Data center: the definition
A data center is a physical facility or dedicated space that houses computing, storage and networking equipment together with the power, cooling, connectivity and operational systems needed to support it.
The key points
- A data center supplies physical infrastructure; a cloud service adds another operating and commercial layer.
- Facility, campus, cloud region and availability zone are different entities.
- Total electrical capacity, IT capacity and currently available capacity must not be added or substituted casually.
- Redundancy labels and certification scope do not replace a workload-specific service and recovery assessment.
The physical layer of a digital service
A cloud application ultimately runs on physical equipment somewhere. Data centers organize that equipment with supporting electrical, mechanical, network and operational systems. Servers perform computation, storage systems retain data, and network equipment connects components and customers. Power and cooling keep the hardware inside its supported operating conditions. The facility is therefore more than a room containing racks. [1]
Imagine a hypothetical AI service with excellent software and access to many GPU cards, but no commissioned location with adequate power and cooling. The cards alone are not a functioning service. Conversely, an empty powered building is not a GPU cloud. Physical readiness, installed IT equipment, software and commercial availability form different parts of a deployable product.
Facility, campus, region and service
A facility is a physical location or building-level entity. A campus can contain several facilities and shared infrastructure. A cloud region is a provider-defined service geography, and an availability zone is another provider-defined deployment boundary. Their precise mapping to physical buildings varies. A region label should not be treated as a precise address or proof of one particular facility without supporting evidence. [1][2]
For an illustrative database, a campus might be listed with 100 MW of eventual capacity and two buildings with 30 MW each in operation. Adding all three numbers would double-count overlapping scopes and mix planned with operating capacity. Preserve the parent–child relationships and the definition of each field before calculating a total. A map can look convincing while making a false claim if those scopes are ignored.
Enterprise, colocation and cloud operating models
An enterprise facility supports an organization's own infrastructure. Colocation commonly supplies space and supporting services for customers' equipment. A cloud provider exposes computing services while operating additional hardware and software layers. These arrangements can overlap physically: a cloud service may operate inside leased colocation space. Ownership of the building, ownership of the servers and responsibility for the application need not belong to the same company. [1]
A buyer should identify who controls each part of the service. Who supplies the servers, patches the host, maintains cooling, provides network access and responds to incidents? The answer changes with the product. A colocation quotation and a GPU-hour listing are not interchangeable offers merely because both relate to the same building.
From incoming electricity to IT load
Electrical infrastructure distributes energy from its supply to the equipment, with conversion, protection and continuity mechanisms appropriate to the design. UPS systems and backup generation address particular interruptions, but their capacities, runtime and operating procedures matter. A single redundancy label does not describe every electrical path or every failure mode. [3][1]
Megawatts measure power, while megawatt-hours measure energy over time. IT power describes the equipment load; total facility power includes supporting consumption. PUE relates facility energy to IT energy over a consistent boundary and period. A facility's total electrical capacity cannot simply be presented as the same amount of available server load without accounting for that distinction. [4]
Suppose a hypothetical site continuously supports 10 MW of IT load and has a PUE of 1.3 for the relevant conditions. Its corresponding total power is approximately 13 MW under that simplified assumption. Over 24 hours, the totals are 240 MWh for IT and 312 MWh for the facility. These are illustrative operating quantities, not a statement that a 13 MW nameplate guarantees 10 MW immediately available to customers.
Rack density changes the engineering problem
Heat must move from processors to a facility heat-rejection system. Air cooling, direct-to-chip liquid cooling and other arrangements have different interface requirements. A facility can support a mix rather than one uniform design. ASHRAE distinguishes technology-side and facility-side cooling arrangements and notes that liquid-cooled systems can still require auxiliary air cooling. [5]
Consider a hypothetical room rated for 1 MW overall. Ten racks at 100 kW each and one hundred racks at 10 kW each have equal aggregate IT power but very different local distribution and heat-removal requirements. Floor space and total megawatts alone do not establish that the high-density arrangement can be installed. Confirm the supported rack-level design and commissioning scope.
Networks and failure domains
A data center includes internal and external connectivity. Internal networks connect servers and storage; external connectivity reaches carriers, clouds and customers. Diversity must be understood physically and operationally, not inferred from the number of commercial network names in a directory. Two services can share an upstream path or a physical route even when their contracts look separate. [1]
As an illustrative resilience question, suppose an application has two network providers but both circuits enter through the same vulnerable route. A local incident could affect both. The useful question is what failures the design survives, not simply how many suppliers appear on the invoice. Similar reasoning applies to power feeds, cooling distribution and replicated application components.
What a Tier classification does and does not say
Uptime Institute's Tier framework distinguishes infrastructure capabilities, including redundant capacity components, concurrent maintainability and fault tolerance at the relevant levels. Certification scope matters: a design assessment, a constructed-facility assessment and operational practices are not identical evidence. The framework should not be casually translated into a universal percentage uptime guarantee for every application in a building. [3]
A resilient facility cannot correct every application defect. A software deployment can fail inside a highly capable building, and a redundant application can sometimes tolerate a component failure in a less elaborate facility. Match facility evidence, service commitments and application recovery design to the actual risk requirement rather than treat one label as a complete guarantee.
Announced, commissioned and available
A capacity announcement can describe a future campus, an expansion phase or a target operating scale. Commissioning establishes that installed systems have passed the relevant acceptance process. Customer availability adds another question about commitments, allocation and the hardware/software actually offered. These states should remain distinct in market research and procurement records.
A hypothetical operator may announce 200 MW for a campus while only 40 MW is commissioned and nearly all of that is already allocated. None of those figures is inherently wrong, but calling all 200 MW immediately available would be misleading. Date, boundary, phase and commercial status are necessary companions to the number.
The central lesson is that data-center capacity is a system-level claim. A credible description connects physical location, electrical and cooling limits, connectivity, commissioning and service responsibility. Understanding those layers lets a reader evaluate both a facility directory and a cloud offering without turning an attractive headline into an unsupported deployment promise.
Check your understanding
Try answering before opening the explanation. Your answers are not collected or scored.
1. Why should campus and building capacities not simply be added?
The scopes may overlap, and planned, operating and available quantities can describe different states. Parent–child relationships and definitions must be preserved.
2. Does 1 MW of room capacity guarantee support for any rack arrangement totaling 1 MW?
No. Rack-level power distribution, heat removal and physical design impose additional constraints beyond the aggregate total.
3. Does a Tier label guarantee an application's uptime?
No. Its infrastructure and certification scope must be understood alongside operational practices, service terms and application design.
Sources & editorial note
Reference documentation is listed below with its recorded check date. Technical statements are attributed; passages framed as our view or recommendation are editorial interpretation. Examples are hypothetical unless explicitly identified otherwise. No independent Kovara hardware testing is claimed.
- IBM · What is a data center? ↗ (opens in a new tab)Infrastructure explanation · Checked 28 September 2026
- AWS · Regions and Availability Zones ↗ (opens in a new tab)Provider documentation · Checked 28 September 2026
- Uptime Institute · Tier Classification System ↗ (opens in a new tab)Certification-body explanation · Checked 28 September 2026
- The Green Grid · Power Usage Effectiveness ↗ (opens in a new tab)Metric definition · Checked 28 September 2026
- ASHRAE · Data centers and telecommunications facilities ↗ (opens in a new tab)Engineering handbook · Checked 28 September 2026
Prepared with AI assistance. Publication authorized by Tommaso Luci; this does not claim independent technical peer review. Kovara Research is the publication label, not a claim of an independent laboratory or a named analyst team.