What Is a Data Center? How the Buildings That Run the Internet Work

What Is a Data Center?

A data center is a purpose-built facility that consolidates computing infrastructure — servers, storage, networking, cooling, and power — in one controlled environment. The goal is to keep that infrastructure running continuously, at high availability, and under physical and digital security controls.

The term covers a wide range of facilities. A small enterprise data center might occupy one room in an office building and hold 20 racks of servers. A hyperscale facility built by Google or Microsoft might cover 500,000 square feet across multiple connected buildings, hold hundreds of thousands of servers, and require its own power substation.

What all data centers share, regardless of size, is the same core architecture: servers that do the computing, storage systems that hold the data, networking equipment that moves data in and out, cooling systems that remove the heat generated by servers, and power systems that keep everything running even when the grid fails.

The physical address of a data center rarely matters to the people using its services. When you use Gmail, Netflix, or ChatGPT, you are consuming compute capacity from data centers you will never see, likely in states like Virginia, Texas, or Oregon.

What Is Inside a Data Center?

Data centers contain five categories of physical infrastructure. Each one is engineered for redundancy, meaning if one component fails, a backup takes over without interrupting service.

Servers

Servers are the computing core. They look like thin, horizontal boxes — 1U or 2U form factors — that slide into racks. Each server contains one or more processors (CPUs), memory (RAM), storage (SSD or HDD), and network interfaces. A standard 42U server rack holds up to 42 of these units. Most enterprise data centers have dozens to hundreds of racks. Hyperscale facilities have tens of thousands.

AI data centers differ from traditional ones in one critical way: GPU density. A rack optimized for AI training might hold 8 NVIDIA H100 GPUs drawing 10 to 30 kilowatts, compared to 3 to 5 kW for a standard compute rack. That higher density drives everything else — more heat, more cooling capacity, more power infrastructure per square foot.

Cooling Systems

Servers generate heat. Removing that heat is one of the most expensive engineering problems in a data center. Three cooling approaches are in common use:

• Air cooling: Computer room air handlers (CRAHs) circulate cool air through cold aisles, past servers, and exhaust hot air into hot aisles. Works well for rack densities below 10 kW/rack.
• Direct liquid cooling: Water or coolant runs through pipes directly attached to server components. Handles densities up to 100 kW/rack. Required for most AI training configurations.
• Immersion cooling: Servers are submerged in a non-conductive fluid. Extremely efficient but requires specialized hardware. Used in a small number of high-density AI facilities.

Cooling accounts for 15 to 20% of data center construction cost and is directly reflected in Power Usage Effectiveness (PUE). For a full breakdown of cooling methods and costs, see our data center cooling systems guide.

Power Infrastructure

A data center needs two things from its power system: reliability and redundancy. Power arrives from the utility grid, passes through transformers, and is conditioned through uninterruptible power supplies (UPS). Diesel generators provide backup power during grid outages, typically capable of running the facility for 24 to 72 hours on stored fuel. Tier III and Tier IV facilities run on redundant power paths so that no single failure interrupts service. Power infrastructure represents 40 to 50% of data center construction cost (Construct Elements, 2025).

Networking

Data enters and exits through fiber optic connections running into the facility from multiple carriers. Inside, switching equipment connects servers to each other and to the outside world. Enterprise data centers typically operate at 10 Gbps inter-server links. Hyperscale facilities use 25 Gbps to 400 Gbps links. AI training clusters require extremely low-latency interconnects between GPUs, often using NVIDIA InfiniBand or NVLink for speeds that Ethernet cannot match.

Physical Security

A data center is a high-value target. Standard security layers include perimeter fencing, biometric access controls, mantrap entry systems, continuous CCTV monitoring, and on-site security personnel. Most carrier-neutral colocation facilities maintain SOC 2 Type II compliance and ISO 27001 certification, which requires documented security controls and annual third-party audits.

The Five Types of Data Centers

Data centers serve different purposes and are built to different specs. The five main types differ in ownership, tenant structure, scale, and the workloads they support.

Enterprise Data Centers

Enterprise data centers are built and owned by a single organization for its own use. A bank, hospital, or government agency might operate its own facility to maintain direct control over hardware, security, and data sovereignty. The tradeoff is cost: building and operating a private data center requires capital investment, specialized staff, and ongoing maintenance regardless of utilization. Many organizations have shifted enterprise workloads to colocation or cloud over the past decade.

Colocation Data Centers

A colocation facility, or colo, is a third-party data center that leases space, power, and cooling to multiple tenants. The tenants bring their own servers. The colo operator provides the building, power redundancy, cooling, physical security, and network connectivity. Equinix and Digital Realty are the two largest colocation providers globally. For businesses that want physical control over their hardware without the capital cost of building a facility, colocation is the standard option.

Hyperscale Data Centers

Hyperscale facilities operate at a scale that makes most enterprise data centers look like server closets. The threshold is generally 5,000 or more servers and 10,000 square feet of floor space per building. AWS, Microsoft Azure, Google Cloud, and Meta operate the largest hyperscale footprints. According to Synergy Research Group, 130 to 140 new hyperscale data centers are built globally every year. As of 2025, there are 1,189 hyperscale facilities worldwide, with 642 of them (54%) in the United States. For a deeper look at how hyperscale infrastructure works, see our hyperscalers explained guide.

Edge Data Centers

Edge data centers are small facilities positioned close to end users, typically in metro areas, to deliver low-latency compute. A 5G tower might connect to an edge facility a few miles away. A content delivery network (CDN) uses edge locations to serve video without routing traffic back to a central data center across the country. Edge facilities typically handle under 1 MW of power and are designed for automation rather than on-site staffing.

Cloud Data Centers

Cloud data centers are operated by cloud providers and accessed entirely remotely. Users rent virtual machines, storage, and services rather than physical hardware. AWS, Azure, and Google Cloud operate cloud data centers in dozens of regions worldwide. The distinction between "cloud" and "hyperscale" is mainly conceptual: cloud data centers are hyperscale facilities viewed from the customer's perspective.

Data Center Tiers: What Tier I, II, III, and IV Mean

The Uptime Institute defines four data center tiers based on redundancy levels, uptime guarantees, and the ability to perform maintenance without taking systems offline. Tier ratings apply to the facility's physical infrastructure, not to the software or services running inside it.

Most enterprise applications can tolerate Tier II or Tier III uptime. Financial trading systems, healthcare records, and AI inference APIs that serve millions of users typically require Tier III or Tier IV. The cost difference between tiers is significant: a Tier IV facility requires twice the power and cooling infrastructure of a Tier I facility for the same compute capacity, and full 2N+1 redundancy means every critical system has two fully independent backups plus a spare.

"The Uptime Institute Global Data Center Survey 2025 reveals an innovative and resilient industry — but one that is also facing rising costs, worsening power constraints and challenges in meeting the demands for AI." (Uptime Institute, 2025)

Most large colocation providers operate Tier III facilities. Tier IV is relatively rare and typically found in government, financial, and mission-critical enterprise deployments. AWS, Google, and Microsoft do not publish Tier ratings for their cloud facilities because their distributed architecture achieves equivalent or higher availability through redundancy across multiple geographic regions rather than within a single building.

How Much Power Do Data Centers Use?

Data centers consumed 415 terawatt-hours (TWh) of electricity globally in 2024, according to the International Energy Agency. To put that in context, 415 TWh is roughly equivalent to the total annual electricity consumption of Spain. The United States accounts for the largest share, with domestic data center consumption estimated at 183 TWh in 2024.

The IEA projects that global data center electricity demand will nearly double to 945 TWh by 2030 — close to 3% of total global electricity generation. US demand is forecast to reach 426 TWh by 2030. The primary driver is AI: training and running large language models requires orders of magnitude more compute than traditional web applications, and that compute is concentrated in GPU-dense data centers with power draws that existing grids were not designed to support.

"Despite Persistent Power Constraints, Hyperscale Growth Accelerates." (CBRE Global Data Center Trends 2025, June 2025)

The Number Most Guides Don't Show

415 TWh across 12,000 data centers gives an average of 34.6 GWh per facility per year. At a commercial electricity rate of $0.07 per kWh, that is approximately $2.4 million in electricity per data center per year, on average. But averages are misleading here. A small enterprise facility drawing 500 kW continuously spends around $306,000 per year on power. A 100 MW hyperscale facility, running continuously at full load, spends approximately $61 million per year on electricity alone, before factoring in cooling overhead.

That electricity cost gap is why hyperscale operators spend billions engineering custom cooling, signing long-term renewable energy contracts, and building facilities near cheap power sources. A 1 percentage point improvement in Power Usage Effectiveness (PUE) at a 100 MW facility saves roughly $600,000 per year.

Power Usage Effectiveness (PUE)

PUE measures how efficiently a data center uses power: a PUE of 1.0 would mean all electricity goes directly to compute, with zero overhead for cooling, lighting, or power conditioning. That is physically impossible, but the industry has improved significantly. Google reports a fleet-wide PUE of 1.10. Microsoft averages approximately 1.19. Industry average across all facility types runs around 1.5, dragged upward by older enterprise facilities built before efficiency became a priority (Uptime Institute, 2025).

AI-dense facilities can temporarily show worse PUE during peak training runs because GPU racks generate heat faster than cooling systems designed for lower-density workloads can remove it efficiently.

Who Operates the World's Data Centers?

The data center industry divides into two segments: hyperscale cloud operators who build for their own use, and colocation providers who lease capacity to enterprise tenants.

Northern Virginia remains the world's largest data center market. The region added 523 MW of new capacity in a single year. Atlanta follows with 1,279.4 MW of total inventory. Outside the US, CBRE identifies Singapore, Frankfurt, Amsterdam, Tokyo, and London as the dominant international markets.

The market value of global data center infrastructure reached $527.46 billion in 2025 and is projected to reach $739.05 billion by 2030, a compound annual growth rate of 6.98% (Cargoson market analysis, 2025). Hyperscalers — AWS, Azure, Google, and Meta — own approximately 40% of the 35 gigawatts of data center capacity currently under construction in North America, per JLL North America Data Center Market Dynamics 2025.

Why Data Centers Are the Foundation of AI

Every AI model you interact with runs inside a data center. ChatGPT, Claude, Gemini, Copilot, Grok — the models behind these products were trained on GPU clusters inside hyperscale facilities, and every query you send is processed by GPU or CPU hardware in the same kind of environment.

The relationship between AI and data centers is creating a feedback loop. AI demand drives new data center construction. New data centers enable larger AI models. Larger AI models drive more user adoption. More users drive more data center demand. The Gartner forecast of $489.5 billion in global data center systems spending for 2025 reflects this cycle: that spending grew 46.8% in a single year, driven almost entirely by AI infrastructure investment.

Three workloads define the AI data center requirement:

• Training: Building a foundation model like GPT-4 or Gemini requires thousands of H100 or H200 GPUs running continuously for weeks or months. Training runs are performed inside dedicated GPU clusters that need extremely low-latency interconnects between chips. For more on the hardware that powers these clusters, see our NVIDIA H100 specs and pricing breakdown and the cloud GPU providers comparison.
• Fine-tuning: Adapting a foundation model to a specific task or domain. Smaller scale than pre-training, but still GPU-intensive. Most enterprise AI deployments involve fine-tuning rather than pre-training from scratch.
• Inference: Running a trained model to respond to user queries. Inference is the most compute-intensive workload in terms of sustained, continuous GPU utilization, because it runs 24 hours a day at scale. OpenAI serving ChatGPT to millions of users simultaneously is an inference problem, not a training problem.

The AI transition is changing the physical specifications of new data centers. Facilities built for traditional web workloads used air cooling designed for 3 to 5 kW per rack. AI training facilities now target 10 to 30 kW per rack or more, requiring liquid cooling infrastructure, heavier power delivery, and reinforced floors. The gap between traditional and AI-optimized data centers, in construction cost and operating complexity, is one of the defining infrastructure stories of 2025 and 2026.

"The scale of investment now required for AI data center builds would have seemed implausible five years ago. A single campus can cost more than the GDP of a small country." (CBRE Global Data Center Trends 2025)

Frequently Asked Questions

Related Articles