How Do Data Centers Work?

By Adam Patti, Chief Executive Officer, VistaShares | November 2026

A data center is a facility that houses computer servers, storage and networking equipment, along with the power and cooling systems needed to keep them running continuously. It takes in electricity and data connections, converts that power into computing work, and removes the heat the equipment produces. AI workloads are changing data center design because they concentrate far more computing power, and heat, into each rack.

What are the main parts of a data center?

Every data center, from a small enterprise server room to a large campus, combines the same basic systems.

Compute. Servers perform the processing. In traditional data centers, most servers use general-purpose processors. In AI data centers, servers are built around accelerators, such as graphics processing units, designed for the parallel calculations that AI models require.

Storage. Storage systems hold the data that applications and models use, ranging from fast solid-state drives to high-capacity storage for archives.

Networking. Switches, routers and cabling connect servers to one another and to the outside world. Fiber connections link the facility to telecommunications networks and to other data centers.

Power. Electrical systems bring utility power into the building and deliver it to each piece of equipment, with backup systems to keep it running during an outage.

Cooling. Cooling systems remove the heat produced by the equipment so that it can operate within safe temperature limits.

Physical security and monitoring. Access controls, fire suppression and building management systems protect the equipment and track its condition.

How does power flow through a data center?

Electricity typically arrives from the utility grid at high voltage, often through a dedicated substation. Transformers step the voltage down, and switchgear routes power through the facility.

Before reaching the servers, power usually passes through an uninterruptible power supply, which uses batteries or other storage to bridge brief interruptions. If utility power fails for longer, backup generators start and carry the load. Power distribution units then deliver electricity to individual racks.

Reliability is central to the design. Facilities are often built with redundant components, so that a single failure does not interrupt service. The Uptime Institute’s Tier classification, from Tier I to Tier IV, is a widely used framework for describing how much redundancy and fault tolerance a facility’s infrastructure provides.¹

A common efficiency measure is power usage effectiveness, defined by The Green Grid as the total energy used by the facility divided by the energy used by the computing equipment.² A value closer to 1.0 means less energy is spent on cooling and other overhead relative to the computing itself.

How are data centers cooled?

All of the electricity used by computing equipment ends up as heat. Removing that heat is one of the main engineering tasks in any data center.

Air cooling. Traditional facilities blow cool air through rows of racks and capture the warmed air for cooling, often using a hot-aisle and cold-aisle layout. Chillers, cooling towers or outside air remove the heat from the building.

Liquid cooling. Liquids carry heat far more effectively than air. In direct-to-chip cooling, liquid flows through cold plates attached to processors. In immersion cooling, servers are submerged in a non-conductive fluid. Liquid cooling allows much higher power density per rack.

Heat rejection. Whatever the method inside the building, the heat must ultimately be released outside, through cooling towers, dry coolers or other systems. Some designs use water, while others are designed to minimize water use.

How are AI data centers different?

AI data centers use the same building blocks but push them much harder.

Higher density. Racks filled with AI accelerators draw much more power than racks of conventional servers. That increases the power delivered to each part of the building and the heat that must be removed.

Liquid cooling as a requirement. At the densities used for AI training, air cooling alone often cannot remove enough heat. New AI facilities are increasingly designed for liquid cooling from the start.

Tightly connected clusters. Training large models requires thousands of accelerators to work together as a single system. That requires high-bandwidth, low-latency networking across the cluster, with large volumes of optical connections.

Training and inference. Training facilities are built to run very large clusters for extended periods and can be located where power and land are available. Inference, which runs trained models to answer user requests, may benefit from being closer to users and is spread across more locations.

What types of data centers are there?

Hyperscale data centers are large facilities built and operated by major cloud and internet companies for their own services.

Colocation data centers are operated by companies that lease space, power and cooling to multiple customers, who install their own equipment.

Enterprise data centers are owned and operated by a single organization for its own use.

Edge data centers are smaller facilities placed near users or devices to reduce latency.

Many large projects combine models. A cloud provider might own some campuses outright and lease capacity in others from colocation developers.

Where do data centers get built?

Site selection balances several factors: the availability and cost of power, the time needed to connect to the grid, land, water, fiber connectivity, tax and permitting conditions, and proximity to users. For large AI facilities, power availability has become the deciding factor in many cases.

That gives regions that welcome data centers, and can deliver power, an advantage. When one location faces grid constraints, local opposition or slower permitting, projects can move to another, either within the United States or abroad. Governments in the Middle East, Europe and Asia have announced programs to build domestic AI computing capacity, which adds to the range of potential locations.³ Our article on the AI supply chain describes these programs and the layers of suppliers involved.

Communities have also raised concerns about the effect of large facilities on electricity bills, water use and land use, and policymakers in several U.S. states are weighing rules on how data centers pay for grid upgrades.

What are the risks for investors in data center infrastructure?

Data center development and the businesses that supply it involve significant capital and long timelines. Key risks include delays in grid connections and equipment delivery, cost overruns, changes in customer demand or technology that alter the type of capacity needed, reliance on a concentrated group of large tenants, financing conditions, and regulatory or community opposition. Investments involve risk, including possible loss of principal.

Frequently asked questions

What is the difference between a hyperscale and a colocation data center?

A hyperscale data center is built and operated by a large cloud or internet company for its own use. A colocation data center is run by a company that leases space, power and cooling to multiple tenants, who supply their own servers. Cloud providers use both, building some capacity and leasing the rest.

What does PUE measure?

Power usage effectiveness compares the total energy a data center uses with the energy consumed by its computing equipment. A value of 1.0 would mean all energy goes to computing, with none used for cooling, lighting or power conversion. Lower values indicate a facility spends less energy on overhead.

Why do AI data centers need liquid cooling?

AI accelerators consume more power and produce more heat in a smaller space than conventional servers. At those densities, moving air cannot remove heat quickly enough. Liquid, which carries heat much more effectively, can be circulated directly to the chips, allowing racks to run at higher power levels.

What happens when a data center loses utility power?

Uninterruptible power supplies take over immediately, using batteries or other storage to keep equipment running. If the outage continues, backup generators start and carry the load until utility power returns. Facilities designed for high reliability include redundant paths so that a single equipment failure does not cause downtime.

What is the difference between AI training and inference facilities?

Training builds a model by processing large datasets on very large, tightly connected clusters, often for weeks at a time, so it can be sited wherever power and land are available. Inference runs a trained model to respond to requests, uses smaller clusters, and may be placed closer to users to reduce delay.

This material is provided for educational purposes only. It is not investment, legal or tax advice, and it is not a recommendation or an offer to buy or sell any security. Information is drawn from sources believed to be reliable as of the date of publication, is not guaranteed as to accuracy or completeness, and may change without notice. Figures attributed to third parties are their estimates. Projections and forecasts are not guarantees of future results. Investors should consult their own financial, legal and tax professionals before making investment decisions. Investments involve risk, including possible loss of principal. Companies in a single industry or theme may be affected by common economic, regulatory or technological developments and may be more volatile than the broad market. References to companies, governments or programs are for illustration only.