Data Centre Colocation Fundamentals: Power, Cooling, Connectivity and Resilience

Introduction

A data centre is a purpose-built facility that houses servers, storage systems and network equipment. A server is a computer that runs applications or provides data; storage systems keep that data; and network equipment moves it between systems and users. The facility supplies the dependable electricity, heat removal, communications links, fire protection, monitoring and physical security that digital services need. An office server room may contain similar equipment, but a data centre is engineered to operate it continuously and at much greater scale.

Colocation means placing equipment that an organisation owns or controls inside a data centre operated by another party. The facility operator supplies the building and shared infrastructure; the customer normally remains responsible for its servers, operating systems, applications, data and backups. Think of an apartment building: the operator looks after the structure, utilities and common security, while each tenant remains responsible for what happens inside its own space.

These concepts matter because usable capacity is a connected system. Empty floor space is useless without power. Power cannot be used safely unless cooling removes the resulting heat. Reliable equipment is still isolated if its network routes fail. Good engineering can also be defeated by weak procedures, poor maintenance, unclear responsibilities or unrealistic growth forecasts.

Colocation models and responsibility

A multi-tenant facility serves several customers in separately controlled areas. A customer may rent a rack, an open metal frame for mounting IT equipment, or a cabinet, an enclosed rack with panels or doors. Larger deployments may use several racks inside a locked cage or a fully enclosed private suite. The choice affects airflow, access, security, scale and how quickly the deployment may grow.

Colocation is not the same as managed hosting or cloud computing. In managed hosting, another party also supplies or operates some IT equipment. Cloud computing is an on-demand service model in which users consume computing resources without managing the underlying physical equipment. Cloud services may themselves run in owned or colocated data centres. Hybrid IT combines customer-owned systems, colocation and cloud services.

The shared-responsibility boundary

The shared-responsibility boundary states who controls each layer. It should cover the building, utilities, racks, cabling, cross-connects, server hardware, operating systems—the core software that controls a server—applications, data, backups, monitoring and incident response. A resilient building does not make a customer's application resilient, and a secure cage does not patch an exposed server.

Space, density and capacity planning

White space is the technical floor area used for active IT equipment. It is only one part of capacity. A deployment simultaneously consumes electrical power, cooling, rack space, structural load, network ports, cable pathways and access space. The smallest available resource becomes the bottleneck. Free floor area with no usable power or cooling is called stranded capacity.

IT load is the electrical demand from servers, storage and network equipment. Rated IT load is the maximum load the complete power and cooling design can support while preserving its intended availability. Installed capacity is the rated capability of infrastructure already installed, committed capacity is already reserved by contract, and used capacity is the measured demand. These numbers are not interchangeable.

Critical IT load is the supported electrical demand of the IT equipment itself, excluding facility overhead such as cooling. Utilisation compares actual demand with usable or rated capacity. Very low utilisation can make power and cooling plant inefficient; very high utilisation leaves little growth or failure margin. Load factor instead compares average demand with peak demand over a defined period, showing how steady or uneven the load is.

Rack density is the IT power concentrated in one rack, usually stated in kilowatts per rack. A kilowatt (kW) is a rate of power; a kilowatt-hour (kWh) is the energy used over time. The distinction is like water: kW describes how fast it flows, while kWh describes how much passed through the pipe. A megawatt (MW) is one thousand kilowatts.

High-density colocation supports racks whose power and heat exceed the site's normal design baseline. There is no durable universal threshold. What matters is whether every link in the electrical and cooling path supports the actual rack. Accelerator-based computing uses specialised processors for highly parallel work. High-performance computing coordinates many powerful systems for intensive technical or analytical work, while an artificial-intelligence cluster coordinates servers that train or run AI models. All three can concentrate far more power, heat and network traffic than conventional business systems.

Planning usable capacity

Capacity planning forecasts space, power, cooling, network and support needs across normal, peak, maintenance and failure conditions. A capacity block reserves a combined unit of these resources. Contiguous capacity means adjacent resources are available for expansion, while pre-provisioned capacity means power, cooling or fibre is prepared before equipment arrives. Both reduce deployment delay, but both risk tying up capital before demand is certain.

• Headroom is deliberately unused capacity for growth, load variation and failure conditions. Too little creates trips and delays; too much can make power and cooling equipment inefficient at partial load.
• Derating reduces nameplate capacity to account for temperature, ageing, maintenance, power quality or reliability margins. Ignoring it overstates what can be delivered safely.
• Phased or modular expansion adds capacity closer to demand. It can reduce stranded investment, but future utility, pathway and equipment interfaces must be reserved and tested.
• Demand forecasting estimates future workload growth. A density roadmap must show not only total MW, but where higher-density racks will sit and how they will be powered and cooled.
• Workload placement chooses the facility, region or cloud best suited to latency, security, regulation, cost and recovery needs. Placement is an architectural decision, not merely a real-estate choice.

Power: from grid to rack

The critical power chain is the complete route electricity follows to the IT equipment. It normally includes the utility connection; transformers that change voltage; switchgear that protects and isolates circuits; an uninterruptible power supply (UPS) that provides near-instant ride-through; stored energy such as batteries; longer-duration standby generation; and power-distribution units (PDUs) that divide and deliver power at facility or rack level. One weak downstream link can limit a much larger upstream supply.

A UPS is the bridge, not the destination: it keeps equipment running while another source takes over. A standby generator can support a longer interruption, but only if it starts, transfers correctly, has usable fuel, meets emissions rules and has been maintained under realistic load. Power quality means electricity remains within safe voltage, frequency, distortion and grounding limits; power can be present yet still be unsafe for equipment.

N, N+1, 2N and dual paths

In redundancy notation, N is exactly the capacity needed for the design load. N+1 adds one spare capacity unit. 2N provides two complete capacity systems. These labels describe capacity, not the whole architecture. Two sets of equipment are not independent if they share a switchboard, control panel, fuel system, cable route or flood zone.

Many servers are dual-corded: one power cord connects to an A path and the other to a B path. Single-corded devices may need a transfer device, which can itself become a failure point. Concurrent maintainability means planned work on a capacity component or distribution path can occur without interrupting the supported IT load. Fault tolerance is the ability to continue operating with certain defined faults present; the protected fault set, operating conditions and assumptions must be stated. Neither term promises that the application above it will stay available.

Grid engagement is long-term coordination with utilities and authorities about connection size, delivery date, demand limits and expansion. Compute equipment can arrive much faster than a substation or transmission connection. Planning only the building can leave racks waiting years for usable electricity.

Cooling and environmental control

Almost every watt used by IT becomes a watt of heat that must be removed. Thermal management moves heat from chips to air or liquid, through facility equipment and finally outside the building. The useful measurement point is the IT inlet condition—the temperature and moisture of air entering the equipment—not merely the average room temperature.

In hot-aisle/cold-aisle layout, rack fronts face each other to form cold intake aisles and rack backs form hot exhaust aisles. Containment uses barriers to keep supply and exhaust air apart. Cool air that returns without entering equipment is bypass air; hot exhaust that reaches an intake is recirculation. Both waste capacity, and recirculation creates local hot spots even when the room appears cool.

Air and liquid cooling

A computer-room air conditioner (CRAC) commonly uses a refrigerant, a fluid that absorbs and releases heat. A computer-room air handler (CRAH) commonly transfers heat into a cool-water loop. A chiller produces that chilled water. A cooling tower may reject heat through evaporation, while a dry cooler rejects it to outdoor air without routine evaporation. An economiser, sometimes called free cooling, uses favourable outdoor conditions to reduce mechanical refrigeration. Each choice shifts electricity, water, climate and maintenance trade-offs.

Liquid cooling carries heat in a fluid rather than relying only on room air. Direct-to-chip cooling brings liquid to cold plates—flat heat exchangers attached to major heat-producing components. A rear-door heat exchanger captures hot rack exhaust. Immersion cooling places equipment in an electrically non-conductive fluid. A coolant distribution unit (CDU) controls flow and transfers heat between the IT liquid loop, the closed pipe circuit serving equipment, and the facility loop, the building-side circuit that carries heat away.

Liquid can carry concentrated heat efficiently, but 'liquid-ready' means more than available pipes. The design needs compatible materials, water or fluid quality, isolation valves, leak detection, drainage, CDU space, monitoring, service procedures and backup heat rejection. Direct-to-chip systems also leave memory, storage, power supplies or network equipment needing residual air cooling.

Environmental limits and failure planning

The recommended environmental envelope is the normal inlet range intended to balance reliability and efficiency. The wider allowable envelope is what compatible equipment can tolerate under defined conditions; it is not automatically the best everyday setpoint. Dew point is the temperature at which moisture condenses, so it helps operators manage condensation risk more consistently than relative humidity alone.

• Thermal monitoring detects hot spots, blocked airflow, lost coolant flow and failing equipment. Poor sensor placement or alarm overload creates false confidence.
• Thermal ride-through keeps heat removal adequate while electrical sources transfer. Servers continue producing heat immediately even when pumps, chillers or controls are changing power source.
• Mechanical redundancy provides spare chillers, pumps, fans or heat exchangers. Shared valves, controls and water loops still need a common-mode analysis: a check for one event that could disable several supposedly independent systems.
• Water stewardship evaluates consumption against local water availability. A design can save electricity through evaporation while worsening water stress.
• Too little cooling causes throttling—automatic slowing to reduce heat—along with shutdowns and shortened equipment life; excessive cooling wastes energy and capital. Condensation, pollutants, corrosion and fluid leaks create different failure paths.

Connectivity and interconnection

A colocation site is also a network junction. Interconnection directly links customer networks, carriers, cloud networks and services within or between data centres. A carrier-neutral facility accommodates several network providers rather than tying customers to one. This creates route choice, but the number of provider names does not prove that their physical paths are independent.

A meet-me room is a controlled space where external networks and customer cabling terminate. A cross-connect is a dedicated physical cable between two ports in the facility or campus. A cloud on-ramp is a dedicated or logically private connection into a cloud network. Peering means two networks exchange traffic directly, often through an internet exchange point, which is shared switching infrastructure for many networks.

Dark fibre is installed fibre that the customer activates with optical equipment, hardware that sends and receives light through the glass cable; a lit service is operated by a network provider. Virtual interconnection provisions logical connections over shared infrastructure through software controls. It can shorten delivery time, but its performance and resilience still depend on real fibre paths, network hardware and secure administration.

A network topology is the arrangement of sites, links and traffic paths. Software-defined networking uses software controls to configure those paths, while interconnection orchestration coordinates connections across facilities or services. The control plane is the software and decision layer that tells the network where traffic should go. Software can make a change faster; it cannot remove the physical dependencies underneath it.

Bandwidth is not latency

Bandwidth is a link's maximum carrying capacity; throughput is the data actually delivered after the network's communication rules, congestion and equipment take their share. Latency is travel and processing time. On a road, bandwidth is the number of lanes, throughput is the vehicles that actually arrive, and latency is the journey time. Jitter is variation in delay, while packet loss means some data units never arrive. A wider road does not make a distant destination nearby.

Physical route diversity means end-to-end paths use separate building entrances, conduits, meet-me spaces, equipment and upstream routes where the risk justifies it. Carrier diversity uses different network providers; it is an additional control, not proof of separate physical paths. One provider can operate genuinely separate routes, while two providers can share a street duct or exchange. A topology drawing, failure test and clear demarcation—the handover point between responsibilities—are stronger evidence than the word 'diverse' on an order form.

Resilience and operations

Availability is the ability to perform when required. Reliability is the likelihood of operating without failure for a period. Resilience is broader: the ability to withstand disruption, continue or degrade safely and recover. Recoverability is the ability and speed of restoration. These qualities overlap, but they are not synonyms.

A single point of failure is one component, path, control or human action whose loss stops service. A common-mode failure is one event that defeats supposedly redundant systems, such as shared firmware, fuel, controls or flood exposure. A failure domain is the set of systems likely to be affected by one fault. Understanding domains matters more than simply counting spare components.

Recovery objectives and geographic diversity

A recovery time objective (RTO) is the longest acceptable restoration time. A recovery point objective (RPO) is the maximum tolerable data loss measured backward in time. Business continuity keeps essential business activities operating, while disaster recovery restores IT services and data. A second facility reduces geographic risk only when applications replicate correctly, network failover works and people regularly test the recovery plan.

Operating discipline

A strong design reaches its potential only through staffing, maintenance and controlled change. A standard operating procedure (SOP) covers routine work. A method of procedure (MOP) details a planned change, including checks, roles, approvals and rollback. An emergency operating procedure (EOP) guides time-critical response to abnormal conditions. Procedures that are vague invite improvisation; procedures that are too complex are often bypassed.

• Commissioning tests whether installed systems perform as designed. Integrated systems testing exercises interactions such as utility loss, UPS ride-through, generator start, cooling recovery, alarms and degraded modes.
• Preventive maintenance follows planned intervals; predictive maintenance uses condition data to act before failure. Deferred work may save money briefly while building hidden risk.
• A network operations centre (NOC) monitors service and coordinates incidents. A building-management system (BMS) controls mechanical systems; an electrical power-monitoring system (EPMS) observes power; and data-centre infrastructure management (DCIM) combines facility and IT capacity data.
• A computerised maintenance-management system (CMMS) tracks assets, work orders and maintenance records. Inaccurate inventories and labels turn even simple changes into risk.
• Remote hands means authorised on-site staff perform physical tasks for an absent customer. Identity checks, exact instructions, change records and stop conditions prevent the wrong cable or device being touched.

A service-level agreement (SLA) defines measurable commitments, exclusions, maintenance treatment, remedies and reporting. Its availability percentage must state the measurement point and period. A service credit may compensate a breach, but it does not restore data, reputation or a missed business deadline. Contracts describe accountability; architecture and operations create resilience.

Security, data location and compliance

Defence in depth uses several independent security layers so one control failure is not enough. Physical layers may include the site boundary, building entrance, staffed reception, an anti-tailgating vestibule that admits one person at a time, a secure room, a cage and a locked cabinet. Multi-factor authentication (MFA) requires more than one type of identity evidence. Least privilege gives people and systems only the access required for their role and time window.

Access logs and video monitoring provide investigation and audit evidence, but their retention and privacy rules must be clear. Tenant segregation prevents one customer from reaching another's space, cabling, consoles, traffic or data. Emergency access needs a controlled break-glass process: strong enough to prevent misuse, but fast enough for a genuine life-safety or service incident.

Physical and cyber systems meet

Operational technology (OT) is the hardware and software that monitors or controls physical processes. In a data centre it includes building, electrical, cooling, access-control and environmental systems. Flat management networks, shared passwords, exposed remote access or unpatched controllers can threaten availability without an attacker ever touching a customer server. Facility controls therefore need network segmentation, secure administration, monitoring and recovery plans.

The customer still protects its own data and applications. The basic goals are confidentiality, integrity and availability: prevent unauthorised disclosure, prevent unauthorised alteration and keep information usable when required. Encryption converts data into a form that unauthorised parties cannot read. Private connectivity reduces public routing exposure but does not replace encryption, identity control, secure configuration, backups or incident monitoring.

Residency, sovereignty and assurance

Data residency is the geographic location where data is stored or processed. Data sovereignty means data is subject to the laws of a jurisdiction. A regulated workload is a system governed by sector or public rules. These requirements can affect site selection, support access, backup location, network paths and the evidence retained.

Compliance zoning physically or logically separates systems with different obligations. It can improve assurance while duplicating controls and leaving fragments of capacity that are hard to reuse. An audit report or certification is scoped evidence that specified controls were assessed for a stated place and period. It does not prove future uptime, secure every customer configuration or transfer the customer's regulatory responsibility.

Sustainability and economics

Data-centre sustainability covers energy, carbon, water, materials, equipment life, land and local impact. No single number captures all of them. Metrics are useful only when their measurement boundary, period, climate and workload are comparable.

Operational resource metrics

Power Usage Effectiveness (PUE) is total data-centre energy divided by IT-equipment energy over the same boundary and period. Standard PUE uses continuous measurement over 12 months; shorter-period or restricted-boundary variants should be labelled as such. A lower number means less facility overhead, and PUE minus 1 shows that overhead relative to IT energy; 1.0 is the theoretical floor. PUE does not measure useful computing, total electricity, carbon, water or resilience. It can improve while total energy rises, and comparisons can mislead when utilisation and boundaries differ.

• Water Usage Effectiveness (WUE) relates water consumption to IT energy. Interpret it with water source, local scarcity, climate and measurement boundary.
• Carbon Usage Effectiveness (CUE) relates use-phase operational carbon-dioxide emissions to IT-equipment energy. It depends on the electricity mix and the emissions accounting method.
• Renewable Energy Factor (REF) reports eligible renewable energy as a share of total data-centre energy under a stated method.
• Energy Reuse Factor (ERF) reports useful energy exported or reused as a share of total data-centre energy. Heat reuse works only when a nearby user needs the available temperature at the right times.
• Embodied impact covers emissions and resource use from constructing buildings and making, transporting and replacing electrical, cooling and IT equipment. Operating metrics do not capture it.

Optimising one measure can worsen another. Evaporative cooling may reduce electrical demand but use more water. Very high redundancy may improve continuity while adding equipment and conversion losses. Renewable-energy contracts may improve annual accounting without matching consumption hour by hour or at the same grid location. Good decisions state these trade-offs instead of turning one metric into a score.

Cost and commercial capacity

Capital expenditure (CapEx) buys long-lived assets such as buildings, substations and cooling plant. Operating expenditure (OpEx) covers ongoing power, space, connectivity, support and maintenance. Total cost of ownership (TCO) combines acquisition, operation, growth, risk and exit over the chosen period. Colocation often replaces direct facility-construction spending with contracted service payments, but the customer still funds its IT equipment and migration; the accounting classification depends on the contract and applicable rules.

• Committed power reserves a set capacity; metered power charges according to measured consumption. Reservation protects growth but can leave paid capacity idle.
• Recurring charges may include space, power reservations, cross-connects and support. Non-recurring charges may include installation, cabling, fit-out and migration.
• A realistic TCO model includes escalation clauses, remote-hands rates, network circuits, testing, equipment refresh, expansion rights, decommissioning and exit work—not only the rack price.
• Site selection must balance power and fibre availability, hazards, permits, workforce, water, land, latency, legal jurisdiction and community impact. A cheap site can be costly if it delays deployment or weakens recovery.

How the concepts connect

A data centre works less like one machine and more like a small city. Power is the utility network; cooling is waste-heat removal; fibre is the road system; security controls the borders; monitoring is the control room; procedures are the operating rules. The service succeeds only when all of them support the same workload at the same time.

Consider a new high-density computing cluster. Its processors raise rack power. Higher power raises heat load, which may require liquid cooling and residual air cooling. The cluster also needs high-throughput, low-latency networking. Its size may require contiguous space and power, while regulated data may constrain location and support access. The change therefore affects utility capacity, pipe routes, CDUs, fibre, security procedures, maintenance skills, recovery design, cost and sustainability together. Replacing servers alone would miss most of the project.

Glossary

This glossary is the reusable keyword set extracted from the anonymous source and normalised to current, generic industry usage. Related terms are combined where that makes the distinction clearer.

Facility, service and capacity terms

Power, cooling and resilience terms

Connectivity, security and operations terms

Sustainability and commercial terms

What to remember

• Colocation combines customer-controlled IT with professionally operated facility infrastructure; the responsibility boundary must be explicit.
• Usable capacity is the smallest simultaneous amount of space, power, cooling, structure, connectivity and support—not the biggest number on a brochure.
• Every watt delivered to computing becomes heat, so power density and cooling design are one planning problem.
• N+1 and 2N describe capacity arrangements. Real resilience also needs independent paths, tested controls, trained people, maintenance and recovery.
• Bandwidth describes network capacity; latency describes delay. Separate network contracts do not prove physically diverse routes.
• A private connection or secure building does not replace application security, encryption, backups or customer incident response.
• PUE is the ratio of total facility energy to IT energy; PUE minus 1 expresses facility overhead relative to IT. The ratio alone does not show absolute energy use, carbon, water, useful computing or resilience.
• Capacity planning must include normal, peak, maintenance and fault conditions, plus the long lead times of utility, construction and network work.
• An SLA defines contractual accountability. Architecture, operations and tested recovery determine what actually happens during a disruption.
• The best decision follows dependencies from business need to workload, rack, grid, heat sink, networks, people, contracts and recovery sites.

Learning and discussion questions

1. Why can a room with empty rack space still have no usable capacity?
2. Which facility responsibilities transfer in colocation, and which remain with the customer?
3. How does a higher rack density change power, cooling, network, cost and operating-skill requirements?
4. Why can two electrical or network paths fail together even when both are labelled redundant?
5. What is the difference between bandwidth, throughput and latency?
6. When might liquid cooling be appropriate, and which new failure and maintenance risks does it introduce?
7. How do RTO and RPO change the design of backups, replication and a second site?
8. Why can a low PUE still accompany high total energy use, carbon emissions or water stress?
9. What evidence would you ask for before accepting a claim of route diversity or concurrent maintainability?
10. How should finance compare reserved capacity, rapid expansion rights and the risk of paying for unused infrastructure?
11. How do data residency and data sovereignty influence workload placement?
12. Choose one critical workload. What is its weakest shared dependency across power, cooling, networking, security, operations and recovery?

Sources and further reading

• European Union — data-centre, operator, customer and colocation definitions
• ISO/IEC 22237-1:2021 — general data-centre concepts and lifecycle requirements
• ISO/IEC 22237-3:2021 — power distribution, measurement and protection
• ISO/IEC 22237-4:2021 — environmental control
• ISO/IEC 22237-2:2024 — building construction and environmental risk protection
• ISO/IEC 22237-6:2024 — physical security systems
• ISO/IEC TS 22237-31:2026 — data-centre infrastructure resilience KPIs
• ISO/IEC TS 8236-2:2025 — facility-provisioning profiles and forecasting methods
• ISO/IEC 30134-2:2026 — Power Usage Effectiveness
• ISO/IEC 30134-9:2022 — Water Usage Effectiveness
• ISO/IEC 30134-8:2022 — Carbon Usage Effectiveness
• ISO/IEC 30134-3:2016 with Amendment 1:2018 — Renewable Energy Factor
• ISO/IEC 30134-6:2021 — Energy Reuse Factor
• ASHRAE Handbook — data-centre environmental conditions and thermal guidance
• US Department of Energy — Best Practices Guide for Energy-Efficient Data Center Design
• US Department of Energy — data-centre cooling-water efficiency guidance
• NIST SP 800-34 Rev. 1 — contingency planning, RTO and RPO
• NIST SP 800-82 Rev. 3 — operational-technology security
• NIST SP 800-53 Rev. 5 — physical, environmental and contingency controls
• NIST SP 800-215 — secure enterprise-network guidance
• Internet Society — internet exchange point definition and resilience benefits
• Telecommunications Industry Association — overview of ANSI/TIA-942-C:2024 infrastructure updates
• ISO 22301:2019 with Amendment 1:2024 — business-continuity management systems
• GHG Protocol Corporate Standard — Scope 1 and Scope 2 emissions