguish a filed application from an approval and an approval from a condition that has been satisfied. Construction sequencing must respect these gates; starting enabling work before authority is clear can strand capital or damage community trust.

Announcements describe intent. Land control provides a site interest. Permits authorize defined work. Utility approval establishes a path subject to conditions. Construction creates assets. Energization delivers electrical capability. Operations demonstrate that the commissioned system can host workload. These states are related but not interchangeable, and a multi-building campus may occupy several at once.

Phasing can reduce risk by aligning each block with committed power, cooling, network, and customers. It can also create stranded interfaces when later blocks change density or never arrive. The master plan must identify which shared assets are built up front, who funds them, and how early phases operate independently. Public communication should state the unit—land, shell, IT megawatts, facility load, or energized capacity—and the date of its evidence.

Abilene illustrates the distinction. Crusoe reported Stargate's first phase live and its first two buildings energized in September 2025; OpenAI later described the site as operating on Oracle Cloud Infrastructure with NVIDIA GB200 systems in April 2026. Those dated operator and developer statements support an operating-state label for part of the campus. They do not turn planned later buildings, performance claims, or full-buildout capacity into measured fact.Claim

Phase labels should describe independently operable systems, not marketing increments. For each phase, reconcile land authority, approved design, utility service, emergency access, cooling and water capacity, network paths, staffing, commissioning, and the customer or workload gate. Shared assets need an allocation rule if later construction is delayed or cancelled. Public reporting should retain both the unit and status date—building shell, connected load, commissioned facility load, or usable IT capacity—so an early permit or energization milestone cannot silently become full-campus operating capacity.

A campus schedule connects design packages, permits, site preparation, foundations, structural work, utility infrastructure, electrical and mechanical equipment, network rooms, controls, racks, and testing. Long-lead items may need orders before every downstream detail is final, creating interface risk. Off-site manufacturing and modular assembly can compress field work, but only when dimensions, utilities, firmware, transport, and acceptance criteria are controlled.

The critical path should be modeled as evidence, not ceremony. Each activity needs an owner, prerequisite, promised date, confidence, and recovery option. Shared infrastructure deserves special attention: a delayed substation or cooling plant can hold multiple completed halls. Conversely, phased designs can release useful capacity early if each block has independent protection, controls, and safe access while neighboring construction continues.

Early site and design work preserves choices; equipment deposits, foundations, and utility upgrades progressively lock the project into a location and architecture. Global data-center investment has reached a scale that makes these sequencing decisions economically significant. Yet an industry total does not validate any single campus. Each project still needs demand evidence, phase-specific financing, contingency, and a credible path to energized use.

Change control is essential because compute density can evolve faster than construction. A later rack design may require different floor loading, busway, coolant temperatures, pipe sizes, or network pathways. Flexible interfaces can absorb bounded change, but unbounded future-proofing wastes capital. Teams should state the design envelope, price options at phase gates, and record which assumptions would force redesign.

A data hall must carry rack weight, preserve equipment clearances, route power and network paths, distribute and return coolant, control leaks, manage fire risk, and allow safe replacement. Floor grids, ceiling heights, loading docks, corridors, and overhead structure are not background architecture; they constrain how quickly dense equipment can be installed and repaired.

Modern AI racks may arrive as integrated systems with prescribed power, cooling, and network interfaces. The hall should standardize those boundaries while containing failures. Leak zones, branch isolation, electrical protection, fire detection, access control, and monitoring need coordinated layouts. Mock-ups can reveal conflicts before repetition multiplies them across rows. The goal is a repeatable cell that remains understandable during an alarm, not only a visually ordered room on opening day.

NVIDIA's DGX GB200 documentation describes one rack-scale platform around an approximately 120-kilowatt envelope. Open Compute Project workstreams discuss future AI rack envelopes from 250 kilowatts toward one megawatt alongside liquid-cooling standards. These are a current product specification and a forward-looking roadmap, respectively—not evidence that every installed rack operates at those levels.

They nevertheless expose the direction of design pressure. Cabling, busways, protection, coolant distribution, floor loading, heat rejection, controls, and technician safety all change as more power concentrates in one service zone. A hall designed around much lower densities cannot be upgraded by replacing servers alone. Facility envelopes should declare the supported rack mix, diversity assumptions, coolant conditions, and degraded-mode limits.

Power may enter through transmission or distribution service, pass through a campus substation, medium-voltage switchgear, transformers, distribution gear, UPS systems or other ride-through devices, busways, rack power shelves, and board-level converters before reaching silicon. The exact architecture varies, but every stage must be rated, coordinated, monitored, maintained, and cooled.

Protection engineering decides which breaker or relay acts during a fault so the smallest safe portion is isolated. Redundancy decides which path carries load during maintenance or failure. Efficiency determines how much purchased energy becomes IT work rather than heat. Power-quality studies address voltage, harmonics, grounding, and fast load behavior. The facility and compute teams must share these assumptions because platform changes alter the electrical system's operating profile.

Trace electrical capacity from the utility delivery point to the computing load with voltage, fault duty, protection, conversion loss, redundancy state, and environmental derating at every interface. A headline service rating can exceed the amount usable by racks after auxiliaries and reserved redundancy, while a downstream bottleneck can strand upstream equipment. The model should represent normal, maintenance, and contingency configurations; identify which components are concurrently maintainable; and show whether controls, batteries, fuel, cooling, or operator response become the limiting dependency during each transition.

Substations contain transformers, breakers, bus, instrument transformers, protection, controls, and communication systems that connect a large load to the wider network. Their design and construction can sit on the campus critical path, while their failure modes can span many halls. Utility ownership boundaries determine who specifies, operates, and repairs each element.

The image shows a military-installation substation, not equipment serving a hyperscale AI campus. It is included as documentary context for the industrial footprint behind digital load. An AI project may use different voltage levels, ownership, redundancy, and layout, but it still depends on heavy electrical infrastructure, trained operators, spare strategy, clearances, and coordinated protection before any server can energize.

When utility service degrades, protection, ride-through, batteries, generators or other backup sources, controls, and load policies must respond in a defined order. Some systems bridge only seconds; others support longer operation. Fuel supply, emissions limits, maintenance state, ambient conditions, starting reliability, and concurrent failures all bound the duration that nameplate backup can provide.

Resilience design starts with service objectives and failure scenarios. It identifies which loads remain, which can be shed, how cooling continues, and when software should checkpoint, migrate, or shut down. Shared dependencies—controls networks, fuel systems, cooling pumps, switchgear rooms—must be included. A duplicate server path is irrelevant if both sides depend on one untested control or heat-rejection system.

Commissioning verifies documents, installation, controls, setpoints, alarms, sequences, performance, and maintainability from individual components through the integrated facility. Functional tests challenge normal and failure modes: utility loss, equipment isolation, transfer, cooling interruption, sensor failure, network loss, and restoration. Temporary load banks or staged IT loads help prove behavior before valuable compute carries the risk.

Results need traceability. Deficiencies are assigned, corrected, retested, and incorporated into as-built records and operating procedures. Operators participate before handoff so they understand the plant's real behavior, not only its design narrative. A hall is energized when power is present; it is operational when the organization can safely run, observe, maintain, and recover it within stated limits.

Electrical work entering processors, memory, networks, storage, and power electronics ultimately appears as heat that must leave the equipment. The path runs from semiconductor junctions through package materials and thermal interfaces to a heat sink or cold plate, then through air or liquid distribution, facility loops, and heat-rejection equipment into the environment. The weakest interface limits safe operation.

Density makes boundaries visible. A facility can have enough total cooling capacity yet fail at one cold plate, hose, manifold, row, pump, or exchanger. Designers therefore specify temperatures, flow, pressure, water quality, allowable approach, redundancy, and control response at every handoff. Compute telemetry and facility telemetry must share a timeline so throttling, hotspots, pump changes, and ambient conditions can be attributed correctly.

A rack power specification bounds design, but actual heat varies with configuration, utilization, power caps, workload phase, and efficiency. Transient changes can be rapid at the electronics while fluid and building systems respond more slowly. Controls must keep temperatures inside safe limits without hunting, starving neighboring branches, or turning every short burst into an unnecessary plant response.

Thermal qualification combines steady state, transients, partial load, maximum ambient conditions, and component failures. It confirms not only that equipment avoids shutdown, but that performance remains predictable. A design that depends on continuous throttling has converted missing cooling into hidden lost compute. A design with excessive headroom may waste capital and pumping or fan energy. The operating envelope should make this tradeoff explicit.

Air cooling moves conditioned air through equipment, captures heat in the exhaust, and returns it to cooling equipment. Containment prevents hot and cold streams from mixing, while fans overcome equipment and room resistance. Air is familiar and serviceable, but fan power, acoustic load, airflow volume, and local hotspot control become more challenging as heat concentrates.

Direct-to-chip liquid cooling places cold plates on the hottest components and carries much of their heat through liquid loops. Other components may still depend on air, making many deployments hybrid. Immersion places equipment in a dielectric fluid and changes component, service, material, and safety assumptions. The correct choice follows heat flux, rack density, climate, water strategy, maintainability, supply support, and the actual platform warranty envelope.

A direct-liquid system couples cold plates, hoses or blind-mate connectors, rack manifolds, coolant distribution units, facility loops, pumps, heat exchangers, controls, and leak response. Each vendor boundary needs fluid specifications, cleanliness, pressure, temperature, materials compatibility, connection procedure, and responsibility for damage. Small impurities or trapped air can compromise narrow passages where the cooling value is highest.

Serviceability must be designed alongside thermal performance. Technicians need isolation, drainage, spill containment, safe access, replacement tools, and verified restart procedures. A redundant pump does not protect a shared manifold, and a leak sensor is useful only if controls and people know what to isolate. The plant should be able to lose a defined cooling component and move to a known power or workload state without surprise.

A coolant distribution unit can exchange heat between a technology loop serving racks and a facility loop connected to the broader plant. It controls flow, pressure, and temperature while pumps and valves adapt to changing branches. This separation helps protect sensitive equipment chemistry and pressure limits, but adds heat-exchanger approach temperature, control dependencies, filters, sensors, and maintenance requirements.

Distribution design balances pipe size, pump energy, redundancy, isolation, and future growth. Branches need enough authority to receive design flow without starving distant or parallel loads. Water quality and material compatibility must cover every wetted component. Commissioning flushes, cleans, tests, balances, and records the loop before attachment to valuable equipment. Operations then tracks differential pressure, temperature, conductivity, filter condition, makeup, and signs of leakage or corrosion.

This photograph documents cooling infrastructure work associated with a government high-performance computing environment. It is not presented as a hyperscale commercial AI data center or as evidence of a particular cooling efficiency. Its role is to show the physical nature of the system: thermal capacity is created through installed mechanical equipment, controlled connections, trained work, and testing.

Retrofits are especially demanding because construction occurs around live services and inherited constraints. Teams must establish isolation boundaries, temporary cooling, contamination control, shutdown sequencing, and rollback. New racks may arrive faster than plant modifications, but energizing beyond verified thermal capacity simply moves the schedule risk into throttling or outage. Cooling readiness should be a formal gate for compute acceptance.

Leak protection is a layered operating system: compatible joints and hoses, controlled installation, pressure testing, accessible isolation, detection placed along credible paths, automatic actions with understood consequences, containment, drainage, response equipment, and trained staff. A detector count does not prove protection if alarms are slow, inaccessible, or mapped to the wrong valve. Audit both nuisance and missed-detection cases, and test whether a local isolation preserves enough cooling elsewhere. Incident records should connect fluid loss, affected equipment, cleanup, downtime, and corrective action.

Cooling towers can reject heat efficiently through evaporation but consume water and require treatment. Dry coolers reject heat to air with less direct water consumption but may need more surface, fan energy, or higher coolant temperatures during hot weather. Hybrid systems move between modes. Chillers, thermal storage, geothermal exchange, or heat reuse may serve particular climates and loads, each with its own power, water, maintenance, and permitting implications.

Selection should use hourly weather and workload conditions, not only annual averages. The design must cover extreme heat, smoke or dust, freezing, noise limits, water availability, discharge rules, treatment chemicals, drift, and visible plume where relevant. Heat reuse is valuable only where a nearby demand, temperature level, commercial arrangement, and reliable operating match exist. The atmosphere is part of the cooling circuit even when it is outside the fence.

Water usage effectiveness usually relates site water consumption to IT energy, but results depend on cooling architecture, climate, utilization, season, water source, and the precise accounting boundary. LBNL estimated a U.S. average site WUE just over 0.36 liters per kilowatt-hour through 2023. That modeled national average is not a default ratio for every project and should not be applied blindly to a different cooling profile.Claim

Project reporting should separate withdrawal from consumption, potable from reclaimed or other sources, and site use from water associated with electricity generation. Monthly or seasonal values can reveal peaks hidden by annual totals. Local basin conditions and competing uses give the volume meaning. A dry-cooled site may reduce direct water use while increasing electricity demand; the tradeoff belongs in one system boundary.

Power usage effectiveness divides total facility energy by IT-equipment energy; it does not measure useful computing work, grid emissions, water stress, or embodied impact. The LBNL series below models a U.S. stock average, so it should not be read as measurements from one operator or a guarantee for a new hall. The 2028 endpoints are projections under efficiency assumptions. Compare facilities only after aligning IT-boundary rules, partial load, climate, redundancy state, and treatment of offices, generation, export, and shared campus systems.

Annual energy describes megawatt-hours consumed over time. Serving a campus also requires dependable capacity during peak and stressed hours, transmission and distribution paths, voltage support, reserves, protection, and equipment that can tolerate the load shape. A power purchase agreement may finance or attribute generation, but it does not by itself prove that physical supply can reach the site whenever the campus operates.

Supply portfolios can include utility generation, bilateral contracts, renewables, gas turbines or engines, nuclear, geothermal, storage, fuel cells, campus microgrids, and backup generators. Each differs in construction time, fuel or weather dependence, emissions, water, operating role, regulation, and ability to serve continuously. Comparisons should state whether a resource supplies annual energy, accredited capacity, short-duration ride-through, backup, or verified load flexibility.

This IEA estimate allocates the physical electricity supplying data centers worldwide in 2024; it is not a contractual procurement ledger and the displayed categories do not total one hundred percent because other sources are outside the selected comparison. Geography and hour matter: annual certificates or contracts can differ from the generators serving load at a particular time and place. Use the mix to explain system exposure, then pair it with regional dispatch, marginal effects, new-build timing, transmission, curtailment, storage losses, and contract terms.

The IEA's 2025 report estimates that data centers used about 415 terawatt-hours worldwide in 2024 and projects about 945 terawatt-hours in its 2030 base case. LBNL's December 2024 report placed U.S. 2028 demand between 325 and 580 terawatt-hours; its June 2026 update instead models a 649-terawatt-hour U.S. Reference Case for 2030 with 521-to-843-terawatt-hour compounded-uncertainty bounds. The later publication uses a different horizon and updated model rather than explicitly superseding the older 2028 series, so both remain visible with their publication dates instead of being blended into one forecast.ClaimClaimClaimClaim

Their shared implication is planning pressure, not inevitability. Hardware efficiency, workload demand, utilization, deployment pace, geographic shifts, and delayed energization can change the result. Grid planners and developers need scenarios that preserve those drivers and show their effect on substations, transmission, generation, and customer costs. The correct response to uncertainty is staged evidence and flexible design, not false precision.

New generation commonly enters interconnection studies that assess network impacts and identify upgrades. In the United States, FERC Order No. 2023 requires cluster studies, stronger project-readiness requirements, and transmission-provider delay penalties for generator interconnection. The rule is effective, but FERC's chair said in June 2026 that transmission-provider compliance remained incomplete, so it is not evidence of one universally implemented schedule.ClaimClaim

Large-load requests have their own utility and regional processes. In June 2026, FERC opened targeted proceedings for six RTOs and ISOs to justify or reform practices spanning flexible service, cost allocation, and co-location. Those show-cause orders start regional tariff work; they do not establish a completed national connection rule. Developers must still align load ramp, substation work, generation milestones, tariffs, security, and commissioning.Claim

LBNL's queue snapshot records proposed U.S. generation and storage seeking transmission interconnection at the end of 2024. It is a development pipeline, not delivered capacity: projects differ in maturity, location, commercial viability, network upgrades, and probability of completion. Generation and storage also are unlike quantities, so the bars should not be added. A data-center plan must identify which resource, network work, study position, cost responsibility, operating date, and contractual right can actually serve its node.

Some workloads can shift in time, move between regions, slow through power caps, or pause after a checkpoint. Batteries can bridge short events, and staged energization can align growth with network readiness. Other workloads have strict response deadlines, data-locality requirements, safety constraints, or expensive interruption costs. The flexible portion is therefore smaller and more conditional than total nameplate load.

To count flexibility, operators need telemetry, control authority, response time, duration, recovery behavior, and an agreement defining when action is called. LBNL's maturity model treats evaluation, measurement and verification plus enabling data infrastructure as distinct capabilities. Tests must show that a reduction at the grid meter persists without shifting the problem into cooling, backup emissions, or later rebound peaks.Claim

Measurement is counterfactual as well as physical. FERC-hosted guidance compares observed load with an estimated baseline, requires uncertainty to be characterized, and treats post-event rebound separately. PJM's current regional rules add product-specific registration, interval-meter or approved submeter data, notification, and settlement. DOE's 2026 monitoring analysis further distinguishes waveform telemetry for high-frequency AI-load oscillations from billing-grade meter evidence.ClaimClaimClaim

A primary field demonstration provides a deliberately narrow measured case: one 256-GPU cluster in a commercial Phoenix facility reduced cluster power by twenty-five percent for three hours while maintaining the study's defined quality-of-service guarantees. It does not establish whole-facility flexible megawatts, availability for other workloads, fleet-scale recurrence, or a settled market resource.Claim

Operations integrates facility control, IT monitoring, security, maintenance, spares, incident response, environmental obligations, utility communication, and change management. Filters load, batteries age, water chemistry shifts, firmware changes, and equipment enters end of support. Deferred maintenance can preserve a short-term uptime metric while eroding the next failure response. Every asset needs ownership, condition evidence, a safe maintenance state, and a replacement path.

Compute refresh cycles can be shorter than building life, so each new generation tests the original envelopes for weight, power, coolant, network, and heat rejection. Teams should preserve as-built records and compare proposed equipment against measured operating headroom. Decommissioned electronics, batteries, refrigerants, generators, and construction materials need secure reuse, recycling, or disposal routes. Lifecycle planning is part of site capacity, not an afterthought.

The campus depends on utilities, public infrastructure, emergency services, workforce, water systems, and regulatory legitimacy. Its effects can include construction traffic, equipment noise, tax revenue, temporary and permanent employment, emissions, water use, and utility upgrades whose costs must be allocated. These outcomes differ by project and cannot be inferred from a global demand trend or a generic jobs multiplier.

The responsible handoff from Plant to Permission is a dated set of project facts: physical status, utility obligations, permit conditions, operating metrics, forecast ranges, risk bearers, and channels for community input. Public claims should distinguish measured operation from forecast and vendor specification. The buildout earns durability when the industrial system remains legible to the people who operate, regulate, finance, and live beside it.