Data Center

A data center is a room, building, or group of buildings that houses computing systems and associated equipment for processing, storing, and exchanging digital information. It combines information technology—including servers, storage devices, and networking equipment—with supporting systems for electrical power, cooling, physical protection, and operational management. Data centers provide the physical foundation for enterprise applications, Internet services, and cloud computing; a cloud is a way of providing computing resources, rather than a substitute for the facilities that house them. (ibm.com)

Data centers range from organizational computer rooms to large campuses and smaller facilities located near users. Their operation brings together computer engineering, telecommunications, building services, and reliability engineering. Increasing demand for digital services and artificial intelligence also connects their development to electricity supply, water availability, and environmental accounting. Understanding data centers therefore requires attention both to their computing architecture and to the facilities, operating practices, and resource systems that sustain them. (ibm.com)

Historical Development

Dedicated computer facilities emerged with room-sized electronic computers in the 1940s. ENIAC, developed by J. Presper Eckert and John Mauchly at the University of Pennsylvania, was completed in 1945 and publicly unveiled in February 1946. It contained approximately 18,000 vacuum tubes, weighed 30 tons, and occupied more than 1,500 square feet. Housing such equipment required a dedicated installation rather than ordinary office accommodation: the machine’s physical footprint and supporting infrastructure were integral to its operation. These installations were predecessors of the data center, although they housed individual computing systems rather than networks of servers. (computerhistory.org)

During the 1950s and 1960s, mainframes became central resources in corporate and government data-processing centers. IBM’s System/360, announced on April 7, 1964, exemplified the transition toward expandable computing platforms. Its initial six processor models covered a fiftyfold performance range while sharing a common architecture, allowing customers to expand capacity without completely rewriting their software. Computer rooms accommodated processors and peripheral equipment as parts of a centrally operated installation. Time-sharing subsequently allowed multiple users at terminals to share processing resources, separating the location of users from the central machine serving them. (ibm.com)

In the 1980s and 1990s, client–server computing distributed application functions between users’ computers and networked servers, including servers providing shared databases. Smaller machines could be deployed separately, but collections of servers also occupied dedicated computer rooms, including spaces previously used for mainframes. The growth of the Internet and World Wide Web expanded demand for externally accessible hosting and network interconnection. Commercial colocation separated ownership of computing equipment from operation of its physical accommodation: customers installed their hardware in facilities supplying space, power, cooling, and connectivity. Equinix, founded in 1998, illustrates the accompanying development of carrier-neutral facilities, where competing networks could interconnect without relying on a single telecommunications carrier. (ibm.com)

Virtualization changed how computing capacity within these facilities was allocated, but its origins preceded commercial cloud services by decades. IBM’s experimental CP-40 system was operating in 1967, and VM/370 was announced on August 2, 1972. VMware brought virtualization to commercial x86 products with VMware Workstation in 1999. By allowing multiple isolated operating-system environments to share physical hardware, virtualization supported consolidation instead of assigning a separate server to every application. Capacity could be allocated and reconfigured through software, reducing dependence on physical installation work and creating opportunities to use otherwise idle resources. (vm.ibm.com)

Commercial cloud computing extended resource pooling into an on-demand service model. Amazon launched S3 storage on March 14, 2006, followed by the EC2 limited beta that August. The initial EC2 offering used Xen virtualization and charged US$0.10 per instance-hour; its advertised configuration included 1.75 GB of memory and 160 GB of local storage. Customers could obtain computing capacity without buying the underlying servers, while the provider operated the physical infrastructure. NIST’s 2011 definition formalized on-demand self-service, resource pooling, rapid elasticity, and measured service as cloud characteristics. Large shared resource pools provided a basis for hyperscale deployment, while smaller edge facilities placed selected services nearer users to reduce communication latency. These developments introduced complementary forms of centralization and geographic distribution rather than eliminating enterprise computer rooms or colocation. (press.aboutamazon.com)

Facility Types and Ownership Models

An enterprise-owned facility supports the organization that owns it rather than primarily selling hosting capacity to outside customers. The enterprise owns the computing equipment and facility infrastructure and is responsible for deployment, monitoring, maintenance, and management. Colocation separates these responsibilities: the customer places its own servers, storage, and networking equipment in space operated by a third party, which supplies supporting services such as power, cooling, physical security, and connectivity. At Equinix, for example, customers can rent cabinets or segregated cages within shared facilities. Enterprise ownership and colocation therefore differ chiefly in who provides and operates the site, not necessarily in the applications or hardware installed there. (ibm.com)

Managed hosting transfers more of the IT operation to a service provider. In a typical arrangement, the customer leases dedicated servers, storage, and networking hardware, while the provider administers and monitors that equipment. In basic colocation, by contrast, customers retain responsibility for their hardware and software; optional technical assistance does not automatically transfer that responsibility. Equinix’s “Smart Hands” service illustrates how on-site assistance can supplement colocation. The contractual scope matters more than the label: providing space, replacing a component on request, and administering an operating system are distinct services that may be purchased separately. (ibm.com)

A cloud-provider facility houses infrastructure used to deliver cloud computing services rather than merely renting equipment space. The provider operates the underlying hardware and service infrastructure, while customer responsibilities depend on the service consumed. For Amazon EC2, AWS manages the physical infrastructure and virtualization layer, but customers manage guest operating systems and applications. For services such as Amazon S3 and DynamoDB, AWS also operates the operating system and platform. Physical location does not uniquely determine this delivery model: AWS Outposts places AWS-managed computing capacity in customer-managed buildings, whose owners must supply suitable facilities, power, and networking. Thus, “on premises” does not necessarily mean enterprise-operated IT. (aws.amazon.com)

Hyperscale describes an approach to deploying and expanding computing infrastructure at very large scale, not a separate ownership category or a synonym for every cloud facility. It emphasizes scale-out capacity, extensive automation, repeatable infrastructure designs, and economies of scale. A large operator may build its own campus or obtain capacity from a colocation provider; hyperscale customers and colocation landlords can therefore participate in the same installation. Conversely, cloud services also operate from smaller facilities. Counts such as 5,000 servers appear in industry descriptions, but a server-count threshold alone does not explain the operational architecture or distinguish cloud delivery from other large-scale computing uses. (ibm.com)

An edge facility is classified by its position relative to users, devices, or sources of data. Its purpose is to place processing nearer those endpoints, reducing communication latency or the need to transport data to a central location. Edge installations may support local industrial processing or content delivery and need not have a particular ownership model or building format. AWS Outposts provides one concrete example of centrally managed cloud infrastructure deployed for local processing and lower latency. “Edge” therefore describes a deployment relationship within distributed computing, rather than establishing who owns the building or administers its servers. (ibm.com)

Modular installations use repeatable building blocks; prefabricated versions assemble and test subsystems in a factory before delivery. Modules can contain IT equipment, electrical infrastructure, cooling, or combinations of these systems. A containerized unit is a particular physical format, commonly using an ISO freight-container enclosure, whereas modular systems also include skid-mounted equipment and purpose-built enclosures. In 2014, Schneider Electric described an Angola installation comprising eight modules, six of them ISO-container modules integrated together. These categories overlap: a containerized installation can serve an enterprise edge location, while prefabricated power modules can support a hyperscale cloud campus. Neither modular construction nor a container enclosure determines the hosting business model. (blog.se.com)

Services, Workloads, and System Boundaries

Transaction processing supports operations such as purchases, bank transfers, and hotel reservations by accepting requests, updating records, and preserving data integrity when many users act concurrently. A database supplies persistent storage and retrieval; it is a component used by applications rather than a workload category confined to transactions. Online transaction processing typically involves numerous short insertions, updates, and deletions, with response times measured in milliseconds. Analytics instead examines accumulated data through operations such as aggregating sales by region and quarter. The same organization may therefore operate both transaction-oriented databases and analytical systems fed by their records, with different processing and storage-access patterns. (ibm.com)

Web services deliver application responses over networks, while content delivery distributes reusable objects such as images and video segments. These functions need not execute in the same facility: a content delivery network caches objects at geographically distributed locations, serving cache hits without requesting the object again from its origin. Amazon CloudFront, for example, distinguishes requests served from its caches from those forwarded to an origin server. Scientific computing presents another workload pattern: climate models, molecular simulations, and engineering calculations can use parallel computing across multiple machines. Tightly coupled calculations exchange intermediate results during execution, whereas loosely coupled jobs can run largely independently. These differences affect whether performance depends principally on computation, storage access, or communication between machines. (docs.aws.amazon.com)

The physical system has two interacting layers. IT equipment comprises servers, storage systems, and networking equipment that process, retain, and transport information. Supporting facility infrastructure supplies electrical power, removes heat, and maintains the physical environment; it includes power-conversion and distribution equipment, cooling plant, and associated controls. Fans and power supplies inside servers remain part of the IT equipment, even though they perform supporting functions within those machines. The US Department of Energy’s July 2024 design guide treats IT systems, environmental conditions, air management, cooling, and electrical systems together because changes in IT demand can produce secondary changes in facility energy use. A reduction in computing-equipment consumption can also reduce the cooling required to remove its heat. (ibm.com)

These boundaries matter when reporting both capacity and resource consumption. Processor count, memory, storage capacity, and network throughput describe different IT capabilities; electrical and cooling capacity describe the facility’s ability to sustain installed equipment. Neither floor area nor a server count alone establishes how much application work a site can perform. Energy accounting similarly distinguishes electricity delivered to IT equipment from total facility consumption, which also includes supporting systems and electrical losses. For illustration, an IT load averaging 1 megawatt continuously for 24 hours consumes 24 megawatt-hours; this is not the site’s total consumption when cooling and power-distribution overheads lie outside that measurement boundary. Comparisons therefore require consistent equipment coverage and measurement periods. (energy.gov)

A physical facility is also distinct from the service deployed across it. Through distributed computing, one application may use computing resources in several facilities, while a single facility may host components of many applications. In cloud computing, service geography can include abstractions larger than individual buildings: Amazon Web Services defines a Region as a separate geographic area and an Availability Zone as one or more discrete data centers. Thus, a service described as regional or multi-zone cannot be equated with one building. Its operating boundary encompasses the participating IT systems and interfacility connections, while each participating site still depends on local power, cooling, and connectivity. Replication and failover determine how the service responds when part of that infrastructure becomes unavailable. (docs.aws.amazon.com)

Site Selection and Geographic Distribution

Electricity availability constrains both the location and the opening date of a data center. A suitable site needs more than inexpensive energy: the local electrical grid must have sufficient generation, transmission, and connection capacity to accommodate a large, sustained load. In its 2025 Energy and AI report, the International Energy Agency (IEA) estimated that grid constraints could delay approximately 20% of data-center capacity planned for construction by 2030. Land availability therefore cannot be assessed independently of utility infrastructure. At Eemshaven in the Netherlands, Google identifies developable industrial land, established energy infrastructure, and a skilled workforce as reasons for selecting the location. (iea.org)

Network access imposes a separate geographic constraint. The availability, bandwidth, quality, and extension cost of telecommunications services affect whether a parcel can support its intended workloads. Land suitability also depends on buildable area rather than acreage alone: topography, soils, drainage, access roads, and construction staging influence how much of a site can actually be developed. Local land-use permissions can further constrain equipment placement and expansion. These factors were explicitly identified in the Telecommunications Industry Association’s December 2019 site-selection paper by Tom Widawsky and Tom Craft. (tiaonline.org)

Local climate and water infrastructure interact with the cooling design. Cool outdoor conditions can increase the hours during which ambient air provides cooling without mechanical refrigeration; temperature and humidity therefore matter alongside annual average temperature. Where heat rejection uses evaporative cooling, reliable water supply becomes another siting constraint, particularly in drought-prone regions. Water quality, storage, and disposal also affect feasibility. Alternative supplies can require substantial off-site works: Google reports investing €45 million in 28 kilometres of pipelines and purification systems to enable industrial-water cooling in Groningen, illustrating how a location’s suitability can depend on infrastructure built beyond the property boundary. (energy.gov)

Hazard assessment concerns both the building and its external dependencies. Floodplain exposure and drainage conditions affect site layout, while earthquakes, severe storms, fires, and extreme temperatures can threaten equipment or interrupt supporting services. Nearby hazardous-material facilities and transport corridors introduce additional risks, including loss of site access after an accident. Geographic separation alone does not establish independence: separate buildings may remain vulnerable to a shared electricity substation, water disruption, or fiber isolation. Amazon Web Services describes its Availability Zones as using separate power infrastructure and different substations, without shared generators or cooling equipment. Zones within a region may be separated by up to approximately 100 kilometres while retaining single-digit-millisecond network latency for synchronous replication. (tiaonline.org)

The resulting geography reflects a tension between concentration and proximity. Large campuses can aggregate capacity, whereas latency-sensitive applications require processing nearer users or devices; the latter requirement motivates smaller edge installations even where utility coverage is less favorable. Concentration also produces grid effects that global energy totals obscure. The IEA’s 2025 analysis projected that data centers would account for less than 10% of worldwide electricity-demand growth between 2024 and 2030 in its Base Case, but found that approximately half of US data centers under development were located in existing large clusters. Additional facilities in such clusters can encounter local connection bottlenecks even when their contribution to global demand growth remains comparatively limited. (tiaonline.org)

Building Layout and Physical Infrastructure

Equipment halls contain the racks of computing, storage, and networking equipment, while supporting spaces accommodate infrastructure with different access and maintenance requirements. Electrical rooms house switchgear and uninterruptible power supplies; mechanical plant areas contain air-handling equipment, pumps, and associated machinery. Service corridors and mechanical–electrical chases provide access and distribution routes alongside the halls. Dedicated telecommunications spaces accommodate incoming carrier connections and internal cable distribution. These functions fall within the coordinated architectural, electrical, mechanical, and telecommunications scope of ANSI/TIA-942-C, published in May 2024. (fm.com)

Within a hall, equipment racks provide repeatable mounting positions and organize equipment into rows. For example, IBM’s 42U Enterprise V2 Dynamic Rack uses the 19-inch mounting format and includes side compartments for power-distribution units and switches, preserving the main mounting space for servers. Rack dimensions alone do not determine row spacing: aisles must also accommodate technicians, open doors, equipment withdrawal, and the turning radius needed to replace cabinets. The US National Institutes of Health’s 2013 design guide explicitly identifies unobstructed passage and removal of racks from within a row as layout requirements. (ibm.com)

Structural design accounts for equipment weight, its distribution, and the loads imposed during installation and servicing. A room-wide average load does not describe every local condition: cabinet supports, equipment movement, and raised-floor components require their own assessment. The NIH guide distinguishes equipment weight, personnel and service-equipment loads, and raised-floor and cable loads. It also reserves vertical space for infrastructure: its preliminary planning allowances include 0.61 m for a raised floor, 0.305 m for lighting and fire-suppression services, and racks at least 2.13 m tall. These are dated planning examples rather than universal dimensions for every facility. (wbdg.org)

Cable pathways and air-distribution spaces interact directly. A raised access floor creates a service void that can carry electrical and network cabling and, in some designs, serve as a supply-air plenum. Overhead trays provide an alternative cable route; the floor therefore need not perform both cabling and cooling functions. The US Department of Energy’s 2024 design guide explains that congested underfloor cabling reduces airflow and disrupts delivery through perforated tiles, producing hot spots. For raised-floor installations, it recommends at least 24 inches, approximately 0.61 m, of effective clear height. Cable management consequently includes removing abandoned cables and keeping equipment air intakes, exhausts, and distribution paths unobstructed. (energy.gov)

Compartmentation limits the spread of fire between equipment and supporting areas. FM’s Data Sheet 5-32 describes fire-rated interior walls extending from the structural floor to the floor above or roof, rather than stopping at a suspended ceiling. Detection must account for rapid airflow and aisle containment: aspirating systems draw air through sampling ports, while high-sensitivity spot detectors monitor particular locations. Suppression options include sprinklers, water mist, and gaseous agents, each with different protection characteristics. Containment panels and other obstructions must therefore be coordinated with both detector placement and suppression coverage. (fm.com)

Leak detection adds a separate layer of protection where cooling water is present. IBM’s room-planning documentation identifies underfloor systems that detect and report the location of fluid leaks. Physical coordination also extends inside cabinets: IBM’s 7965-S42 cooling manifold occupies the rack’s right side and makes the corresponding power-distribution-unit pockets unavailable. Such interfaces mean that architectural clearances, pipe routes, cable pathways, power equipment, and fire-protection arrangements cannot be designed independently; their spatial relationships determine whether equipment remains accessible and protective systems remain effective. (ibm.com)

Computing Hardware and Storage

Rack-mounted servers combine processors, main memory, local drives, network interfaces, and management electronics in a chassis designed for installation and servicing within an equipment rack. The central processing units (CPUs) execute operating-system and application instructions; memory modules hold working data, while attached drives provide persistent storage. A representative configuration is Dell’s PowerEdge R760: a 2U, two-socket server whose fourth-generation Intel Xeon configuration supports up to 56 cores per processor and 32 DDR5 memory slots, with a specified maximum of 8 TB. It uses error-correcting-code memory and offers configurations containing SAS, SATA, or NVMe drives, alongside embedded or expansion-card network interfaces. These are configuration limits rather than components necessarily installed in every server. (i.dell.com)

A baseboard management controller provides a management path distinct from the host operating system. Implementations such as Dell’s iDRAC expose hardware status and remote-management functions through out-of-band interfaces, allowing administrators to inspect equipment and perform supported recovery operations without relying on the application software running normally. The controller complements, rather than replaces, the server’s main processors and application-facing network interfaces. (downloads.dell.com)

General-purpose CPUs accommodate varied instruction sequences and control-intensive work, whereas accelerators devote more resources to particular computational patterns. A graphics processing unit (GPU) achieves high throughput by executing many threads concurrently, making it suitable for sufficiently parallel numerical calculations and matrix operations used in machine learning. GPU acceleration normally remains part of a heterogeneous system: the CPU runs host code, launches device work, and coordinates data movement between host and accelerator memory. Consequently, performance depends on parallelism and data transfer as well as arithmetic capability. (docs.nvidia.com) Accelerator servers can also concentrate substantial electrical demand into one chassis. NVIDIA’s DGX H100 datasheet, for example, specifies eight H100 GPUs with 640 GB of aggregate GPU memory, dual x86 CPUs, 2 TB of system memory, and approximately 10.2 kW maximum system power. That maximum is an equipment specification, not a statement of continuous operating consumption. (nvidia.com)

Storage attachment and storage access are separate classification axes. Direct-attached storage connects drives to an individual server without an intervening storage network; networked storage makes resources available to hosts over network connections. A storage area network commonly presents block devices, whereas network-attached storage commonly presents shared files. (ibm.com) Block access exposes addressable blocks upon which a host can construct a filesystem or place database data. File access exposes named files within directories, commonly through NFS or SMB. Object access addresses data objects by identifiers through an application programming interface, with metadata accompanying the stored content. Block, file, and object therefore describe interfaces and organization, not mutually exclusive physical media: an object service can ultimately store its contents on drives inside ordinary storage servers. (ibm.com)

Storage capacity, throughput, latency, and redundancy describe different properties. Capacity measures the quantity stored; throughput measures bytes transferred per second; input/output operations per second (IOPS) counts completed operations; latency measures the time taken by an operation. Request size connects throughput and IOPS, while concurrency and queue depth influence whether a system reaches its advertised limits. Large transfers can exhaust a throughput limit before reaching the maximum IOPS, and lightly concurrent applications may achieve neither maximum. (cloud.ibm.com) Redundancy consumes additional storage through mechanisms such as mirroring or parity: usable capacity is therefore distinct from the sum of installed drive capacities. RAID levels differ in their capacity overhead, read and write behavior, and tolerance of drive failures; RAID 0 provides striping without fault-tolerant redundancy. RAID is not a backup, since keeping an array operational after a drive failure does not provide an independent historical copy of its data. (ibm.com)

Networking and External Connectivity

Data center networks carry east–west traffic between servers and storage systems and north–south traffic entering or leaving the facility. A request from an external client may trigger several internal exchanges among application servers, caches, and a database, so external bandwidth alone does not describe network demand. Internal communication supporting distributed computing can require substantial aggregate bandwidth and consistently low latency; bulk transfers place greater emphasis on sustained throughput. External service latency also includes the network path between the facility and its users. These are workload-dependent requirements rather than a fixed distinction between “fast” internal and “slow” external networks. (ibm.com)

Switches connect servers and forward traffic within the facility, while routers forward packets between IP networks and toward external destinations; a single switching platform may perform both functions. Physical connections include copper cables for suitable short links and optical fiber with optical transceivers for higher-capacity or longer-distance connections. Structured cabling organizes these links through backbone cables, distribution panels, and replaceable patch cords, separating permanent cable infrastructure from equipment connections. This permits servers and switches to be rearranged without replacing every intervening cable. (cisco.com)

A traditional hierarchical network has access switches connecting servers, aggregation switches combining access networks, and a core connecting aggregation blocks and external networks. A leaf–spine network instead connects each leaf switch to every spine switch within a two-tier fabric. Servers attach to leaves; traffic between different leaves passes through a spine, producing a consistent leaf–spine–leaf path. Layer 3 equal-cost multipath routing distributes flows among available spine paths, allowing parallel links to carry traffic rather than reserving them solely as inactive backups. Additional spines increase fabric capacity, while additional leaves provide more server-facing ports, subject to available switch ports and uplinks. (cisco.com)

Port speeds do not by themselves establish usable network capacity. Oversubscription compares aggregate server-facing bandwidth with uplink bandwidth: Cisco’s 2024 fabric-design white paper gives an example with forty-eight 25-Gbit/s downlinks and eight 100-Gbit/s uplinks, yielding $1{,}200/800 = 1.5:1$. Simultaneous traffic destined beyond that leaf can therefore exceed its uplink capacity. Redundant paths provide alternative routes after failures, but physical diversity is a separate property: two connections can share a cable route or intermediate infrastructure. Equinix, for example, explicitly distinguishes ordinary in-building cross-connects, which are not guaranteed diverse routing, from separately routed campus connections. (cisco.com)

External connectivity joins customer equipment to telecommunications carriers, private networks, cloud platforms, or Internet service providers. A cross-connect is a dedicated physical cable connection between parties within a facility; it supplies the connection, not automatically an Internet service. Internet transit provides onward connectivity through a provider, whereas peering exchanges traffic between participating networks. Peering can use a direct private cross-connect or a shared Internet exchange point switching platform. A data center is the physical facility housing equipment; an exchange is an interconnection platform that may occupy one or several facilities, and a carrier network extends beyond them. (docs.equinix.com)

Interfacility links connect separate data centers using arrangements such as leased transport services, managed metropolitan optical connections, or dark fiber operated with the customer’s transmission equipment. These links carry replication, storage transfers, and other communication between sites. Their bandwidth determines achievable transfer rates, while distance and intermediate equipment contribute latency. Synchronous storage replication illustrates the distinction: when a write requires acknowledgment from the remote system, intersite round-trip delay enters the write-completion time, even if the link has ample bandwidth. (equinix.com)

Virtualization and Resource Orchestration

A hypervisor, or virtual machine monitor, mediates access to a server’s processors, memory, storage, and network interfaces, presenting each virtual machine (VM) with virtual hardware on which a separate guest operating system runs. It schedules virtual processors on physical processors, maintains memory separation, and mediates device access. Bare-metal hypervisors operate directly on hardware, whereas hosted implementations run above a host operating system. This abstraction allows applications with different operating-system requirements to share one physical server without sharing a guest kernel. NIST’s Guide to Security for Full Virtualization Technologies, published in January 2011, distinguishes these architectures and describes the hypervisor as a central isolation boundary. (nist.gov)

Containers instead provide operating-system-level isolation: application processes have separated views of resources but ordinarily share the host kernel. On Linux, namespaces isolate resources such as process identifiers and network interfaces, while control groups account for and constrain resource consumption. Container images package applications and user-space dependencies rather than a complete guest kernel. Consequently, containers can avoid the overhead of booting and maintaining a separate operating system for every application instance, but a shared-kernel vulnerability can affect multiple containers. NIST’s September 2017 Application Container Security Guide also describes containers running inside VMs, combining container-based deployment with a separate virtualization boundary between groups of workloads. (nvlpubs.nist.gov)

Resource orchestration coordinates these execution environments across a pool of servers. Provisioning translates a request into operations such as selecting an image, reserving resources, creating an instance, attaching storage, and connecting network interfaces. Placement matches requirements against available capacity and policy constraints; pooling does not make physically incompatible or exhausted resources interchangeable. OpenStack illustrates this division of responsibility: Nova provisions compute instances, Placement tracks resource inventories and allocation candidates, Glance supplies images, and Neutron provisions network connectivity. Cinder provides persistent virtual block devices whose lifetimes are independent of individual compute instances, allowing an application’s storage allocation to survive replacement of its VM. (docs.openstack.org)

In container orchestration, Kubernetes schedules Pods—groups of one or more containers—onto worker nodes. CPU and memory requests inform placement, while limits constrain execution: CPU limits are enforced through throttling, whereas memory limits can result in out-of-memory termination. One Kubernetes CPU unit represents one physical core or one virtual core, depending on the node; a request of 500m therefore denotes half a CPU unit, not a fixed processor frequency. Requests and limits distinguish capacity promised for scheduling from permitted consumption. They also help control resource contention between applications sharing a node, although they do not themselves establish complete security isolation between tenants. (kubernetes.io)

Software-defined infrastructure extends programmatic control across compute, storage, and networking. Compute services expose instance creation through APIs; storage services expose logical volumes independently of a particular server; virtual switches and network overlays provide logical connectivity over physical links. These abstractions allow provisioning workflows to coordinate several resource types without manually configuring every device for each workload. They also relocate important security boundaries into software: traffic between VMs on the same host may remain within a virtual switch rather than pass through an external physical firewall. NIST SP 800-125B, published in March 2016, consequently treats virtual-network segmentation, path redundancy, firewall-based traffic control, and VM traffic monitoring as distinct configuration concerns. (docs.openstack.org)

Management systems possess privileges beyond those of ordinary workloads: depending on their role, they can create or delete guests, change virtual connectivity, and access stored images or snapshots. Compromise of this management plane can therefore cross otherwise separate application or tenant boundaries. Hypervisor isolation, administrative authorization, and virtual-network policy address different parts of that exposure; a functioning VM boundary does not compensate for an administrator granting access to the wrong tenant. NIST’s virtualization guidance identifies restricted management access, protected management communications, and separation of administrative responsibilities as controls specific to this privileged infrastructure layer. (nvlpubs.nist.gov)

Electrical Supply and Power Distribution

A conventional electrical path begins at the utility connection to the electrical grid, passes through transformers and switchgear, and continues through uninterruptible power supplies (UPS), distribution panels or busways, and rack power distribution units to IT equipment. Transformers reduce the supply voltage; switchgear provides switching, isolation, and fault protection. The precise order and number of voltage transformations vary with site design. Rack distribution supplies servers, storage, and network equipment, whose power supplies convert incoming alternating current into the direct current required by electronics. Cooling and other mechanical loads also require electrical capacity, although they need not all follow the same UPS-backed path as IT loads. (energy.gov)

An online, double-conversion UPS rectifies incoming AC to DC and uses an inverter to produce AC for the protected load. A connected battery supplies the DC link when the input fails, allowing the inverter to continue operating without a transfer interruption. This arrangement also conditions voltage and frequency during normal operation. Battery storage provides ride-through, rather than an unlimited substitute for the utility: its duration depends on usable stored energy and connected demand. As an illustrative energy balance, a usable 100 kWh store supplying 1 MW would last six minutes before conversion losses and operating reserves are considered. Increasing battery energy capacity extends runtime; increasing the inverter’s power rating alone does not. (origin-edge.eaton.com)

Standby generators provide longer-duration outage support. Following a utility failure, the UPS carries protected loads while generators start and reach acceptable voltage and frequency. An automatic transfer switch monitors the sources and connects the emergency supply once it is stable; it later transfers the load back when normal supply is restored under the controller’s specified conditions. The transfer switch selects a source but does not itself store energy, so it does not replace the UPS’s bridging function. Generator-backed operation remains limited by fuel availability, equipment condition, and the ability to supply all essential supporting loads, rather than merely the server circuits. (eaton.com)

Conversion and distribution consume part of the incoming electricity. The U.S. Department of Energy’s July 2024 design guide reported double-conversion UPS efficiencies of 95% or higher in 2023, compared with approximately 85–90% in the 1990s. At 95% efficiency, delivering 1 MW requires approximately 1.053 MW at the UPS input, leaving about 53 kW as losses. Transformer losses and conductor heating add further demand. Electrical redundancy also affects loading: parallel UPS units sharing a load may operate well below their individual ratings, changing their efficiency. Extra installed equipment therefore does not imply proportionately greater useful output or electricity consumption. (energy.gov)

Installed capacity, operating demand, and energy consumption describe different quantities. Capacity specifies a rated capability, commonly in MW for real power or MVA for apparent power; operating demand is the power actually drawn at a particular time. Annual energy is accumulated demand, $E=\int P(t)\,dt$. Thus, a hypothetical facility averaging 10 MW throughout a 365-day year consumes 87.6 GWh, irrespective of whether its installed capacity is substantially larger. Capacity figures must also identify whether they describe IT loads, the entire facility, or redundant equipment totals. New capacity cannot necessarily be energized when construction finishes: utility studies, substations, transmission reinforcement, and equipment procurement can govern the connection schedule. In its 2025 Energy and AI report, the International Energy Agency estimated that grid constraints could delay around 20% of global data-center capacity planned for construction by 2030; this was a conditional assessment, not an observed delay rate. (energy.gov)

Cooling and Thermal Management

Nearly all electrical energy consumed by IT equipment ultimately becomes heat, so a rack drawing 20 kW produces approximately 20 kW of thermal load. For sensible cooling, heat removal follows $\dot Q=\dot m c_p\Delta T$, where $\dot m$ is coolant mass flow, $c_p$ its specific heat capacity, and $\Delta T$ its temperature rise. Increasing rack power therefore requires greater coolant flow, a larger temperature rise, or a different cooling arrangement. Concentrated loads can exceed local cooling capacity even when the room’s total capacity appears sufficient; equipment nameplate power is not necessarily its actual heat output. (handbook.ashrae.org)

Air-cooled racks commonly alternate front-to-front cold aisles and back-to-back hot aisles. Cold aisles supply equipment intakes; hot aisles collect exhaust. Doors, roofs, and partitions contain one airflow stream, while blanking panels close unused rack positions, limiting exhaust recirculation and supply-air bypass. Computer-room air handlers (CRAHs) use fans to pass return air across chilled-water coils before supplying it to equipment. The chilled-water loop carries heat to a chiller or an economizer; in a water-cooled chiller plant, a separate condenser-water circuit carries heat onward to cooling towers. These arrangements transport heat through successive heat exchangers rather than simply “producing cold.” (energy.gov)

Economization reduces compressor operation when outdoor conditions permit. Air-side economization introduces suitably filtered outdoor air and exhausts warm indoor air, subject to humidity and contamination limits. Water-side economization uses a heat exchanger to transfer heat from the chilled-water circuit to a cooler tower-water circuit, partly or entirely bypassing the chiller. Its availability depends on outdoor conditions and the temperatures required by the equipment. “Free cooling” therefore means reduced refrigeration work, not zero energy use: fans, pumps, and controls still operate. (energy.gov)

Liquid-based approaches capture heat at different points. Rear-door heat exchangers place liquid-cooled coils in rack exhaust paths, retaining air cooling inside servers. Direct-to-chip systems circulate coolant through cold plates attached to CPUs, GPUs, or other components; remaining components may still require airflow. Immersion systems instead surround electronics with electrically nonconductive fluid. Single-phase immersion transfers heat without boiling, whereas two-phase immersion boils the fluid and condenses its vapor. (energy.gov) A coolant distribution unit (CDU) commonly separates the equipment coolant loop from facility water through a heat exchanger and provides pumping, temperature regulation, and monitoring. Manifolds, hoses, and quick-disconnect fittings distribute coolant to racks and servers; collected heat still requires outdoor rejection or another heat sink. (handbook.ashrae.org)

Environmental limits concern conditions at equipment inlets, not merely a room thermostat. ASHRAE’s 2021 guidelines specify a recommended inlet-air temperature range of 18–27 °C for classes A1–A4; allowable operating envelopes are wider and class-dependent, and are distinct from recommended conditions. (energy.gov) Moisture control addresses corrosion, electrostatic discharge, and condensation. Condensation can form when exposed coolant surfaces fall below the surrounding air’s dew point. ASHRAE’s Water-Cooled Servers guidance describes maintaining equipment-loop supply water at least 2 °C above room dew point, providing a margin for sensor uncertainty; a CDU can regulate this temperature independently of colder facility water. (ashrae.org)

Internal coolant circulation is distinct from water consumption during heat rejection. A closed equipment loop repeatedly transports the same fluid and does not inherently require evaporation. An associated cooling tower, however, rejects heat through evaporation and requires makeup water; additional water leaves through blowdown used to control dissolved minerals. Dry coolers reject heat to outdoor air without that evaporative process, although their performance depends on ambient temperature and the required coolant temperature. Consequently, neither “air-cooled servers” nor “liquid-cooled servers” alone establishes a facility’s water requirements: the downstream heat-rejection system is also decisive. (energy.gov)

Availability, Redundancy, and Disaster Recovery

A single point of failure is a component or dependency whose loss can interrupt the supported service despite other equipment remaining operational. Two servers connected to the same power distribution unit, for example, remain vulnerable to that unit’s failure. A failure domain is the set of resources potentially affected by a particular failure: it may encompass a rack, an electrical distribution path, a cooling system, or an entire site. Fault isolation limits propagation beyond these boundaries. Common-mode failures defeat apparent independence when redundant systems share an exposure, such as flooding, a common controller, or the same defective software update. AWS’s fault-isolation documentation describes both separation of generators and cooling equipment between Availability Zones and staggered software deployments to reduce correlated failures. (docs.aws.amazon.com)

Redundancy notation describes capacity relative to the load. N denotes the capacity required to support the design load; N+1 adds one capacity component beyond that requirement. In an illustrative installation requiring four equally rated 250 kW uninterruptible power supply modules to support 1 MW, five modules provide N+1 capacity: one can become unavailable while four continue supplying the load. 2N provides two complete capacity systems, each capable of supporting the full load. These arrangements describe different protection levels, not interchangeable counts of spare machines. N+1 module redundancy does not necessarily duplicate downstream switchgear or distribution paths, and increasing demand beyond the remaining modules’ capacity removes the intended redundancy. (product-help.schneider-electric.com)

Equipment redundancy is distinct from concurrent maintainability and fault tolerance. Concurrent maintainability means that each required capacity or distribution component can be deliberately isolated for maintenance without interrupting IT operation. Fault tolerance additionally addresses unplanned failures within the specified design scope, preventing their effects from interrupting the supported IT load. An extra pump cannot make a cooling installation concurrently maintainable if servicing a shared pipe requires shutting down every pump; similarly, duplicated electrical equipment cannot isolate a fault that propagates through a common connection. Uptime Institute therefore evaluates system outcomes and topology rather than assigning a classification from component counts: both N+1 and 2N arrangements can support concurrently maintainable designs. (journal.uptimeinstitute.com)

Site availability concerns the continuity of facility services, especially power and cooling; application availability concerns whether users can successfully access a functioning service. An application can fail while its building remains operational because a database, network dependency, or software component is unavailable. Conversely, an application distributed across independent locations can continue serving users after one site fails. Failover transfers processing or traffic to surviving resources, but its effectiveness depends on failure detection, accessible data, sufficient surviving capacity, and functioning dependencies. AWS’s published controls illustrate this distinction by describing both facility redundancy and the distribution of core applications across sites with enough remaining capacity to absorb a site failure. (docs.aws.amazon.com)

Disaster recovery addresses restoration after disruptions that exceed normal continuity mechanisms. Backups retain recoverable historical states; replication maintains additional copies of changing data; failover activates an alternative processing environment. Replication can shorten interruption and reduce lost updates, but it can also reproduce accidental deletion or logical corruption. Historical backups or point-in-time recovery therefore address consequences that a current replica alone may not resolve. Recovery also requires application code, configuration, and supporting infrastructure, not merely copied user data. AWS’s disaster-recovery documentation explicitly separates backup-and-restore, partially provisioned standby environments, and fully operating multisite arrangements by their recovery characteristics. (docs.aws.amazon.com)

Two objectives distinguish restoration speed from data loss. The recovery time objective (RTO) specifies the target recovery interval; the recovery point objective (RPO) specifies the point in time to which data must be recoverable, commonly expressed as an acceptable data-loss interval. For example, a service with a two-hour RTO and a 15-minute RPO has separate targets for restoring operation and recovering recent records: rapid restart does not compensate for losing several hours of transactions. NIST’s Contingency Planning Guide for Federal Information Systems, published in 2010 by Marianne Swanson and colleagues, places these objectives within business-impact analysis and recovery planning, with exercises and recovery testing used to evaluate the implemented capability. (csrc.nist.gov)

Standards, Ratings, and Certification

Uptime Institute’s Tier classification system describes four levels of site-infrastructure capability, with each successive level incorporating the preceding requirements. Tier I—Basic Capacity provides dedicated IT space, an uninterruptible power supply, continuous cooling, and an engine generator for extended utility outages. Tier II—Redundant Capacity Components adds spare power and cooling components, such as pumps or UPS modules, but does not establish concurrent maintainability throughout the distribution system. Tier III—Concurrently Maintainable requires that every capacity component and distribution element needed to support the IT environment can be removed from service for planned maintenance without interrupting IT operation. Tier IV—Fault Tolerant adds the ability to withstand an individual equipment failure or distribution-path interruption without affecting operation. These are performance outcomes rather than mandatory equipment recipes: a redundant component count alone does not establish a Tier level. (journal.uptimeinstitute.com)

ANSI/TIA-942, developed by the Telecommunications Industry Association, addresses a broader set of physical-infrastructure requirements than its title, Telecommunications Infrastructure Standard for Data Centers, might suggest. Its scope includes architecture, telecommunications, power, cooling, fire protection, physical security, safety, and monitoring. Revision C, dated May 2024 and announced on May 9, 2024, provides a dated example of this scope. Changes included incorporation of edge-data-center material, recognition of additional cable media and connectivity requirements, revised provisions for floor loading and minimum cabinet width, and changes to annexes concerning site selection and space. It applies across enterprise, colocation, and hyperscale facilities rather than defining a separate standard for each ownership model. (tiaonline.org)

TIA’s four ratings are Rated-1—Basic Site Infrastructure, Rated-2—Redundant Capacity Component Site Infrastructure, Rated-3—Concurrently Maintainable Site Infrastructure, and Rated-4—Fault Tolerant Site Infrastructure. Although their names resemble Uptime’s categories, TIA ratings and Uptime Tiers are separate classifications with different criteria and certification arrangements; an award under one does not constitute certification under the other. TIA’s program uses TIA-licensed certification bodies, whereas Uptime Institute certifies facilities against its own Tier Standard. A standard specifies requirements; a rating identifies the assessed performance category; certification records conformity established through the relevant assessment process. Neither a numerical designation nor a statement that a facility was “designed to” a category substitutes for that assessment. (tiaonline.org)

Certification also distinguishes what was examined. Uptime’s Tier Certification of Design Documents evaluates the documented design, while Tier Certification of Constructed Facility examines installed infrastructure and includes demonstrations of the required functionality. For example, Tier III demonstrations must establish that baseline power and cooling capacity remain available in every required maintenance configuration—not merely that spare equipment appears on drawings. Its subsequent Tier Certification of Operational Sustainability evaluates management practices, staffing, training, operating conditions, and maintenance. TIA likewise distinguishes document-based design certification from facilities certification involving on-site inspection. These voluntary assessment programs therefore provide different evidence at different lifecycle stages: design conformity, implemented capability, and operational practice are related but not interchangeable findings. (uptimeinstitute.com)

A facility classification is not a universal prediction of application uptime. Uptime Institute removed references to “expected downtime per year” from its Tier Standard in 2009, and its topology standard does not assign availability predictions to Tier levels. Converting a Tier designation into a fixed annual outage allowance consequently misrepresents its meaning. The classification concerns site infrastructure; software configuration, IT-system architecture, and operational practices introduce additional failure mechanisms. Even infrastructure meeting the specified topology outcomes must be operated and maintained appropriately, which is why operational evaluation is distinct from approval of drawings or verification of construction. (journal.uptimeinstitute.com)

Physical Security, Cybersecurity, and Data Protection

Physical protection uses successive access boundaries rather than treating admission to the building as permission to reach every server. Perimeter fencing, guarded entrances, intrusion detection, and surveillance control access to the site; authenticated doors then restrict entry to equipment areas. Permissions can be limited by role, location, and duration, with separate authorization for visitors and contractors. AWS, for example, documents time-bound access requests requiring a business justification, restrictions to approved areas, retained access logs, and camera coverage of server-room access points. These records support investigation by connecting an individual’s authorization with recorded entry events; they do not replace preventive barriers. (aws.amazon.com)

Cybersecurity also depends on protecting the management plane: the interfaces through which administrators configure hosts, networks, storage, and virtual machines. These interfaces have greater authority than ordinary application connections. NIST’s January 2011 Guide to Security for Full Virtualization Technologies, authored by Karen Scarfone, Murugiah Souppaya, and Paul Hoffman, describes restricted administrative access, dedicated management networks, and protected management communications. Firewall rules limit which systems can reach administrative interfaces, while role-based privileges distinguish observation from configuration authority. Hypervisor and guest-system patching address different software layers; updating one does not update the other. A compromised hypervisor is particularly consequential because it mediates access to resources shared by several virtual machines. (csrc.nist.gov)

Tenant isolation requires controls within shared hardware as well as at the external network boundary. A hypervisor separates virtual-machine execution and mediates access to processor, memory, storage, and networking resources. NIST’s June 2018 guidance on server-based hypervisor platforms identifies runtime isolation and security-preserving communications as explicit hypervisor responsibilities. Network segmentation and traffic filtering further restrict communication between workloads. Monitoring must also account for traffic that remains inside a host: two virtual machines connected through a virtual switch can communicate without passing through a physical network sensor. Virtual-switch monitoring or other appropriately placed controls are therefore necessary where inspection of that traffic is required. (csrc.nist.gov)

Facility security and application-data protection have different responsibility boundaries. In cloud computing, AWS’s shared-responsibility model assigns the provider responsibility for physical facilities, hardware, and the underlying virtualization infrastructure, while an infrastructure-as-a-service customer manages the guest operating system, application software, and relevant access and firewall settings. Managed services shift additional operational tasks to the provider, so the boundary varies with the service. Encryption protects data in transit and at rest, but its configuration, key management, and access permissions remain distinct controls. An encrypted storage service does not by itself establish that every user or application granted access is authorized to receive the plaintext. (aws.amazon.com)

Data-bearing equipment remains within the security boundary during repair, transfer, and retirement. Asset identification, restricted handling, and documented transfers establish a chain of custody; sanitization records connect each device with the action performed and its outcome. NIST SP 800-88 Revision 2, published in September 2025, distinguishes clear, which protects against recovery through ordinary interfaces; purge, which makes recovery infeasible using advanced laboratory techniques while potentially preserving reuse; and destroy, which additionally makes the medium unusable for storage. Ordinary overwriting cannot reliably reach every previously used location in flash devices with spare cells and wear levelling. Cryptographic erasure instead sanitizes the keys needed to recover encrypted data, subject to conditions concerning encryption coverage and key handling. Verification checks whether the operation completed successfully; validation determines whether its outcome is sufficient for the data’s confidentiality requirements. (nvlpubs.nist.gov)

Operations, Monitoring, and Workforce

Data center operations divide responsibility among interdependent teams. Facility engineers operate and maintain electrical distribution, backup power, cooling, and building controls; IT operators manage server and storage health, operating systems, and workload execution. Network staff maintain switching, routing, connectivity, and network configurations. Security personnel oversee physical access and surveillance, while on-site technicians install equipment, replace failed components, inspect cabling, and perform physical tasks requested by remote operators. These boundaries vary by organization, but coordination is essential: replacing a server, for example, involves identifying the correct asset, authorizing the intervention, disconnecting the appropriate connections, and verifying restored service. Google describes monitoring and controls spanning servers, storage, networking, electrical distribution, mechanical cooling, and security, with site engineering teams working alongside operations personnel. (ibm.com)

Monitoring connects physical conditions with equipment and service status. Building-management systems supervise mechanical plant and environmental controls; electrical monitoring follows supply status, equipment loading, and power-system alarms. IT-management tools observe server health, storage, network performance, and application behavior. Data center infrastructure management (DCIM) connects these domains by associating equipment inventories and rack locations with power, cooling, space, and device-health information. The resulting view helps distinguish, for example, a hardware fault from a shared infrastructure problem affecting several machines. A 2015 paper by Matthew Hubbell and colleagues described a monitoring implementation combining approximately 5,000 environmental sensors with more than 3,500 IT data points, enabling visualization from a command center over 100 miles away. Such integration supplements rather than makes identical the functions of facility controls and IT monitoring. (ibm.com)

Preventive maintenance translates equipment requirements into scheduled inspections, servicing, testing, and recorded completion. Uptime Institute’s operational criteria require fully scripted maintenance processes, rather than reliance on technicians’ memory. Documentation distinguishes standard operating procedures for established activities, methods of procedure for particular interventions, and emergency operating procedures for abnormal conditions. A method of procedure specifies the sequence of actions that changes the state of critical equipment; it operates within change control, which coordinates authorization, timing, affected systems, and verification. This matters when maintenance temporarily removes redundancy: an otherwise routine intervention can expose the remaining equipment to a failure that the normal configuration could withstand. (uptimeinstitute.com)

Incident response adds explicit coordination when normal operations break down. Google’s published incident-management guidance separates command, operational response, and communications so that technical recovery does not depend on every responder independently choosing priorities. After recovery, a post-incident review records impact, chronology, contributing conditions, and corrective actions, including improvements to detection and coordination—not merely replacement of the failed component. Google’s published postmortem examples include an August 11, 2014 incident in which faulty retirement automation selected all machines in a satellite environment rather than one rack for disk erasure. The example illustrates how an operational tool can enlarge an intervention’s scope and why reviews examine automation and safeguards as well as individual actions. (sre.google)

Staffing determines whether monitoring, maintenance, and emergency procedures can actually be executed. Coverage includes routine operating work, planned maintenance, and unexpected interventions; training requirements increase with site complexity and availability objectives. Uptime Institute’s 2024 resiliency survey found that 39% of 418 respondents reported a significant, serious, or severe IT-service outage involving human error during the preceding three years. Among respondents identifying causes of such outages, failure to follow procedures was selected by 48%, and incorrect processes or procedures by 45%; respondents could select up to three causes. These are survey responses, not proportions of all outages, but they distinguish two operational weaknesses: inadequately defined work and failure to execute correctly defined work. (journal.uptimeinstitute.com)

Capacity Planning, Construction, and Asset Lifecycle

Capacity planning translates forecast workloads into a time-dependent infrastructure requirement. Transaction rates, concurrent users, storage growth, and processing deadlines establish the computational demand; workload measurements and hardware benchmarks then inform processor, memory, storage-throughput, and network requirements. The resulting equipment configuration determines rack occupancy and the electrical and cooling loads that the facility must support. Planning distinguishes the initial installation from expected growth and ultimate capacity, rather than treating the building’s maximum capacity as an immediate requirement. Schneider Electric’s planning framework, published in 2011, explicitly separates initial IT load, minimum and maximum final loads, and the anticipated ramp-up period. (blog.se.com)

These requirements must remain consistent across physical systems: spare rack positions do not represent usable capacity if their power circuits, cooling provision, or network connections are already exhausted. As an illustrative calculation, 100 racks operating at 10 kW each impose a 1 MW IT load, before cooling equipment and electrical losses are included; concentrating the same load into 25 racks raises the average to 40 kW per rack without reducing the aggregate heat-removal requirement. Consequently, capacity plans specify both total load and its spatial distribution. The US Department of Energy’s July 2024 design guide emphasizes that IT equipment choices affect the size of supporting systems and that higher-density equipment can require direct liquid cooling rather than simply additional floor space. (blog.se.com)

Development proceeds from site acquisition and confirmation of service availability through concept design, detailed engineering, procurement, construction, testing, and operational handover. Commissioning begins with requirements and design review, rather than only after construction: it verifies that installed equipment and interacting systems perform as intended. Tests progress from individual components to coordinated operation, using load banks to reproduce electrical demand and heat release before production equipment is fully deployed. Simulated failures reveal how power, cooling, and controls respond together. Handover includes test records, operating documentation, training, and resolution of outstanding defects. Phased construction can bring successive halls or infrastructure blocks into service as demand develops; each addition requires testing of its interfaces with the operating installation. (journal.uptimeinstitute.com)

The building and fixed service infrastructure commonly accommodate several generations of IT equipment. Hardware refreshes therefore create a compatibility problem as well as a replacement task: a hall with sufficient aggregate electrical capacity may lack suitable rack-level distribution, airflow, pipework, or heat-rejection arrangements for a denser installation. Retrofitting can require rearranging racks, replacing distribution equipment, or introducing coolant distribution systems while adjacent equipment remains operational. Replacement intervals are not universal technical limits. In July 2022, Microsoft reassessed its server and network equipment and increased their estimated accounting useful lives from four to six years, effective in fiscal 2023, citing software efficiencies and technological advances; this did not establish a six-year physical failure threshold. (energy.gov)

Asset management follows equipment beyond its first deployment. Inventory records connect equipment identity, location, configuration, and service status with subsequent transfers, testing, repair, and retirement. Serviceable equipment may be redeployed, while components harvested from retired servers can supply repairs elsewhere in the fleet. Microsoft’s April 2025 circular-hardware white paper describes tracking components through shipping, receiving, testing, storage, and disposition, with recertified parts returned to active service. Its Circular Centers, first established in Amsterdam in 2020, supported the reported reuse of more than 3.2 million components in 2024. Final decommissioning routes remaining equipment and materials into reuse or recycling streams, with traceability maintained through external recovery partners; controls for data-bearing media form a separate security requirement. (datacenters.microsoft.com)

Economics and Infrastructure Supply Chains

Data-center investment separates capital expenditure on land, buildings, electrical and mechanical infrastructure, and IT equipment from recurring operating expenditure. The facility budget includes construction, transformers, switchgear, backup power, cooling plant, and installation; the IT budget includes servers, storage, and networking equipment. Operating costs include electricity, staff, maintenance contracts, replacement parts, connectivity, and software subscriptions. These categories have different replacement schedules and cost drivers: a completed building does not eliminate the need to replace servers or maintain its supporting plant. An IBM Research study published in 2009 distinguished building construction, power-and-cooling infrastructure, server acquisition, electricity, and server management as five principal components of computing cost. (research.ibm.com)

Utilization determines how much productive work shares the cost of installed capacity. Purchasing equipment for occasional peaks leaves capital committed during quieter periods; pooling workloads with different demand patterns can spread that cost across more activity. Economies of scale also arise through bulk hardware purchasing and the distribution of procurement, engineering, and operational expertise across a larger equipment base. These effects are distinct: a large installation can obtain purchasing discounts while still containing underused servers. Redundancy adds another economic layer, because spare equipment, duplicate connections, and recovery capacity require investment and maintenance without necessarily increasing ordinary production throughput. Comparisons therefore depend on both workload utilization and the availability requirements being met. (d0.awsstatic.com)

Electricity links operating expenditure to equipment loading and facility efficiency. Both computing equipment and supporting cooling and electrical systems consume energy, so the electricity bill cannot be inferred from server ratings alone. As an illustrative calculation rather than a reported facility measurement, a constant total facility demand of 10 MW uses 87.6 million kWh in a 365-day year; at an assumed flat energy price of USD 0.10/kWh, the energy charge would be USD 8.76 million. This excludes demand charges and other tariff components. Staffing and maintenance remain separate costs: continuous operation requires monitoring, technical coverage, servicing, and repairs even when application demand falls. (energy.gov)

Ownership models redistribute these obligations rather than remove them. An enterprise operating its own facility finances and manages both the building infrastructure and IT equipment. In conventional colocation, the provider supplies the premises and supporting infrastructure while the customer owns and manages its servers, storage, and network equipment; additional management services may be contracted separately. Managed hosting transfers dedicated hardware provision and specified administrative tasks to the service provider. With cloud computing, customers purchase computing services rather than construct the underlying facility, while providers recover infrastructure and operating costs through service charges. These arrangements change the customer’s upfront investment, recurring payments, and responsibilities, but do not make the physical resources costless. (ibm.com)

Expansion depends on supply chains and construction schedules as well as demand for computing. Servers require an operational building, cooling, electrical distribution, and an energized utility connection before they can deliver service. Equipment availability can constrain that sequence: the International Energy Agency’s World Energy Investment 2025 reported procurement times of two to three years for transmission cables and up to four years for large power transformers. These are power-system equipment lead times, not universal delivery times for every data-center component. Separately, the IEA’s 2025 Energy and AI analysis estimated that grid constraints could delay around 20% of global data-center capacity planned for construction by 2030. That figure described a prospective connection risk, not an observed cancellation rate; strong customer demand and available financing do not by themselves establish when new capacity can begin operating. (iea.blob.core.windows.net)

Energy, Water, and Environmental Accounting

Power usage effectiveness (PUE) measures facility overhead relative to IT equipment energy over the same reporting period:

The numerator includes IT equipment, cooling, electrical conversion and distribution losses, and other supporting loads within the facility boundary; the denominator covers servers, storage, and networking equipment. Comparisons require consistent meter locations, reporting periods, and treatment of shared infrastructure. A hypothetical annual total of 12 GWh supporting 10 GWh of IT consumption gives a PUE of 1.2. PUE does not measure useful computation: its denominator is energy, not completed transactions or scientific calculations. It therefore needs complementary measures of equipment utilization and computational output per unit of energy. (energy.gov)

In its April 2025 report Energy and AI, the International Energy Agency estimated that data centers consumed 415 TWh of electricity in 2024, approximately 1.5% of global electricity consumption. This was an estimate of past consumption, not a complete worldwide metered census. The report separately projected about 945 TWh in 2030 under its Base Case. That projection was conditional on assumptions about equipment deployment, utilization, efficiency, and demand; it was not an observed outcome or an unconditional forecast. Global totals also obscure concentration: the report estimated that the United States accounted for 45% of data-center electricity consumption in 2024, China 25%, and Europe 15%. (iea.org)

Water accounting distinguishes withdrawal, the volume taken from a source, from consumption, the portion not returned because of processes such as evaporation. Cooling-tower evaporation is consumption, whereas discharged water is not automatically counted as consumed. Reports must identify which quantity their “water use” represents. Site-based water usage effectiveness (WUE) divides annual site water use by annual IT equipment energy, commonly in liters per kilowatt-hour. A source-based boundary additionally includes water associated with electricity production. Thus, low on-site water use does not establish a low total water footprint; electricity-related impacts occur outside the facility and depend on its power supply. (usgs.gov)

Carbon inventories similarly depend on organizational and lifecycle boundaries. Operational emissions include on-site fuel combustion and emissions associated with purchased electricity; embodied emissions arise from producing construction materials and manufacturing equipment. Under the GHG Protocol’s 2015 Scope 2 Guidance, location-based electricity accounting uses average emissions factors for the relevant geographic area, while market-based accounting reflects qualifying contractual instruments, including electricity contracts and energy-attribute certificates. These answer different questions: a contractual renewable-energy claim is not a measurement of the physical generation supplying a facility at each hour. Carbon usage effectiveness expresses operational emissions relative to IT energy, but does not by itself describe embodied impacts. (ghgprotocol.org)

Cooling and heat recovery introduce trade-offs between these accounts. Evaporative heat rejection consumes water but can reduce electricity requirements; dry cooling avoids that evaporation, while its energy performance depends on climate and operating temperatures. Waste heat can supply nearby heating loads, with higher coolant-return temperatures increasing opportunities for direct use. A heat pump can raise heat to a more useful temperature but adds electricity consumption. The Department of Energy’s 2024 design guide treats recovered energy separately from PUE, using energy reuse effectiveness rather than subtracting exported heat from the conventional PUE numerator. (energy.gov)

A life-cycle assessment also considers hardware manufacture, reuse, and end-of-life processing rather than equating environmental performance with operating efficiency alone. Microsoft’s April 2025 circular-hardware documentation, for example, describes routing decommissioned servers and components through Circular Centers for internal reuse, other electronics supply chains, or technician-training programs. It distinguishes these routes from material recycling and identifies repair using harvested components as a tracked activity. Such accounting makes the destination and continued use of retired hardware visible alongside electricity and water indicators. (datacenters.microsoft.com)

AI and High-Performance Computing Infrastructure

Distributed machine learning training couples accelerators through repeated computation and communication. In synchronous data-parallel training, workers process different portions of a batch and combine parameter gradients before updating their model replicas; PyTorch’s DistributedDataParallel, for example, groups gradients into buckets and overlaps their all-reduce communication with backward computation. Consequently, accelerator count alone does not determine training speed: communication delays can leave otherwise available processors waiting. A tightly coupled job also needs its participating processes to run together, making coordinated resource allocation important rather than treating each accelerator as an independently scheduled server. These requirements connect AI training to the scheduling and communication problems of parallel computing. (docs.pytorch.org)

Inference executes an already trained model, but its infrastructure requirements depend on how that model is served. For autoregressive large language models, prefill processes the input prompt, whereas decoding generates successive output tokens. Long prefills are commonly compute-intensive; decoding can instead be constrained by memory bandwidth and the capacity occupied by model weights and cached attention state. Interactive serving therefore balances concurrent requests, memory allocation, and response latency rather than simply minimizing the completion time of one training job. NVIDIA Dynamo illustrates an approach in which separate accelerator pools handle prefill and decoding, transferring the attention key–value cache between them. This introduces network traffic of its own, and separating the pools is not universally advantageous: low concurrency, transfer overhead, and uneven pool utilization can favor combined execution. (docs.nvidia.com)

An accelerator cluster combines GPUs or other specialized processors with interconnects and storage that sustain their collective operation. NVIDIA’s DGX GB200 reference architecture, for example, places 72 GPUs in one rack-scale NVLink domain and uses InfiniBand for communication between nodes and racks. Its management architecture distinguishes the compute fabric from a dedicated storage fabric. Shared storage supplies datasets and checkpoints, while local NVMe devices can cache or stage data to reduce repeated network reads. Checkpointing creates a different bottleneck from dataset ingestion: snapshots can reach terabytes, and synchronous checkpoint writes pause training progress. Thus, aggregate storage capacity, read throughput, checkpoint-write performance, and interconnect bandwidth are separate constraints on useful accelerator utilization. (docs.nvidia.com)

Concentrating this equipment changes the unit of facility planning from an individual server to an integrated rack and its supporting infrastructure. In its October 2024 description of the GB200 NVL72 design, NVIDIA specified a rack cooling requirement of 120 kW and direct liquid cooling; the DGX implementation combines liquid cooling with residual air cooling. This is a product-specific example, not a universal AI-rack rating. Such deployments require electrical distribution and heat-removal capacity at the rack’s location, not merely sufficient unused capacity elsewhere in the building. The reference architecture accordingly coordinates compute racks, power shelves, cooling, networking, and facility-management integration. Deployment planning must accommodate these interacting systems rather than assume that new accelerators can simply replace lower-density servers in existing rack positions. (developer.nvidia.com)

Future electricity requirements depend on both workload growth and the resources needed per completed task. Hardware and software efficiency, utilization, model adoption, and the location and pace of deployment are therefore scenario assumptions rather than fixed outcomes. The International Energy Agency’s 2025 Energy and AI analysis illustrates this sensitivity: its 2035 projections for all data centers, not AI alone, ranged from approximately 700 TWh in the Headwinds Case to more than 1,700 TWh in the Lift-Off Case. Its High Efficiency Case reached about 970 TWh while meeting the same service demand as the Base Case. These differences reflect alternative assumptions about efficiency, demand, supply chains, and deployment constraints; they are not measurements of future consumption. (iea.org)

References

1. February 14: John Mauchly and J. Presper Eckert Unveil The ENIACcomputerhistory.org
2. What Is a Data Center? | IBMibm.com
3. The S/360: A turning point in mainframe historyibm.com
4. The IBM System/360 | IBMibm.com
5. Time-sharing | IBMibm.com
6. 20 Years of Being Customer-Inspired and Customer-Drivenblog.equinix.com
7. IBM: VM 50th Anniversaryibm.com
8. VMware Celebrates Seven Years of Continual Innovation and Executionnews.broadcom.com
9. Amazon-Web-Services-Launchespress.aboutamazon.com
10. Amazon EC2 Beta | AWS News Blogaws.amazon.com
11. Final Version of NIST Cloud Computing Definition Publishednist.gov
12. What is Colocation? Data Center Comparison & FAQs | Equinixequinix.com
13. Colocation Products and Services | Equinix Product Documentationdocs.equinix.com
14. Shared Responsibility Model - Amazon Web Services (AWS)aws.amazon.com
15. What is AWS Outposts? - AWS Outpostsdocs.aws.amazon.com
16. What is a hyperscale data center? | IBMibm.com
17. Hyperscale vs Colocation Data Centers | IBMibm.com
18. OLAP vs. OLTP: What’s the Difference? | IBMibm.com
19. Use Amazon CloudFront Origin Shield - Amazon CloudFrontdocs.aws.amazon.com
20. Caching and availability - Amazon CloudFrontdocs.aws.amazon.com
21. What is High-Performance Computing (HPC)? - HPC Explained - AWSaws.amazon.com
22. Scenarios - High Performance Computing Lensdocs.aws.amazon.com
23. Best Practices Guide for Energy-Efficient Data Center Designenergy.gov
24. Cooling Water Efficiency Opportunities for Federal Data Centersenergy.gov
25. AWS Regions and Availability Zones - AWS Regions and Availability Zonesdocs.aws.amazon.com
26. AI and energy security – Energy and AI – Analysis - IEAiea.org
27. Eemshaven, Netherlands – Google Data Center Locationdatacenters.google
28. Site Selection Considerations for Edge Data Centerstiaonline.org
29. Availability Zones - AWS Fault Isolation Boundariesdocs.aws.amazon.com
30. Energy demand from AI – Energy and AI – Analysis - IEAiea.org
31. Executive summary – Energy and AI – Analysis - IEAiea.org
32. TIA-942tiaonline.org
33. NIH Sustainable Data Center Design Guidewbdg.org
34. Cooling recommendations for the roomibm.com
35. Model 7965-S42 water cooling manifold (Feature codes ECR3 and ECR4)ibm.com
36. PowerEdge R760dell.com
37. In-band or out-of-band: Advantages of iDRAC and iSM compared to OMSAdownloads.dell.com
38. Systems managementi.dell.com
39. CUDA Programming Guidedocs.nvidia.com
40. 2. Programming Model — CUDA Programming Guidedocs.nvidia.com
41. NVIDIA DGX H100 — Datasheetnvidia.com
42. What Is Data Storage?ibm.com
43. What Is a Storage Area Network (SAN)?ibm.com
44. Object vs. File vs. Block Storage: What’s the Difference?ibm.com
45. What Is File Storage?ibm.com
46. Cloud Object Storage as a Service: IBM Cloud Object Storage from Theory to Practiceredbooks.ibm.com
47. SDN and its Role in Automating & Scaling in the Data Center - Ciscocisco.com
48. Cisco Data Center Infrastructure 2.5 Design Guide - Data Center Architecture Overviewcisco.com
49. Enterprise Networks Structured Cabling | CommScopecommscope.com
50. Fiber Cable Assemblies for Data Centers | CommScopecommscope.com
51. Cisco Massively Scalable Data Center Network Fabric Design and Operation White Papercisco.com
52. Cross Connect | Equinix Product Documentationdocs.equinix.com
53. Internet Exchange | Equinix Product Documentationdocs.equinix.com
54. Internet Exchange FAQs | Equinix Product Documentationdocs.equinix.com
55. Physical Interconnection | Equinixequinix.com
56. Data Center Interconnect Design Guide for Virtualized Workload Mobility with Cisco, EMC, and VMware - Cisco Virtualized Workload Mobility Introductioncisco.com
57. Guide to Security for Full Virtualization Technologiesnist.gov
58. NIST Special Publication 800-125, Guide to Security for Full Virtualization Technologiesnvlpubs.nist.gov
59. Application Container Security Guidenvlpubs.nist.gov
60. OpenStack Compute (nova)docs.openstack.org
61. Nova System Architecturedocs.openstack.org
62. Scheduling, Preemption and Evictionkubernetes.io
63. Resource Management for Pods and Containerskubernetes.io
64. Compute service architecturedocs.openstack.org
65. Secure Virtual Network Configuration for Virtual Machine (VM) Protectionnvlpubs.nist.gov
66. Best Practices Guide for Energy-Efficient Data Center Designenergy.gov
67. UPS Fundamentals & FAQs | How to choose the right backup power solutionorigin-edge.eaton.com
68. Automatic transfer switches (ATS) fundamentals | Eatoneaton.com
69. CHAPTER 20. DATA CENTERS AND TELECOMMUNICATION FACILITIEShandbook.ashrae.org
70. Water-Cooled Servers Common Designs, Components, and Processesashrae.org
71. REL 10. How do you use fault isolation to protect your workload?docs.aws.amazon.com
72. AWS Fault Isolation Boundariesdocs.aws.amazon.com
73. UPS system redundancy typesproduct-help.schneider-electric.com
74. Myths and Misconceptions Regarding the Uptime Institute’s Tier Systemjournal.uptimeinstitute.com
75. AWS Fault Isolation Boundariesdocs.aws.amazon.com
76. Data Center Controlsaws.amazon.com
77. Disaster recovery options in the clouddocs.aws.amazon.com
78. Recovery Point Objective - Glossarycsrc.nist.gov
79. Recovery Time Objective - Glossarycsrc.nist.gov
80. SP 800-34 Rev. 1, Contingency Planning Guide for Federal Information Systemscsrc.nist.gov
81. Explaining the Uptime Institute’s Tier Classification System (April 2021 Update)journal.uptimeinstitute.com
82. TIA Unveils Revision C of the Global Data Center Infrastructure Standardtiaonline.org
83. Telecommunications Industry Association 942 Certifications & Ratingstiaonline.org
84. ANSI/TIA-942 Audit Servicestiaonline.org
85. Uptime Institute Publications — Tier Certificationuptimeinstitute.com
86. Tier Certification for Modular and Phased Constructionjournal.uptimeinstitute.com
87. Tier Certification Operational Sustainabilityuptimeinstitute.com
88. AWS Data Centersaws.amazon.com
89. SP 800-125, Guide to Security for Full Virtualization Technologiescsrc.nist.gov
90. NIST Special Publication 800-125, Guide to Security for Full Virtualization Technologiesnvlpubs.nist.gov
91. SP 800-125A Rev. 1, Security Recommendations for Server-based Hypervisor Platformscsrc.nist.gov
92. Amazon Web Services: Risk and Compliance - AWS Whitepaperdocs.aws.amazon.com
93. Guidelines for Media Sanitizationnvlpubs.nist.gov
94. What is Data Center Management? | IBMibm.com
95. Data Center Facilities Technician, Controls — Google Careersgoogle.com
96. Advancing security – Google Data Centersdatacenters.google
97. Google shares data center security and design best practicescloud.google.com
98. IBM Control Desk documentationibm.com
99. Big Data Strategies for Data Center Infrastructure Management Using a 3D Gaming Platformarxiv.org
100. Data Center Management and Operations Criteria - Uptime Instituteuptimeinstitute.com
101. The Making of a Good Method of Procedurejournal.uptimeinstitute.com
102. Start With the End in Mind!journal.uptimeinstitute.com
103. Google SRE - Incident Management Guidesre.google
104. Google SRE - Postmortem Practices for Incident Managementsre.google
105. Data center staffing is key to operating a data centerjournal.uptimeinstitute.com
106. 5 Essential Components of Data Center System Planningblog.se.com
107. Improve Project Success Through Mission Critical Commissioningjournal.uptimeinstitute.com
108. Microsoft 2022 Annual Reportmicrosoft.com
109. Four pillars of a successful circular datacenter hardware programdatacenters.microsoft.com
110. Sustainable by design: Innovating for zero wastemicrosoft.com
111. True value: Assessing and optimizing the cost of computing at the data center levelresearch.ibm.com
112. The Economics of AWSaws.amazon.com
113. The Economics of the AWS Cloud vs. Owned IT Infrastructured0.awsstatic.com
114. Cloud Financial Management with AWSaws.amazon.com
115. World Energy Investment 2025iea.blob.core.windows.net
116. Water-Use Terminology | U.S. Geological Surveyusgs.gov
117. Water Usage Effectiveness (WUE): A Green Grid Data Center Sustainability Metricthegreengrid.org
118. GHG Protocol Scope 2 Guidanceghgprotocol.org
119. 2022 Environmental Sustainability Reportcdn-dynmedia-1.microsoft.com
120. Circular Datacenter Infographic Spreaddatacenters.microsoft.com
121. DistributedDataParallel — PyTorch documentationdocs.pytorch.org
122. Performance Tuning Guide — PyTorch Tutorialsdocs.pytorch.org
123. Slurm Workload Manager — Gang Schedulingslurm.schedmd.com
124. Disaggregated Serving — NVIDIA Dynamo Documentationdocs.nvidia.com
125. Key Components of the DGX SuperPODdocs.nvidia.com
126. Overview — NVIDIA Mission Control Software with GB200/GB300 NVL72 Systems Administration Guidedocs.nvidia.com
127. Storage Architecture — NVIDIA DGX SuperPODdocs.nvidia.com
128. NVIDIA Contributes NVIDIA GB200 NVL72 Designs to Open Compute Projectdeveloper.nvidia.com
129. DGX SuperPOD Architecturedocs.nvidia.com