What Is a Data Center? Types, Components, and How They Work

A data center is a purpose-built facility that houses servers, storage, and networking equipment along with the power, cooling, and security systems those systems need to run without interruption. It is the physical infrastructure behind essentially every website, app, and AI model — ranging from a single rack in a shared colocation cage to a 100+ MW hyperscale campus dedicated to one cloud provider.

Key takeaways

• Five main types: enterprise/on-premises, colocation, hyperscale, edge, and cloud-provider facilities — distinguished by who owns the building and who uses the compute inside it.
• Four Uptime Institute tiers (I–IV) classify reliability, from a single non-redundant power path (99.671% target uptime) to fully fault-tolerant dual paths (99.995%), per Uptime Institute.
• Core components are the same everywhere: power (utility feed, generators, UPS), cooling (CRAC/CRAH, chilled water, or liquid cooling), compute/network racks, and physical security/fire suppression.
• Global capacity is projected to nearly triple to 219 GW by 2030, with AI workloads driving ~70% of new demand, per McKinsey.
• Around 11,000–12,000 data centers exist worldwide across roughly 180 countries by industry-tracker counts (Statista), though the figure depends heavily on what is counted as a “data center.”
• Coloprice’s catalog tracks 2,773 verified colocation and wholesale facilities in 19+ countries, sourced only from operator disclosures — a narrower, source-verified subset of the global total.
• Global data center capex has been raised above $1 trillion for 2026 and is forecast to exceed $3 trillion annually by 2030, according to Dell’Oro Group.

For live pricing by market, see the Colocation Price Index; for individual facilities, see the data center catalog.

The core definition

At its simplest, a data center is a room, building, or campus engineered to keep IT equipment running continuously. That means four things exist together in every data center, regardless of size:

1. IT equipment — servers, storage arrays, and network switches, typically mounted in standardized racks.
2. Power infrastructure — a utility feed, backup generators, and uninterruptible power supplies (UPS) that bridge the gap between a power failure and generator start.
3. Cooling infrastructure — computer room air conditioning/handling (CRAC/CRAH) units, chilled-water plants, or, increasingly for AI racks, direct-to-chip and immersion liquid cooling.
4. Physical security and fire suppression — badge/biometric access control, CCTV, and clean-agent or pre-action fire systems designed not to damage equipment during a false alarm.

What separates a data center from an office server closet is redundancy and formal design intent: a data center is engineered from the start to a specific uptime target, with defined maintenance procedures and (usually) third-party audited compliance. See our data center certifications guide for how that gets verified.

The five types of data centers

Type	Who owns it	Who uses it	Typical size
Enterprise / on-premises	The company itself	Single organization	Server room to a few MW
Colocation	Third-party operator	Multiple tenants, shared building	1 rack to 50+ MW campus
Hyperscale	Cloud/AI company or built-to-suit for one	Single tenant, often the owner	50–1,000+ MW campus
Edge	Telco, operator, or enterprise	Local/regional workloads needing low latency	Sub-1 MW, often <100 kW
Cloud-provider (public cloud)	AWS, Azure, Google Cloud, etc.	Thousands of customers via virtual instances	Runs inside hyperscale facilities

The practical decision most buyers face is not “what is a data center” but which of these models fits a given workload — a comparison we cover in colocation vs. cloud, bare metal vs. colocation, and hyperscale vs. colocation. Edge deployments get their own tradeoffs, detailed in edge data centers.

How data centers are classified: the Uptime Institute tiers

Reliability is standardized through four tiers, each building on the one below:

Tier	Redundancy	Concurrently maintainable	Target uptime	Max annual downtime
Tier I	Single, non-redundant path	No	99.671%	~28.8 hours
Tier II	N+1 redundant components	No	99.741%	~22 hours
Tier III	Multiple independent distribution paths	Yes	99.982%	~1.6 hours
Tier IV	Fault-tolerant, dual paths	Yes	99.995%	~26 minutes

Source: Uptime Institute Tier Classification System. Most colocation RFPs from enterprise and AI buyers specify Tier III as the floor; Tier IV is reserved for the highest-criticality financial, healthcare, or government workloads given the added construction cost. Our tiers explained guide breaks down what each tier actually requires on-site.

How power and cooling actually work

Utility power arrives at the facility and is split across at least one UPS-backed path. If utility power fails, generators — sized to carry full IT and cooling load — start within seconds, while the UPS bridges the gap on batteries. Cooling removes the heat that IT load generates, historically via chilled air (CRAC/CRAH units pushing air through a raised floor or overhead ducting) and increasingly, for AI racks running 60–130+ kW, via direct-to-chip or immersion liquid cooling — covered in our liquid cooling guide and power density guide.

Efficiency is measured as Power Usage Effectiveness (PUE): total facility power divided by IT power. A PUE of 1.5 means for every 1 kW of IT load, the facility draws 0.5 kW more for cooling, power conversion losses, and lighting. Modern hyperscale campuses run PUE near 1.1–1.2; older retail colocation facilities often sit at 1.4–1.6.

Who actually operates data centers

Ownership splits into a handful of recognizable groups. Wholesale/retail colocation operators (Equinix, Digital Realty, NTT, STT GDC, and hundreds of regional players) build and lease space to multiple tenants — several are structured as REITs, detailed in our data center REIT analysis. Hyperscalers (Amazon, Microsoft, Google, Meta) increasingly build and own their own campuses rather than leasing, especially for AI training clusters. Neoclouds (CoreWeave, Lambda, and similar GPU-focused providers) lease shells or wholesale capacity from operators and resell GPU compute by the hour — a model contrasted with traditional hyperscalers in our neocloud vs. hyperscaler guide. And a long tail of enterprises still run their own on-premises facilities, though most net-new capacity since 2020 has gone to colocation and hyperscale rather than enterprise-owned space, as workloads shift to outsourced models.

Financing these buildings is itself a distinct market — debt, equity, and build-to-suit development financing, covered in how AI data centers get financed — because a single hyperscale campus can now cost well into the billions of dollars before a single server is racked.

The scale of the market in 2026

Data centers have moved from back-office infrastructure to one of the largest categories of global capital spending:

• Global data center capex has been raised above $1 trillion for 2026 and is forecast to exceed $3 trillion annually by 2030 — nearly double Dell’Oro’s own January 2026 outlook of $1.7 trillion.
• McKinsey estimates cumulative 2025–2030 investment need at $6.7 trillion, of which $5.2 trillion is specifically for AI capacity.
• The colocation segment alone is worth $91–104 billion in 2025, projected to reach $165–204 billion by 2030 (14–16% CAGR), per MarketsandMarkets and Grand View Research.
• Global capacity is set to grow from roughly 82 GW in 2025 to 219 GW by 2030, with AI-ready capacity rising from ~54% to ~71% of that total, per McKinsey.

That growth is why vacancy has collapsed in major markets — North American primary-market vacancy hit a record-low 1.4% in H1 2026, per CBRE — and why grid connection, not floor space, is now the binding constraint on new supply almost everywhere.

How many data centers exist, and where

Industry aggregators (Statista, Data Center Map-style trackers) put the global count near 11,000–12,000 facilities across roughly 180 countries as of mid-2026. That number is soft: it depends entirely on what gets counted, and most aggregators mix verified operator-disclosed facilities with scraped or self-reported listings of uncertain provenance.

Coloprice takes a narrower, source-verified approach: the data center catalog currently lists 2,773 facilities across 19+ countries, built only from operator websites, press releases, BOI/investment-board filings, and Uptime Institute/EPI registries — never copied from other directories. Where an operator does not publish a figure (power capacity, certification, commissioning date), the field is left null rather than estimated. That makes the catalog smaller than aggregator counts by design, but every entry traces to a primary source, visible on each facility’s page and summarized on /sources/.

What to do with this

• Choosing between models: if you need dedicated space with operational control, start with colocation vs. cloud. If your workload is bursty or you don’t want a multi-year commitment, cloud or GPU-cloud (compared in our GPU cloud comparison) usually wins on flexibility.
• Picking a tier: default to Tier III unless your workload has zero tolerance for planned maintenance windows — Tier IV’s added redundancy rarely justifies its cost premium outside regulated, mission-critical use cases.
• Sizing power needs: if you’re deploying AI hardware, plan around 60–130+ kW per rack and confirm liquid-cooling readiness before signing, not after.
• Benchmarking price: compare any quote against live market rates on the Colocation Price Index and current market statistics before negotiating, or request a benchmarked comparison through Coloprice’s quote service.