Data center types: enterprise, colocation, hyperscale, and edge

What a data center type is, and why it sets everything else

A data center type is the category a facility falls into based on who owns and uses it and how big it is. The type is not a label you stick on at the end. It is the decision that sets the size, the redundancy, the design, and the way the place is operated and paid for, and it is made before the first drawing. Call a facility enterprise, colocation, hyperscale, or edge and you have already said a great deal about its power, its tenants, its uptime target, and who is standing in the white space at two in the morning.

This guide is the map. It sorts the types, says what each one is and where it fits, and points to the deeper guides where a type earns its own walkthrough. The edge and micro data center guide covers the small, distributed end in full. The white space and gray space guide covers how any of these buildings is zoned inside. Here the job is the classification itself: the two axes, the types along each, and how to pick the one that matches the need.

Get the type wrong and everything downstream is wrong with it. A business that builds its own hall when it should have leased a cabinet, or rents cloud when it needed a controlled room on its own floor, pays for that mismatch for years.

The two ways data centers are classified

Two questions sort data centers, and mixing them up is the most common confusion in the whole subject. The first is who owns and operates the place and who uses it. That axis gives you enterprise, colocation, cloud, and hyperscale, an operating-model split. The second is how big the facility is and where it sits, which spans a single edge box near the user up to a hyperscale campus measured in hundreds of megawatts. That axis is about scale and location.

The two axes are not the same list, and they overlap on purpose. Hyperscale shows up on both, because it names an operating model, a self-building cloud-scale operator, and a scale, the largest sites going. Edge names a location and, usually, a small size. A colocation facility can be retail or wholesale, can be modular in how it was built, and can house an AI hall inside it. A type is rarely one pure thing.

So read any facility on both axes at once. Ask who owns it and who uses it, then ask how big it is and where it lives. The answer to both, together, is the actual type.

The enterprise data center

An enterprise data center is a facility a single organization builds and runs for its own IT, not to sell space to anyone else. It is the traditional corporate data center: the company owns or leases the room, owns the gear, and operates it for its own applications. A bank's processing center, a hospital system's clinical data room, a manufacturer's plant-floor compute, all enterprise.

This is the single-tenant, owner-operated end of the field. The organization carries the whole cost and the whole risk, and it gets full control in exchange. Control is the reason an enterprise room survives in an era when most new capacity goes elsewhere: regulated data, latency to the business, a workload that has to live on the company's own floor, or gear too specialized to hand to a landlord.

The trend has been away from it. The enterprise share of US server energy fell hard over the last decade as workloads moved to colocation and hyperscale, by some measures from well over half in the mid-2010s to roughly a tenth by 2023. The enterprise room did not vanish. It got more specific. What stays on-premises now is there for a reason, and what is not has already moved out.

What is a colocation data center?

A colocation data center, or colo, is a building a provider owns and operates to rent space, power, and cooling to multiple customers who bring their own IT gear. The provider runs the shell and the plant: the power chain, the cooling, the security, the connectivity. The tenant racks its own servers and keeps control of them. It is the rent-the-room, own-the-gear model.

The unit of sale is space and power. A cabinet, a locked cage of several cabinets, or a private suite, sold with a power allotment and a cooling commitment and wrapped in a service-level agreement. The SLA is the product. It states the uptime the provider commits to, the power and temperature it will hold, and what happens when it misses. Colocation is the most common category of commercial facility worldwide, numbering in the thousands of sites across more than a hundred countries.

The multi-tenant nature drives the whole design. Many customers share one building without sharing access to each other's gear, which sets the cage-and-cabinet security model and the way the white space is zoned. The white space and gray space guide covers that layout in full. The colo split into retail and wholesale is the next section.

Retail versus wholesale colocation

Colocation divides into retail and wholesale, and the line is roughly the size of the footprint. Retail colocation sells smaller units, cabinets and cages, to many tenants in a shared hall. Wholesale sells large dedicated blocks, whole data halls or megawatt-scale capacity, to a few big customers. A common rule of thumb puts the break around ten cabinets or about 100 kW: under that is retail, over it is wholesale, though providers draw their own lines.

The two read differently as a business. Retail is a packaged service: a flat monthly rate that bundles the rack, the power, the cooling, and often remote-hands support, on a short term that can run month to month. It suits a small or mid-size company that needs a few cabinets in a carrier-dense building. Wholesale is closer to a long lease: large space, power metered and billed by what the tenant draws, on terms that run five years or more, with the tenant handling most of its own operations. Its customers are the cloud providers, content networks, and large enterprises taking down whole halls.

Pick by footprint and term, not by name. A growing tenant can outgrow retail packaging and find wholesale economics, while a wholesale-scale need dressed up as retail cabinets costs more than it should.

What is a hyperscale data center?

A hyperscale data center is a very large facility built and run by a cloud or platform operator to serve massive, scalable workloads, the kind operated by companies like AWS, Google, Microsoft, and Meta. The defining traits are size, standardization, and automation. One design gets repeated across many halls and many sites, built to a template, run by software, and staffed thin for the load it carries.

Scale is the headline. The industry commonly draws the hyperscale line somewhere around 100 MW of capacity and up, with campuses reaching several hundred megawatts and gigawatt-scale sites now being planned. These are mostly self-built: the operator designs the facility around its own hardware and operating model rather than renting a generic hall, though hyperscalers also lease wholesale capacity when speed matters more than ownership.

Hyperscale is where most of the growth is. The count of operational hyperscale facilities roughly doubled over five years to more than a thousand by the end of 2024, with the US holding around half of global capacity. Hyperscale plus colocation now accounts for the large majority of server energy use, a share still climbing. When people say the data center boom, this is mostly what they mean.

Cloud and managed data centers

Cloud and managed facilities are sorted by what they deliver rather than how big they are. A cloud data center is a facility a provider runs to deliver computing as a service, infrastructure, platform, or software, that customers consume over the network and never see. Most large cloud facilities are hyperscale, so cloud and hyperscale overlap heavily, but the words point at different things. Hyperscale is the scale and build model. Cloud is the service delivered out of it.

Managed services sit alongside. A managed service provider, or MSP, runs the gear and the operations on the customer's behalf, whether in the provider's facility, in a colo, or even on the customer's own floor. The customer offloads the operating burden and keeps the application. The line between colocation, managed hosting, and cloud is a spectrum of how much the provider runs versus how much the customer does, racking your own box in a colo at one end and consuming a service with no hardware in view at the other.

The thing to hold onto is the consumption model. Cloud is capacity you rent by the hour and never own. That changes the economics and the trades involved, which the operating-model and cost sections cover.

What is an edge data center?

An edge data center is a small facility placed close to where data is created or used, instead of in a distant central hall, so the compute answers with low latency. Edge sites run from a single ruggedized rack up to a few cabinets, often unstaffed and managed remotely, holding a few kW to tens of kW. The point is to put the processing near the user, the machine, or the sensor.

Micro data center is the closely related term for the smallest self-contained units, a sealed cabinet with its own power, cooling, and monitoring built in. Most edge deployments are micro-scale by definition. They show up at retail stores, factory floors, cell tower bases, and hospitals, anywhere a round trip to a regional cloud is too slow or too expensive to backhaul.

This guide keeps edge brief on purpose, because it has its own deep walkthrough. The edge and micro data center deployment guide covers the forms, the lights-out operation, the cooling and power in one box, the connectivity and security, and the commissioning that proves it. Read that one for the build. This section just places edge on the map of types.

Modular and prefabricated data centers

Modular, or prefabricated, describes how a data center is built rather than who uses it, so it cuts across the other types. A modular data center is engineered, integrated, and tested in a factory, then shipped to site as finished modules for final assembly and commissioning. Instead of building everything in place, you build the power module, the cooling module, and the IT module in a controlled plant and bolt them together on a pad.

The forms run by size. The containerized POD packs racks, power, and cooling into a standard 20 or 40 foot container that arrives ready to run, a term that has become industry shorthand. Above that sit larger prefabricated skids and rooms that assemble into a full hall. The same approach serves an edge box in a parking lot and a hyperscale hall on a greenfield campus, which is why modular is a delivery method, not a scale.

Speed is why it has taken over so much new work. Factory build runs in parallel with sitework, so a modular project can stand up in months where a ground-up build takes far longer, and a containerized AI pod can land in weeks when the equipment lead times allow. For AI capacity racing to come online, that schedule is often the deciding factor.

The scale spectrum, edge to hyperscale

The types also line up as a size map, smallest distributed unit to largest campus. Reading the spectrum tells you roughly what power, footprint, and operating model to expect before you know anything else about a site. Treat the numbers as industry ranges, not hard cutoffs. Operators draw their own lines and the bands overlap.

At the small end sits the edge or micro unit, a few kW to tens of kW near the user. Enterprise rooms cover a wide middle, a server closet up to a few megawatts for a large corporate facility. Colocation runs from hundreds of kW for a retail footprint up to tens of megawatts for a wholesale building. Hyperscale tops the spectrum, commonly 100 MW and up, with campuses in the hundreds of megawatts and gigawatt-scale sites on the drawing board. The AI hall is reshaping these bands from the inside, packing far more power into the same floor than the band ever assumed.

The AI and GPU data center, the new type

The AI data center is the type that did not have a name a few years ago and now drives the conversation. It is a facility built for AI and high-performance computing, training and inference, where the racks are packed with GPUs and the power and cooling are an order of magnitude denser than a traditional hall. People call it the AI factory, and the name fits. It exists to turn power into trained models and served inference.

Density is what sets it apart. A conventional rack might draw a handful of kW. A current AI training rack can draw well over 100 kW, with single reference designs landing in the 120 to 130 kW range, and the heat flux at the chip is far past what moving air can carry away. That is why liquid cooling has gone from exotic to default for AI gear. Industry forecasts have the majority of AI servers shipping with liquid cooling within the next year or two, up from a small fraction in 2024.

The AI hall can live inside any of the other types. A hyperscaler builds it at campus scale, a colo fits out a high-density suite for it, and a specialist operator builds nothing but. The high density resets the design assumptions for power, cooling, floor loading, and redundancy no matter whose building it sits in.

Redundancy expectations by type

Redundancy is the other thing the type sets, because each type carries a different uptime expectation and a different way of meeting it. The common reference is the Uptime Institute Tier system, which rates a facility's concurrent maintainability and fault tolerance from Tier I to Tier IV. The TIA-942 standard offers a parallel rating for the facility infrastructure. Both are owner and certification frameworks, not building codes, so the project requirement controls what gets built.

A retail colo and a wholesale hall commonly target the concurrently maintainable level, often described as Tier III, because the SLA promises tenants they can keep running through maintenance. An enterprise room is sized to its own business case and can land anywhere. The interesting shift is at the top. Hyperscale and large AI operators often build physical redundancy leaner per hall and put the resilience in software, spreading a workload across many sites so any one hall can fail without taking the service down. The edge unit may run lower physical redundancy still, with the network providing the fallback. Match the redundancy to where the risk is actually carried, not to a tier number for its own sake.

How the design differs by type

The white space, the power chain, the cooling, and the security all change with the type, which is why you cannot copy one building's design onto another. Hyperscale leans on standardization: one repeated hall design, high-density power distribution, and a layout tuned for the operator's own hardware, built to be assembled fast and run by software. The colo is built for many tenants at once, so the floor is zoned into cages and suites, the power is metered per tenant, and the security separates one customer from the next.

The enterprise room is built around one organization's applications, often older, often mixed, and rarely as dense or as standardized as a purpose-built operator hall. The edge box compresses the whole facility into one enclosure. The AI hall throws out the density assumptions entirely, with liquid cooling and reinforced floors that a traditional air-cooled hall never needed.

The mechanics of the layout itself, the data hall versus the support plant, the aisles, the raised floor or slab, are covered in the white space and gray space guide. The point here is that the type drives those choices. A design that is right for a wholesale colo is wrong for a hyperscale self-build, and both are wrong for a closet at the edge.

The operating model: own, lease, or consume

Underneath the types is a simpler split: do you own the facility, lease space in someone else's, or consume capacity as a service. That choice is mostly a capex-versus-opex decision and it follows the type. Enterprise is own-and-operate, a capital build the organization carries on its own books and runs with its own staff. Colocation is lease: the provider owns the building and the plant, the tenant rents space and power and runs its own gear, turning a capital build into a recurring cost. Cloud is consume: no building, no gear, just capacity rented by the hour.

The further you move from owning toward consuming, the less you control and the less you carry. Owning gives full control and full risk. Consuming gives speed and flexibility and hands the risk to the provider. Leasing sits between, the tenant keeping its hardware and handing off the building.

Most organizations end up with a mix rather than one pure model, which the selection and hybrid sections cover. The operating model matters here because it decides who builds, who maintains, and who the trades actually work for on a given site.

Security and multi-tenancy

Security design splits hard on one question: is the building single-tenant or multi-tenant. A single-tenant enterprise or hyperscale facility serves one organization, so the security boundary is the building perimeter and the access control is internal policy. A multi-tenant colo serves many customers in one building, so it has to keep each tenant out of every other tenant's gear while letting them all into the same hall.

That multi-tenancy is what drives the cage, the locked cabinet, and the layered access in a colo. A customer badges through the perimeter, into the data hall, and then into its own cage or to its own cabinet, and never has a path to anyone else's equipment or to the gray-space plant. The provider's staff reaches the plant without ever touching a tenant's racks. The zoning that makes this work is part of the white space and gray space layout.

The edge site flips the problem. It is usually unstaffed and physically exposed, sitting in a back room or a yard, so its security leans on the enclosure, remote monitoring, and tamper detection rather than on guards and badge tiers. Match the security model to the tenancy and the staffing. A colo's layered access is overkill at the edge, and a single lock is not enough for a shared hall.

What drives the location of each type

Where each type gets sited follows from what it is for, and the drivers pull in different directions. Hyperscale chases power, land, and climate. A campus drawing hundreds of megawatts goes where it can get cheap, plentiful, and increasingly clean power, a large parcel of land, and a climate or water situation that makes cooling affordable, which is why these sites cluster in a handful of regions far from any city.

Edge does the opposite. It chases proximity to the user or the machine, because the whole reason it exists is low latency, so it sits in the metro, at the tower, on the factory floor, near where the data is made. Colocation lands in carrier-dense metro markets, near the network interconnection and the customers who need it. The enterprise room sits near the business it serves, often on a corporate campus or close to the operation that depends on it.

So the same question, where do you put it, has a different answer for every type. Power and land for hyperscale, latency for edge, interconnection for colo, the business for enterprise. Siting a facility against the wrong driver is an expensive mistake to discover after the slab is poured.

The cost and business model by type

The money works differently across the types, and the difference is bigger than most line items. An enterprise build is a capital project: the organization pays for the building, the plant, and the gear up front, then carries the operating cost, and the facility is an asset and a liability on its own books. Colocation converts that into a recurring charge. Retail colo bundles space, power, and cooling into a flat rate per cabinet, while wholesale meters the power and bills the tenant for what it draws on a long lease.

Hyperscale economics run on self-build at scale. The operator standardizes the design, buys power and equipment in enormous volume, and drives the cost per unit of compute down in a way no smaller builder can match. That scale advantage is part of why workloads keep moving to hyperscale and colo and away from owned enterprise rooms. Cloud is the pure opex case, capacity rented by the hour with no asset at all.

The pricing question for a buyer is rarely just the sticker. It is the total over the life of the workload, against the control and the risk you are willing to carry. A cheap cloud bill that runs for years can cost more than a building you own, and a building you own can sit half-empty. The cost model and the type are the same decision.

How AI is reshaping the mix of types

AI demand is the force bending the whole field right now, and it is changing the mix of types more than any trend in years. The hyperscalers are building AI campuses at a scale that pulls most of the new capacity toward the top of the spectrum, and they are doing it fast enough that schedule, not cost, often drives the decision. That is why modular and prefabricated build has surged for AI work. A pod that lands in weeks beats a hall that takes years.

A new category has grown up alongside the hyperscalers. The neocloud, a specialist operator renting GPU capacity as a service, GPU-as-a-service, has moved from a niche to a real part of the market, building dense, liquid-cooled facilities tuned for nothing but AI and HPC workloads. They compete with the hyperscalers on flexibility and regional placement rather than on raw scale.

AI is also pushing compute back toward the edge for inference, because a model often has to answer where the data is. So the same trend grows both ends at once: enormous training campuses at the top and small inference units at the edge. The density it brings, and the liquid cooling that density forces, is the thread running through all of it.

Choosing own, colo, cloud, or hybrid

For an organization deciding where to put its workload, the choice is the old make-versus-buy-versus-rent question in a new form: build your own enterprise room, lease colocation space, consume cloud, or mix them. Start from the workload, not the building. What has to stay on your own floor for regulation, latency, or control goes to enterprise or a private cage. What needs to scale fast or run anywhere goes to cloud. What sits between, steady gear you want to own but a building you do not, goes to colo.

The honest decision weighs control against speed and capital against operating cost. Owning gives the most control and ties up the most capital. Cloud gives the most speed and the least control, and the bill never stops. Colocation keeps your hardware and hands off the building. Most organizations land on more than one of these, because few workloads all want the same answer.

Match the type to the need, run by run, the way you would size a feeder to the load. The expensive errors are the mismatches: building a hall for a workload that belonged in the cloud, or renting cloud for a steady base load that an owned room would have carried for less.

The hybrid and multi-data-center strategy

Most estates of any size end up hybrid, running some workloads on their own enterprise floor, some in colocation, and some in cloud, because no single type fits everything an organization does. The steady, regulated, or latency-bound work stays close and owned. The bursty or experimental work goes to cloud. The middle, owned gear that does not need an owned building, goes to colo. The strategy is not a failure to choose. It is choosing per workload.

Spreading across multiple facilities also buys resilience. A workload split across two colos or across cloud regions survives the loss of any one site, the same software-resilience pattern the hyperscalers use, applied by a smaller operator. The cost is complexity. More places to manage, more network between them, more boundaries to secure.

The discipline is to keep deciding deliberately rather than letting the mix sprawl. Workloads drift to whatever was easy to stand up, and an unmanaged hybrid turns into paying for capacity in three places when one would have carried it. Review the placement on a cadence, the way you would review any portfolio.

What each type means for the trades

For the people who build and run these places, the type decides the job. A hyperscale campus is a mega-build: huge electrical and mechanical scope, repeated halls, heavy rigging, and an integrated commissioning effort that runs for months, with the same details repeated enough that the crews get fast. A colocation fit-out is a different rhythm, a tenant build-out inside an existing shell, racking, power whips, cooling tie-ins, and the cage work, often on a tenant's schedule and inside a live building with other customers next door.

The edge deploy is its own trade. The box is built and tested in a factory, so the field work is the pad, the power and network drop, the set, and the commissioning, often at a remote or unstaffed site with no one to call. Modular work splits between the factory floor, where the integration happens, and the site, where the assembly and final commissioning happen.

The AI hall has rewritten the mechanical scope across all of them. Liquid cooling means plumbing the white space, coolant distribution units, manifolds, and leak detection that an air-cooled hall never carried, and the electrical scope grows with the density. A crew that knows air-cooled halls is learning new work fast.

What to document when you classify a facility

When you sort a facility or plan one, write down what type it is and why, because the type carries the assumptions that everything downstream depends on. The record is what keeps a project from drifting, a colo fit-out designed as if it were an enterprise room, or an AI hall sized with air-cooled numbers.

Capture the type on both axes, who owns and operates it and who uses it, plus the scale in power and footprint, the redundancy target and the framework behind it, the operating model, and the one trait that defines this facility against the others. If it is multi-tenant, note the tenancy model and the security boundary. If it is AI or high-density, note the cooling method and the rack power, because those are the numbers that break the usual design.

Common mistakes

• Confusing the ownership types with the scale types, and treating hyperscale as only a size or only an operating model when it is both.
• Treating every data center the same, copying one type's design, redundancy, or cost model onto another.
• Mismatching the type to the need, building an enterprise hall for a workload that belonged in cloud, or renting cloud for a steady base load an owned room would carry cheaper.
• Ignoring the emerging types, planning as if AI and edge were edge cases when AI density and edge placement now drive the market.
• Setting the wrong redundancy for the type, paying for Tier IV where software resilience across sites was the real answer, or running an edge box too lean for what depends on it.
• Sizing an AI or GPU hall with air-cooled power and floor-load assumptions, then discovering the density needs liquid cooling and reinforced floors after the design is set.
• Reading a facility on one axis only, so a wholesale colo gets confused with a hyperscale self-build, or a micro unit with a full enterprise room.

Field checklist

Standards and references

The classifications in this guide are industry conventions, not a single published standard, so treat the names and the scale numbers as the way the market talks rather than as fixed definitions. Different operators and analysts draw the lines at different places, and the power bands shift as the technology moves. Where a number matters, confirm it against the specific project and the operator's own definitions.

Two frameworks do carry weight on the facility side. The Uptime Institute Tier system rates concurrent maintainability and fault tolerance, Tier I through Tier IV, and is the common language for redundancy expectations, especially in colocation SLAs. The TIA-942 standard covers data center facility infrastructure and offers a parallel rating. ASHRAE Technical Committee 9.9 publishes the thermal guidelines that govern the operating environment, and BICSI covers the cabling and telecommunications infrastructure. None of these defines the data center types themselves. They govern how a facility of any type is built and run.

The scale figures, the hyperscale 100 MW line, the retail-versus-wholesale break, the AI rack power, come from industry reporting and shift every year. Cite them as current industry ranges and verify against the project before you design to them.

Units, terms, and conversions

The same facility gets described in a few different vocabularies across a real-estate listing, an engineering drawing, and an operator's own terms, so the words shift even when the thing does not.

Capacity is given in megawatts (MW) of IT or critical load, sometimes in kilowatts (kW) for smaller sites and gigawatts (GW) for the largest campuses. Footprint shows up in square feet or square meters of white space, and increasingly in power density, kW per rack or watts per square foot, because the power, not the floor area, is what now caps a building. Match the term to the source, and confirm whether a power figure is IT load, critical load, or total facility draw, because those are different numbers.

Enterprise data center

A facility an organization owns and runs for its own IT, single-tenant and owner-operated

Colocation (colo)

A multi-tenant facility renting space, power, and cooling to customers who bring their own gear

Hyperscale

A very large, standardized, self-built facility run by a cloud or platform operator

Edge / micro

A small facility placed near the data for low latency, often unstaffed and remotely managed

Hyperscale line

The industry's rough threshold for hyperscale, commonly around 100 MW and up

Neocloud / GPUaaS

A specialist operator renting GPU capacity as a service for AI and HPC workloads

Uptime Tier

The Uptime Institute rating of facility redundancy, Tier I through Tier IV

Related tools

Calculators and readiness checks for this work

Compare your options

FAQ

People also ask

• What is an edge data center?
• What is white space in a data center?
• What is colocation?
• What kind of batteries do data centers use?
• What is a rack unit (U)?