Data Center

A data center is a purpose-built facility, or a dedicated part of a facility, that houses and interconnects information technology and telecommunications equipment together with the power, environmental-control, safety, and security systems needed to operate it. The International Electrotechnical Commission definition emphasizes centralized accommodation, interconnection, and operation of the equipment, plus the infrastructure needed to provide the intended availability and security.^[1] This makes a data center more than a room of computers: it is a coupled system whose computing, electrical, thermal, network, physical, and operational layers must work together.

Data centers support services ranging from an organization's internal applications to cloud computing, scientific simulation, search, video delivery, and artificial intelligence. An "AI data center" is not a separate class in the principal facility standards. It is usually a data center configured for workloads that use many accelerators, move large datasets, and may require dense, tightly connected clusters. Those requirements can change rack power, cooling, network, storage, and failure-management designs, but the underlying engineering disciplines remain those of data-center design and operation.^[2][3]

Scope and terminology

The boundary used for a data center matters. A campus may contain several data-center buildings, substations, generators, cooling plants, offices, and network meet-me rooms. A building may contain several independently operated computer rooms. Conversely, a small edge site may place a few racks in a telecommunications shelter or other mixed-use building. Comparisons are meaningful only when they use the same physical boundary, measurement point, time period, and workload scope.^[4]

Common labels describe operating and ownership models rather than universal size classes:

Scroll sideways for more →

There is no single, standards-based power or rack-count threshold that makes a site "hyperscale." Capacity figures may refer to utility connection capacity, installed electrical equipment, maximum IT power, contracted customer power, or actual metered load. These quantities are not interchangeable. A planned campus, a building under construction, an energized shell, and a facility filled with operating servers are also different project states.

Development of the modern data center

Early organizational computing often placed mainframes and storage in dedicated computer rooms. As networked services expanded, operators developed larger facilities with independent power distribution, continuous cooling, fire protection, access control, and multiple telecommunications paths. Commercial colocation separated facility operation from ownership of the computing equipment, while cloud operators standardized hardware and software across fleets of sites.

The warehouse-scale-computer model describes a large data center as one computer rather than as an unrelated collection of machines. In that model, workload placement, server design, networking, power delivery, cooling, and fault recovery are co-designed because decisions in one layer affect the others.^[3] This systems view is particularly relevant to machine learning clusters, in which accelerators may communicate continuously during training and storage must feed data to many workers at once.

Large facilities do not eliminate smaller ones. Edge computing places services closer to demand when latency, data locality, intermittent wide-area connectivity, or regulatory constraints justify the loss of some economies of scale. The European Telecommunications Standards Institute describes multi-access edge computing as cloud-computing capability at the edge of an access network, with proximity, low latency, and high bandwidth among its intended properties.^[5]

Facility architecture

IT equipment

The IT layer includes compute nodes, storage, network switches, management systems, and the racks, cabling, and firmware that connect them. General-purpose processors remain important for operating systems and application logic. Graphics processing units and other AI accelerators provide high-throughput matrix and vector computation. Storage can include local devices, distributed object or file systems, and separate high-performance storage networks.

Equipment specifications are not facility measurements. A component's thermal design power, a server's nameplate rating, its measured peak draw, and its long-run average draw answer different questions. Designers use measured workload profiles and conservative engineering allowances to size power and cooling, then confirm behavior during commissioning. Simply adding every nameplate value can overstate expected demand, while relying on average utilization can leave inadequate capacity for simultaneous peaks, startup conditions, or failure scenarios.^[6]

Electrical path

A typical electrical path begins at a utility service or on-site source and passes through switchgear, transformers, transfer equipment, uninterruptible power supplies, distribution panels or busways, and rack-level power distribution before reaching the IT power supplies. The exact topology varies by site. Generators, batteries, fuel cells, or other sources may support interruption ride-through or longer operation, subject to fuel, emissions, maintenance, and permitting constraints.^[4]

Power conversion and distribution introduce losses and heat. Higher distribution voltages can reduce current and conductor losses for a given power level, but insulation, protection, maintainability, and equipment compatibility still govern the design. Protective devices must isolate faults selectively so that a downstream failure does not unnecessarily remove a wider portion of the facility.

Redundancy is commonly expressed with shorthand such as N, N+1, or 2N:

• N is the capacity required to serve the design load.
• N+1 provides one additional unit beyond the number needed at that condition.
• 2N provides two complete capacity sets.

The label alone is not proof of resilience. Common controls, shared switchgear, fuel systems, cooling dependencies, or cable routes can create a common point of failure. A useful review follows every power and control path from source to workload under maintenance, component failure, and emergency conditions.

Thermal management

Nearly all electricity used by IT equipment becomes heat inside the facility. Thermal management transports that heat from chips and other components to the outdoor environment or to a heat-reuse system. The path can contain several stages: chip-to-cold-plate or chip-to-air transfer, server fans or liquid loops, room air or facility water loops, heat exchangers, pumps, chillers, cooling towers, dry coolers, and controls.^[4][7]

Air-cooled rooms commonly arrange racks into hot and cold aisles. Containment reduces mixing between supply air and hot exhaust, allowing cooling equipment to operate closer to the temperatures the IT equipment requires. Air-side or water-side economizers can reduce compressor operation when outdoor conditions permit. Their suitability depends on climate, air quality, humidity, water chemistry, equipment class, and operating objectives.

Liquid cooling moves heat with a liquid closer to high-power components. Major arrangements include rear-door heat exchangers, direct-to-chip cold plates, and immersion systems. A facility may combine liquid cooling for accelerators and processors with air cooling for memory, storage, networking, and power components. Liquid cooling does not by itself establish water consumption: a closed internal loop can ultimately reject heat through a dry cooler, an evaporative system, a district-energy connection, or another heat sink. Water use and energy use therefore depend on the complete heat-rejection design, climate, controls, and operating point.^[8]

Designers must specify fluid quality, materials compatibility, leak detection, isolation valves, pumping, water treatment where applicable, and thermal behavior during loss of power or flow. The Open Compute Project's liquid-cooling guidance also treats commissioning, maintenance access, control integration, and thermal ride-through as system requirements rather than details to be added after equipment selection.^[8]

Network and storage fabric

A data-center network connects servers to each other, to storage, to management systems, and to external networks. Large facilities often use multi-stage Clos topologies built from many switches rather than one very large central switch. Google's published Jupiter network history describes how a centrally controlled Clos-based fabric was scaled over successive generations while maintaining high bisection bandwidth.^[9] The design principle is broader than one operator: scale-out networks divide capacity across multiple paths and use routing or scheduling to balance traffic and tolerate failures.

AI clusters can contain several logically distinct fabrics. A scale-up fabric connects accelerators within a server or rack; NVLink is one vendor technology for that role. A scale-out fabric connects many compute nodes, often using InfiniBand or high-speed Ethernet. Separate networks may carry storage traffic, management, telemetry, and external service traffic. Whether these networks are physically separate or logically isolated depends on performance, security, cost, and failure-domain requirements.

Network capacity cannot be summarized by the peak rate of one port. Relevant properties include end-to-end throughput, oversubscription, latency, congestion behavior, collective-communication performance, route diversity, cable reach, transceiver power, and the effect of a failed link or switch. Storage design likewise considers sustained read and write rates, metadata load, checkpoint bursts, data durability, and recovery time rather than only installed capacity.

Controls, monitoring, and safety

Building-management and electrical-monitoring systems supervise power, cooling, environmental conditions, and alarms. Data-center infrastructure management systems can combine facility measurements with rack or asset data. Useful telemetry includes energy, power quality, breaker state, temperatures, humidity where relevant, fluid flow and pressure, leak detection, equipment status, and network health. Alarm design must distinguish actionable states from noise and must preserve monitoring during the events it is intended to diagnose.

Fire protection is based on applicable codes, occupancy, construction, and equipment risks. NFPA 75 addresses protection of information-technology equipment, including construction, utilities, detection, suppression, emergency procedures, and recovery.^[10] A design may use pre-action sprinklers, clean-agent systems, aspirating smoke detection, or combinations, but no one technology replaces a site-specific fire analysis and legally required protection.

Availability and resilience

Availability is an end-to-end service property. A facility can maintain electrical power while an application is unavailable because of software, storage, or network failure. It can also lose part of a facility while a distributed service remains available through workload migration or replication. Facility topology and software resilience should therefore be designed together.

The Uptime Institute Tier system is one framework for evaluating critical facility topology. Its current descriptions are outcome-based and cumulative:^[11]

Scroll sideways for more →

The Tier Standard does not prescribe a particular technology. Uptime Institute removed mappings between Tier levels and expected annual downtime in 2009, and the current topology standard does not assign availability percentages to the four levels.^[11] Tables that claim, for example, that every Tier III facility provides a fixed percentage of uptime misstate the framework. Certification of a design, a constructed facility, and operational sustainability are also distinct evaluations.

Other standards use broader risk-based classifications. ISO/IEC 22237-1 covers general concepts for data-center facilities and infrastructure, including availability, security, and energy-efficiency classification over the planned lifetime.^[2] Selection should follow the business impact of failure, recovery objectives, regulatory duties, and cost rather than a presumption that the highest redundancy is always appropriate.

Failure domains

A failure domain is a set of resources that can be lost together. Examples include equipment behind one breaker, racks sharing a top-of-rack switch, rows using one cooling loop, or buildings served by one substation. Replicas placed in the same failure domain do not protect against that domain's loss. A 2016 study of replicated tasks in cloud data centers modeled correlated power and network failure domains, including racks, and showed why scheduling and replica placement should not assume independent machine failures.^[12]

Resilience reviews should trace:

• utility, on-site generation, energy storage, and fuel dependencies;
• switchgear, UPS, distribution, and rack power paths;
• pumps, heat exchangers, controls, and heat-rejection paths;
• network, storage, timing, name-service, and management dependencies;
• authentication, monitoring, and remote-access services;
• spare parts, staff access, and vendor support;
• software recovery, data restoration, and workload relocation.

A backup system that has never been tested under representative conditions is only an asserted capability. Integrated systems testing, load-bank testing, failover exercises, restoration drills, and documented acceptance criteria expose interactions that component tests can miss.

Operation and maintenance

Design intent can be defeated by operating practice. Capacity changes, temporary cabling, disabled alarms, deferred maintenance, or an undocumented control change can introduce failure paths after commissioning. Operators use methods of procedure for planned work, change review, peer checking, maintenance windows, rollback plans, and post-change verification.

Uptime Institute's 2025 outage analysis continued to identify power as the leading cause of impactful data-center outages. It also reported that IT and network failures accounted for 23 percent of impactful incidents in its survey data and emphasized the role of procedures in human-error outages.^[13] These are survey findings rather than universal failure rates, but they support a durable conclusion: redundant equipment is not a substitute for controlled operations and tested recovery.

Capacity management should combine facility and IT measurements. A site can have unused utility capacity while a particular busway, cooling loop, cable route, or network tier is full. Conversely, power capping, workload scheduling, or equipment replacement can release usable capacity without expanding every upstream system. Forecasts should identify the constraint, the time resolution of measurements, and the assumptions about utilization and coincidence.

Security

Data-center security includes physical, environmental, network, system, and supply-chain controls. ISO/IEC 22237-6 addresses protection against unauthorized access, intrusion, internal fire and environmental events, and external events.^[14] NIST Special Publication 800-53 provides a wider control catalog that includes access control, audit, contingency planning, incident response, physical and environmental protection, system and communications protection, and supply-chain risk management.^[15]

Physical measures can include site perimeters, vehicle controls, guarded or electronically controlled entrances, identity verification, visitor management, cameras, compartmentalized access zones, secure loading areas, and locked racks or cages. Their selection follows threat modeling and legal requirements. Highly visible perimeter security is ineffective if contractors, remote-management systems, spare equipment, or media handling remain uncontrolled.

Cybersecurity extends below the hosted applications. Facility controllers, firmware, baseboard-management interfaces, orchestration systems, and remote vendor connections can affect availability or expose privileged access. Segmentation, authenticated administration, least privilege, signed updates where supported, vulnerability management, logging, protected recovery credentials, and tested incident-response procedures all matter. Security controls also require safe emergency access: a mechanism that prevents authorized operators from responding to a fire or electrical event can create a different risk.

Metrics and measurement

No single metric describes efficiency, sustainability, or useful work. A measurement program begins by defining the facility boundary, IT boundary, time period, metering points, data completeness, and treatment of on-site generation or exported energy.

Scroll sideways for more →

ISO/IEC 30134-2:2026 defines and standardizes PUE measurement and reporting, including treatment of mixed-use buildings, on-site generation, and unmeasured energy.^[16] A PUE closer to 1 means less facility overhead per unit of measured IT energy, but a lower PUE does not necessarily mean less total energy. A lightly loaded old facility could improve PUE while consuming more energy than a smaller, busier one; two sites can also report different values because their boundaries and climates differ.

ISO/IEC 30134-9 defines WUE for the use phase and introduces measurement categories and reporting rules.^[17] Site WUE can omit water consumed in electricity generation unless that indirect use is separately assessed. It also does not capture embodied water in construction and equipment. For this reason, comparisons should identify whether they include direct on-site water, source water, electricity-related water, or a life-cycle boundary.

ERF measures the fraction of data-center energy delivered for a useful external purpose, such as heating another building or process.^[18] Heat-reuse performance depends on temperature, distance, seasonal demand, infrastructure, and the energy needed to upgrade or transport the heat. A nominal connection is not equivalent to year-round useful reuse.

Server utilization can reveal stranded IT capacity, but the standardized ITEUsv metric uses CPU utilization and explicitly excludes servers whose main computation is performed by auxiliary accelerators.^[19] An AI facility therefore needs workload-relevant measures such as accelerator occupancy, memory and interconnect utilization, completed jobs, latency or service-level compliance, and useful work per unit of energy. These measures are application-specific and should not be combined into a universal score without disclosing their definitions.

AI and high-density computing

AI facilities often run deep learning training, model evaluation, data preparation, or online serving. These workloads have different duty cycles. Large training jobs can keep many accelerators and network links active for long periods, then produce storage bursts during checkpointing. Online services can be latency-sensitive and variable. Batch inference may be schedulable. Treating all AI demand as a constant load or as the profile of one benchmark obscures these differences.

Rack power is equipment-specific

High rack density is a design input, not a definition of an AI data center. NVIDIA's published DGX H100 data-center guide lists a maximum system power of 10.2 kW and a four-system rack design of 40.8 kW.^[20] A later NVIDIA GB200 NVL72 rack-scale system is specified at approximately 120 kW and uses liquid-cooling manifolds and rack busbars.^[21] These are bounded examples for particular vendor systems. They do not establish a universal 40 kW, 100 kW, or 120 kW threshold for AI racks.

The examples show why a refresh can affect more than the electrical feed. Higher-power equipment can change floor loading, rack layout, busway and breaker sizing, cooling-distribution units, pipe routes, water temperatures, leak detection, cable reach, transceiver count, service clearance, and acoustic conditions. A site qualified for one accelerator generation may need substantial changes for another even if aggregate campus capacity appears sufficient.

Cluster interdependence

Tightly coupled jobs are sensitive to the slowest or failed worker. Network congestion, a storage stall, thermal throttling, or one unreliable node can reduce useful work across a much larger allocation. Cluster design therefore considers:

• balanced accelerator, host, memory, network, and storage throughput;
• collective-communication patterns and congestion control;
• checkpoint frequency, storage burst capacity, and restart time;
• placement across electrical, cooling, network, and fire compartments;
• health monitoring and removal of marginal hardware;
• scheduler behavior during partial capacity or maintenance;
• power transients and the response of UPS, storage, and controls.

The objective is completed work within reliability and service constraints, not maximum instantaneous accelerator allocation. A facility can report an efficient PUE while expensive accelerators wait for data or communication. It can also improve useful throughput through software, scheduling, or model changes without adding facility load.

Power variation and flexibility

AI training and use can create rapid changes in power, according to the International Energy Agency's 2026 assessment.^[22] The magnitude and timing depend on hardware, workload, orchestration, and power-management settings. Designers evaluate step loads, harmonics, voltage behavior, UPS modes, generator response, and the thermal effect of transitions. Batteries or other storage can support ride-through or smooth some changes, but their role must be stated: backup duration, power-quality support, peak management, and grid response impose different sizing and control requirements.

Some batch workloads can be shifted in time or place, curtailed, or checkpointed during grid constraints. Interactive services and tightly coupled jobs may offer less flexibility. Claims about demand response should therefore specify the workload fraction, notice time, duration, rebound, data movement, and service consequences. A nameplate connection is not evidence that the entire load can be dispatched like a generator.

Energy use and electricity systems

Data-center electricity use is growing quickly, but current measurements and projections must be distinguished. The IEA estimates that global data-center electricity consumption rose 17 percent in 2025 and reached about 485 TWh. Its central projection roughly doubles that amount to 950 TWh in 2030, around 3 percent of global electricity demand.^[22] The 2030 figure is a modeled central projection, not a measured outcome or a sum of guaranteed projects. The IEA identifies efficiency, adoption, applications, finance, chip supply, electricity equipment, and grid connections as sources of uncertainty.

Growth is geographically concentrated. A modest share of global electricity can create a large local requirement where projects cluster behind the same substations or transmission paths. Connection requests can also overstate eventual load because projects compete for sites and power, facilities fill gradually, and not all announced capacity is built. Planning agencies should separate proposed, permitted, financed, under-construction, energized, and operating capacity.

For the United States, Lawrence Berkeley National Laboratory's 2025 update uses a bottom-up model based on planned IT-equipment shipments, device energy, cooling simulations, facility type, and location. It gives a 2030 reference case of 649 TWh, or 11.8 percent of US electricity, and a compounded uncertainty range of 521 to 843 TWh, or 9.5 to 15.3 percent.^[23] The wide range is part of the result. Quoting the reference case without the model assumptions and uncertainty would imply more precision than the study provides.

Earlier global research showed that rapid growth in data-center computing did not translate one-for-one into electricity growth because server, storage, network, virtualization, and facility efficiency improved.^[24] That historical result does not prove future demand will remain flat, nor does a high project pipeline prove every load will materialize. Scenario analysis is preferable to extrapolating a single hardware, workload, or construction trend.

Grid connection and site power

Large new loads can require substations, transmission upgrades, generation, and long-lead equipment. Site selection therefore includes more than an electricity price. Relevant questions include available firm and non-firm capacity, connection date, fault level, transmission constraints, generation mix, curtailment terms, backup emissions, cost allocation, and the risk that infrastructure is built for load that arrives later or not at all.

The US Department of Energy notes that rate design for large loads can allocate upgrade costs, address stranded-asset risk, and create incentives for flexibility.^[25] These decisions belong to utilities and regulators, not facility engineers alone. On-site generation can reduce dependence on one connection or accelerate a limited phase of a project, but it introduces fuel supply, air emissions, noise, water, maintenance, safety, and permitting constraints. It is not automatically faster, cheaper, lower-carbon, or independent of the grid.

Water, emissions, and material impacts

Environmental accounting should separate at least four scopes:

1. electricity and fuel consumed at the site;
2. water consumed directly for cooling or other facility uses;
3. upstream impacts of electricity, fuels, and water supply;
4. embodied impacts of buildings, generators, batteries, cooling equipment, servers, accelerators, and network equipment.

Cooling creates an energy-water tradeoff. Evaporative heat rejection can reduce electricity use under suitable conditions but consumes water. Dry heat rejection can reduce direct water consumption but may use more electricity or limit performance during hot weather. Hybrid systems change mode according to weather or water constraints. The best choice depends on watershed conditions, power-system emissions, climate, reliability, and the value assigned to each impact.

Mytton's review of data-center water consumption argues for accounting that includes both direct cooling water and indirect water associated with electricity, with transparent organizational and geographic boundaries.^[26] A low site WUE does not prove a low total water footprint if the electricity supply is water-intensive. Conversely, a higher site WUE in one climate cannot be judged without considering its electricity and water sources.

Carbon claims require similar boundary care. A contractual renewable-energy purchase can support renewable generation while the facility continues to draw from the local grid at hours when other generation is marginal. Annual market-based accounting, location-based grid emissions, hourly matching, backup fuel use, and embodied emissions are different measures. Reports should state which one they use instead of describing a facility as simply "carbon-free."

Equipment manufacture and replacement also matter. Accelerator systems contain semiconductors, memory, metals, circuit boards, cooling hardware, and power electronics. Short refresh cycles may improve performance per unit of computation while increasing manufacturing and end-of-life impacts. Life-cycle conclusions require a defined functional unit, such as a completed workload under stated quality and latency constraints, rather than a comparison of device count alone.

Heat reuse can reduce another user's energy demand when temperatures, distance, and demand profiles align. Low-grade waste heat may require heat pumps, networks, and a nearby year-round customer. Reported reused energy should be metered and should account for energy used to deliver or upgrade it, rather than treating all rejected heat as useful.

Location and community effects

ISO/IEC 22237-2 treats building construction and site selection as part of data-center infrastructure design. It covers environmental hazards, access, intrusion, fire, water, building configuration, and construction quality.^[27] A site assessment commonly examines:

• grid and fuel availability, connection risk, and long-lead equipment;
• carrier diversity, fiber routes, latency, and cable landing or backbone access;
• flood, wildfire, seismic, wind, extreme temperature, and air-quality hazards;
• water availability, source resilience, discharge, and watershed stress;
• land, floor loading, transport, noise, lighting, generator emissions, and visual effects;
• construction labor, operations staffing, spare parts, and emergency response;
• taxes, permits, zoning, reporting, and community obligations.

Risk is not captured by distance from a hazard alone. Two "diverse" utility feeds may share a transmission corridor; two fiber paths may enter the site through one duct; separate water supplies may depend on one treatment plant. Due diligence follows upstream dependencies and verifies route diversity.

Communities can bear infrastructure cost, noise, water demand, emissions, land-use change, and construction disruption while receiving different levels of employment and tax benefit. Transparent project descriptions should state peak and expected electricity demand, water source and maximum use, backup generation, construction phase, equipment fill assumptions, and who pays for grid or water upgrades. Confidentiality may limit detail, but vague aggregate campus figures are a poor substitute for clearly bounded quantities.

Regulation is evolving. Under Commission Delegated Regulation (EU) 2024/1364, operators of data centers with installed IT power demand of at least 500 kW report specified information and performance indicators to a European database. The required data include energy, water, temperature, waste-heat, renewable-energy, and capacity information, with aggregate publication provisions.^[28] This is a reporting threshold for that regulation, not a universal definition of a large data center.

Edge and modular facilities

Edge sites trade some centralization benefits for proximity. They can support telecommunications, industrial control, content delivery, sensor processing, and latency-sensitive services. Constraints often include limited space, power, cooling, physical supervision, and maintenance access. Remote sites need robust monitoring, standardized replaceable units, secure remote administration, and a realistic plan for failures that require a visit.

Modular construction can package power, cooling, or IT capacity in repeatable units and add them in phases. "Modular" does not necessarily mean temporary or containerized; it can describe factory-built electrical rooms, cooling skids, or standardized data halls. Benefits can include parallel construction and tested repetition. Risks include interfaces between modules, capacity stranded in fixed increments, proprietary dependencies, and later equipment that exceeds the original module's density or fluid requirements.

Small edge or modular facilities still require fire, electrical, environmental, and security analysis. Their limited scale may make full component redundancy uneconomic, so service resilience may instead come from workload replication across sites. That strategy works only if sites do not share the same upstream network, grid, software, or management failure.

Evaluating a data-center claim

Data-center announcements and comparisons often combine incompatible quantities. A reproducible evaluation asks:

1. What exists? Confirm whether the subject is proposed, permitted, financed, under construction, energized, equipped, or operating.
2. What is the boundary? Identify campus, building, room, rack, IT equipment, or service, and whether shared infrastructure is included.
3. What does capacity mean? Separate grid connection, electrical nameplate, redundant capacity, contracted capacity, maximum IT load, and measured demand.
4. What workload is assumed? State hardware generation, utilization, duty cycle, redundancy, and whether figures are peak, average, or energy over time.
5. How are power and cooling supported? Trace normal and failure paths, control dependencies, fuel, water, and heat rejection.
6. Are metrics comparable? Match boundary, period, category, climate, and data completeness for PUE, WUE, emissions, and utilization.
7. What is measured and what is projected? Give scenarios and uncertainty for future energy, construction, revenue, or capacity rather than presenting them as facts.
8. What is omitted? Check upstream electricity and water, embodied equipment, community infrastructure, and correlated failure domains.

This method avoids two common errors: treating a vendor configuration as an industry-wide threshold and treating a project announcement as operating capacity.

References

1. ^International Electrotechnical Commission. "data centre." *IEC 60050 - International Electrotechnical Vocabulary*. ieclib17.iec.ch/...en%3A741-01-501
2. ^International Organization for Standardization. "ISO/IEC 22237-1:2021, Data centre facilities and infrastructures - Part 1: General concepts." iso.org/...78550
3. ^Luiz Andre Barroso, Urs Holzle, and Parthasarathy Ranganathan. *The Datacenter as a Computer: Designing Warehouse-Scale Machines*. 3rd ed., 2018. doi.org/...978-3-031-01761-2
4. ^US Department of Energy, Federal Energy Management Program. *Best Practices Guide for Energy-Efficient Data Center Design*. July 2024. energy.gov/...-energy-efficient-data-center-design
5. ^European Telecommunications Standards Institute. "ISG MEC: Multi-access Edge Computing." etsi.org/...mec
6. ^Xiaobo Fan, Wolf-Dietrich Weber, and Luiz Andre Barroso. "Power Provisioning for a Warehouse-sized Computer." 2007. research.google/...-for-a-warehouse-sized-computer
7. ^ASHRAE. "Chapter 20: Data Centers and Telecommunication Facilities." *ASHRAE Handbook - HVAC Applications*. handbook.ashrae.org/...a23_ch20_si
8. ^Open Compute Project. *Advanced Cooling Facilities Reference Design Guidance*. 2024. opencompute.org/...sign-guidance-white-paper-pdf-1
9. ^Arjun Singh et al. "Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google's Datacenter Network." *ACM SIGCOMM*, 2015. research.google/...in-googles-datacenter-network-2
10. ^National Fire Protection Association. "NFPA 75: Standard for the Fire Protection of Information Technology Equipment." nfpa.org/...75
11. ^Matt Stansberry. "Explaining the Uptime Institute's Tier Classification System." Uptime Institute, updated April 2021. journal.uptimeinstitute.com/...assification-system
12. ^Mina Sedaghat, Eddie Wadbro, John Wilkes, Sara De Luna, Oleg Seleznjev, and Erik Elmroth. "DieHard: Reliable Scheduling to Survive Correlated Failures in Cloud Data Centers." *16th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGrid)*, 2016, pp. 52-59. research.google/...-failures-in-cloud-data-centers
13. ^Uptime Institute. *Annual Outage Analysis 2025*. 2025. uptimeinstitute.com/...annual-outage-analysis-2025
14. ^International Organization for Standardization. "ISO/IEC 22237-6:2024, Data centre facilities and infrastructures - Part 6: Security systems." iso.org/...82250
15. ^National Institute of Standards and Technology. *Security and Privacy Controls for Information Systems and Organizations*, SP 800-53 Rev. 5. 2020. csrc.nist.gov/...final
16. ^International Organization for Standardization. "ISO/IEC 30134-2:2026, Data centres key performance indicators - Part 2: Power usage effectiveness (PUE)." iso.org/...30134-2
17. ^International Organization for Standardization. "ISO/IEC 30134-9:2022, Data centres key performance indicators - Part 9: Water usage effectiveness (WUE)." iso.org/...77692
18. ^International Organization for Standardization. "ISO/IEC 30134-6:2021, Data centres key performance indicators - Part 6: Energy Reuse Factor (ERF)." iso.org/...71717
19. ^International Organization for Standardization. "ISO/IEC 30134-5:2017, Data centres key performance indicators - Part 5: IT Equipment Utilization for servers (ITEUsv)." iso.org/...66934
20. ^NVIDIA. *NVIDIA DGX SuperPOD Data Center Design Featuring DGX H100 Systems*. 2023. docs.nvidia.com/...latest
21. ^NVIDIA. "DGX GB200 Hardware Overview." *NVIDIA DGX GB200 User Guide*. docs.nvidia.com/...hardware
22. ^International Energy Agency. *Key Questions on Energy and AI*. 2026. iea.org/...key-questions-on-energy-and-ai
23. ^Sarah Josephine Smith et al. *United States Data Center Energy Usage Report: 2025 Update*. Lawrence Berkeley National Laboratory, 2026. doi.org/...P1RP4F
24. ^Eric Masanet et al. "Recalibrating global data center energy-use estimates." *Science*, vol. 367, no. 6481, 2020, pp. 984-986. doi.org/...science.aba3758
25. ^US Department of Energy. "Electricity Rate Designs for Large Loads: Evolving Practices and Opportunities." 2025. energy.gov/...evolving-practices-and-opportunities
26. ^David Mytton. "Data centre water consumption." *npj Clean Water*, vol. 4, article 11, 2021. doi.org/...s41545-021-00101-w
27. ^International Organization for Standardization. "ISO/IEC 22237-2:2024, Data centre facilities and infrastructures - Part 2: Building construction." iso.org/...82248
28. ^European Commission. "Commission Delegated Regulation (EU) 2024/1364 on the first phase of the establishment of a common Union rating scheme for data centres." 14 March 2024. eur-lex.europa.eu/...oj