The Types of Data Centers, Side by Side

"Data center" covers everything from a locked closet behind a dentist's front desk to a campus that wants a gigawatt. Anyone who uses one word for all of it ends up comparing a corner store to a regional distribution warehouse. The useful questions are who owns it, who uses it, what work it does, and how much power it draws. Those four answers sort almost every facility into one of the types below.

The big comparison table

Sizes in this table are typical ranges drawn from the named examples and industry reporting. They are rules of thumb. No standards body defines "hyperscale" by megawatts. Treat the edges of each range as soft.

Type	What it does	Typical IT power	Typical floor area	Usual owner	Examples (as of October 2026)
On-premises enterprise	Runs one organization's own systems: payroll, records, phones, files	Under 0.1 MW to about 5 MW	A room to about 50,000 sq ft	The organization itself: a bank, hospital, university, state agency	A county government server room; a hospital system's main data center
Retail colocation	Rents racks, cages or small suites to many tenants; often hosts internet exchanges	About 1 to 30 MW per building	About 20,000 to 300,000 sq ft	Colocation companies, many organized as REITs	Equinix lists 282 data centers in 70-plus metros
Wholesale colocation	Leases whole halls or buildings to one large tenant	About 10 to 100+ MW per building	About 100,000 to 500,000+ sq ft	Developers and REITs, often backed by private equity	Digital Realty lists 300-plus data centers in 55-plus metros
Hyperscale (self-built)	Runs a giant company's own services and cloud	Tens to hundreds of MW per campus	Hundreds of thousands to millions of sq ft	Amazon, Microsoft, Google, Meta, Apple, Oracle and a few others	Synergy Research counted 1,136 hyperscale facilities at the end of 2024 (published March 2025)
Cloud region and availability zones	A cluster of hyperscale buildings arranged for fault isolation	Each AZ is one or more buildings; regions reach hundreds of MW or more	Spread over several sites	Cloud providers	AWS: 124 AZs in 39 Regions
Edge	Small sites close to users or devices for low delay or local caching	A few kW to about 5 MW	A cabinet to about 10,000 sq ft	Telecom carriers, CDNs, tower companies	Cache servers at a carrier hub; a micro site at a cell tower
AI training cluster	Tens or hundreds of thousands of tightly linked GPUs building models	100 MW to over 1 GW per campus	About 1 to 4+ million sq ft per campus	Hyperscalers, AI labs, and partners such as Oracle and Crusoe	Stargate Abilene, Texas: 1.2 GW planned, about 4 million sq ft; Microsoft Fairwater, Wisconsin: 1.2 million sq ft on 315 acres
AI inference	Answers live AI requests from users and apps	Tens to hundreds of MW, often inside cloud regions	Varies	Cloud providers, AI companies	Usually hosted in existing cloud regions near population centers
High-performance computing (HPC)	Scientific simulation, weather, weapons and energy research	About 1 to 40 MW per system	Usually one large machine room	National labs, universities, governments	Frontier at Oak Ridge, about 24.6 MW; El Capitan at Livermore, about 29.7 MW
Crypto mining	Runs specialized machines that compete to validate blockchain transactions	Tens to hundreds of MW	Often simple metal sheds or containers on large acreage	Publicly traded miners and private operators	Riot Platforms, Rockdale, Texas: 700 MW
Modular or containerized	Factory-built power, cooling or IT modules shipped and set on a pad	A few hundred kW to a few MW per module	Shipping-container to small-building size	Anyone needing speed: miners, militaries, remote sites, AI startups	Modules stacked at oil fields, mines or temporary sites

Enterprise, colocation and the slow move out of the basement

For decades most companies ran their own server rooms. That is changing. Uptime Institute's 2026 survey summary says that, for the first time, third-party facilities and services host a larger share of IT workloads than enterprise-owned data centers (counted by facility, not total capacity). A hospital may keep a small room on site for systems that must work even if the internet link fails, then rent space at a colocation building and buy cloud services for the rest.

Colocation splits into two businesses. Retail colocation is like an apartment building: lots of tenants, each with a few racks or a caged area, sharing power and cooling. These buildings are where networks meet, so they tend to sit in big metros. Wholesale colocation is like leasing a whole warehouse: one tenant, often a hyperscaler, takes 10, 30 or 100 MW at a time on a long lease. When a developer you have never heard of files plans for a 300 MW campus, it is usually a wholesale play with a big tenant behind it.

Cloud regions and availability zones

Cloud providers group buildings so that one fire, flood or power failure does not take down a whole service. AWS calls the group a Region and the separate pieces Availability Zones. As of October 2026 it describes an AZ as one or more discrete data centers with redundant power, networking and connectivity, says each Region has at least three AZs, and says AZs are physically separated by a meaningful distance while staying within 100 km of each other. Microsoft and Google use similar terms. For anyone watching a local project, this explains why a cloud company might buy three parcels several miles apart instead of one big one.

AI training versus inference

These two jobs use similar chips and very different buildings. Training builds a model by running enormous numbers of calculations across tens or hundreds of thousands of GPUs that must talk to each other constantly. The GPUs need to sit close together on very fast links, so training wants one giant campus with a huge, steady power supply. It does not care much where that campus is, because it barely talks to users.

Inference is the model at work: answering a chat, summarizing a document, generating an image. Each request is small, but there are billions of them, and users notice delay. Inference tends to live inside existing cloud regions near people and fiber. As AI products spread, a larger share of AI electricity goes to inference, which may push more AI capacity back toward metro areas.

HPC, crypto mining and modular sites

Supercomputers at national labs are the ancestors of AI clusters. Frontier at Oak Ridge National Laboratory drew about 24.6 MW in the June 2026 TOP500 list, and El Capitan at Lawrence Livermore about 29.7 MW. Those are single machines in purpose-built rooms. A modern AI training campus can draw 40 times as much.

Crypto mining sites look like data centers from the road but behave differently. The machines do one task, need little network connection, and can shut off in minutes. EIA's preliminary estimate, published in February 2024, put U.S. bitcoin mining at 0.6 to 2.3 percent of U.S. electricity in 2023. Miners often sign deals to curtail during grid emergencies and get paid for it. Several mining companies have converted sites to AI hosting since 2024, because they already hold what AI builders want most: an energized grid connection.

Modular or containerized data centers are built in a factory as boxes of power gear, cooling or racks, then trucked to a site. They trade some efficiency and flexibility for speed. You see them at remote energy sites, military bases and temporary deployments, and as add-ons at bigger campuses.