How Data Centers Actually Make Money

A data center can cost billions of dollars before its first customer workload begins running. The land, electrical connection, substations, generators, cooling equipment, fiber routes, security systems, and building must all be financed and constructed in advance.

Yet the operator usually does not make money by selling servers.

It makes money by converting physical infrastructure into dependable, contract-backed capacity. Customers pay for a combination of space, electrical power, cooling, connectivity, security, and operational reliability. Large facilities can then add interconnection services, installation work, remote technical support, and other recurring products.

The result is closer to a mixture of commercial real estate, utility infrastructure, telecommunications, and subscription revenue than to an ordinary technology-product business.

Understanding that combination explains why power availability can matter more than empty floor space—and why a full data center can still produce a disappointing return if it was financed or priced badly.

Three ways companies obtain data-center capacity

The first distinction is between owning computing equipment and owning the facility around it.

A hyperscaler such as Microsoft, Amazon, Google, or Meta may build and operate its own campus. In that case, the data center supports the company’s cloud or digital services rather than earning rent from outside tenants.

But technology companies also lease capacity from specialist operators. Those arrangements generally fall into two broad categories.

Retail colocation

A colocation customer installs its equipment inside a cabinet, cage, or private suite operated by another company.

The customer owns or controls the servers. The operator supplies the building, electrical distribution, cooling, physical security, network access, monitoring, and on-site operations.

The customer may begin with a few cabinets and expand over time. Contracts can combine charges for reserved space, power allocation, power consumption, connections, installation, and support services.

Equinix describes infrastructure offerings that are generally billed according to the space and power consumed by a customer. Its products include secure cabinets and private cages, along with interconnection and managed-infrastructure services (see Sources, [1]).

Wholesale and hyperscale leasing

A much larger customer may reserve an entire data hall, building, or campus phase.

These deals are often expressed in megawatts rather than individual cabinets. The operator may construct capacity for a particular customer and recover the investment through a long lease.

Digital Realty reports that leases providing one megawatt or more of power generally have terms of five to more than ten years (see Sources, [2]).

A long contract can provide predictable future rent. It can also create customer-concentration risk: losing or repricing one large lease may affect a property more than losing one cabinet-sized customer.

Owner-operated hyperscale campuses

A hyperscaler that owns its facility avoids paying an outside landlord, but it must fund and manage the complete project itself.

That includes construction, power procurement, cooling, network infrastructure, maintenance, staffing, and the risk that demand changes before the facility earns an acceptable return.

The decision to own or lease therefore depends on capital, speed, location, control, technical requirements, and the availability of suitable external capacity.

The data-center revenue stack

A data-center operator can earn several types of revenue from the same customer relationship.

| Revenue layer | What the customer receives | | ---------------------- | ---------------------------------------------------------------------------------- | | Space and capacity | A cabinet, cage, suite, data hall, building, or contracted power allocation | | Electrical service | Power delivery, backup systems, metering, and facility electrical infrastructure | | Cooling | Heat removal and environmental control for the customer’s equipment | | Interconnection | Physical or virtual connections to networks, clouds, partners, and other customers | | Installation | Initial cabling, equipment setup, and deployment work | | Operational support | Monitoring, maintenance access, and remote technical assistance | | Managed infrastructure | Additional services that help operate the customer’s deployment |

Some charges recur monthly. Others occur when equipment is installed, changed, or supported.

The attraction of the model is not one large payment when the building opens. It is the possibility of collecting contract-backed revenue repeatedly while customers continue operating inside the facility.

Why recurring revenue matters

Data centers have high fixed costs.

The operator must finance the site, building, cooling system, generators, electrical equipment, and network infrastructure whether a particular cabinet is occupied or not. Depreciation, property costs, maintenance, security, and staffing continue even when utilization is weak.

Recurring contracts help match those long-lived assets with continuing cash inflows.

Equinix reported that recurring revenue represented more than 90% of its total revenue during each of the previous three years. Its customers were generally billed monthly under contracts lasting one to five years, followed by one-year renewal periods.

For 2025, Equinix reported approximately $8.739 billion in recurring revenue. Of that amount, approximately $6.475 billion came from colocation and $1.655 billion from interconnection (see Sources, [1]).

Those figures belong to one operator and should not be treated as an industry-wide margin or revenue mix. They do, however, demonstrate how a major data-center platform can combine physical capacity with a meaningful second stream of connectivity revenue.

Power becomes inventory

In an ordinary warehouse, the central inventory measure may be square feet.

In a data center, floor space without usable electrical capacity has limited commercial value. A customer cannot operate a dense cluster of servers simply because a room is empty.

The operator needs enough power from the grid, substations and switchgear capable of delivering it, backup systems, and cooling equipment capable of removing the resulting heat.

This makes sellable capacity a combination of several constraints:

• Available utility power
• Electrical-distribution capacity
• Cooling capacity
• Floor space
• Structural and rack-density limits
• Network connectivity
• Redundancy and reliability requirements

Digital Realty explicitly states that its estimate of square feet available for lease considers factors including available power, support space, and common areas (see Sources, [2]).

That is why power can behave like inventory. A facility may own land and buildings but still be unable to sign another large customer until additional megawatts can be delivered.

The grid constraint discussed in Why AI Data Centers Have a Power Problem Nvidia Can't Solve is therefore also a revenue constraint. A delayed substation or transmission connection can postpone the moment when an operator can turn construction spending into billable capacity.

Interconnection creates a second business

Customers place equipment in a data center partly because of what they can connect to.

A bank may need access to market-data providers. A software company may need several cloud platforms. A content company may need internet exchanges and network carriers. An enterprise may need private connections between its own infrastructure and external services.

The operator can sell physical cross-connects, network ports, and virtual connectivity between these participants.

This creates a different economic layer from renting space.

A cabinet occupies a physical footprint. An interconnection service uses the data center’s position as a meeting point between customers, networks, and cloud providers. Once a facility contains a useful collection of participants, that ecosystem can attract additional customers.

Equinix describes some customers as “magnets” because their presence can draw other businesses into the same facilities. It also states that a more balanced customer base can help generate additional interconnection revenue (see Sources, [1]).

This suggests a network effect: the value of joining a facility can increase when more relevant organizations are already present.

The effect is not automatic. It depends on location, customer mix, available networks, reliability, and the cost of connecting. But it helps explain why two buildings with similar power and floor space may not have equal commercial value.

Utilization determines whether capacity earns money

A completed data hall is not the same as a profitable data hall.

Operators commit capital before all capacity is necessarily leased. Revenue improves as more cabinets or megawatts become billable, while many property and operating costs already exist.

Equinix reported a cabinet-utilization rate of approximately 77% at the end of 2025. The company defines that metric as billed cabinet space divided by total cabinet capacity (see Sources, [1]).

That percentage is specific to Equinix and is not an industry benchmark. Its importance is conceptual: utilization connects constructed capacity to current revenue.

There are also several meanings of "sold" capacity.

A customer may sign a contract before the data hall is ready. The space may be contracted but not yet generating revenue. It may be physically complete but awaiting equipment installation. It may have servers installed but operating below the power level originally reserved.

Investors and customers should therefore distinguish between:

• Capacity planned for future construction
• Capacity currently under construction
• Capacity completed and available
• Capacity contractually committed
• Capacity billed and producing revenue
• Power actually consumed by operating equipment

Treating these categories as interchangeable can create an exaggerated impression of near-term earnings or power demand.

PUE affects operating economics

Power Usage Effectiveness, or PUE, compares total data-center energy with the energy delivered to IT equipment.

The relationship is:

PUE = total facility energy ÷ IT equipment energy

A PUE of 1.5 means that delivering one unit of energy to computing equipment requires another 0.5 units for cooling, power distribution, lighting, and other facility overhead.

A result closer to 1.0 indicates that a larger share of the facility’s electricity reaches the IT equipment. The Green Grid provides the industry’s measurement guidance for applying PUE to different kinds of data-center facilities (see Sources, [3]).

Lower overhead can improve the business in several ways. It can reduce facility electricity use for a given IT load. Where total site power is constrained, it may also leave more of the electrical envelope available for revenue-generating computing equipment.

But PUE is not a complete profitability measure.

It does not show whether cabinets are occupied, whether lease prices cover construction costs, whether financing is expensive, whether servers perform useful work, or whether the facility is located where customers need it.

A highly efficient empty building can still be a poor investment.

Use the AI Data Center Power & Cost Calculator to see how IT load, overhead, utilization, operating time, PUE, and electricity prices affect facility energy requirements and cost.

A simple revenue lens

The base economics of contracted power can be expressed without assuming a market price:

Annual base revenue = contracted kilowatts × monthly price per kilowatt × 12

That result is not profit.

The operator must subtract electricity and other operating expenses, maintenance, property costs, employees, taxes, insurance, financing costs, and the continuing investment required to keep the facility competitive.

Interconnection and support services may add revenue. Installation work may create non-recurring revenue. Contract escalators may increase payments over time. Power expenses may be included in the price, separately metered, or passed through under the customer agreement.

The structure of the contract matters as much as the headline number of megawatts.

A long lease can protect revenue visibility, but a poorly priced long lease may lock the operator into weak returns while electricity, equipment, and financing costs rise.

Why location changes the economics

The best data-center site is not necessarily the cheapest land.

An operator must balance several resources:

• Sufficient and dependable power
• Grid-connection timing
• Fiber routes and network carriers
• Proximity to customers and cloud regions
• Latency requirements
• Construction and operating labor
• Cooling conditions and water availability
• Permits, taxes, and local regulation
• Exposure to floods, heat, storms, and other hazards
• Space for future expansion

A location with abundant land but no near-term power may sit idle. A power-rich location without the required network connectivity may not attract the intended customers. A central interconnection hub may support high-value connectivity services even when its property costs are greater.

This is why the physical infrastructure mapped in The Hidden Economy Behind AI Data Centers and the energy strategies examined in Why Big Tech Is Becoming an Energy Company for AI ultimately meet inside the operator’s business model.

The financial risks behind the recurring revenue

Long contracts can make data-center revenue appear bond-like, but the assets still carry substantial risk.

Construction costs can exceed the original budget. Grid connections can arrive late. A customer can delay deployment, default, or choose not to renew. New hardware can require a power density or cooling design the building was not created to support.

Large customers can also create concentration risk.

Digital Realty reported that its three largest customers represented approximately 26% of portfolio annualized recurring revenue at the end of 2025. Its largest 20 represented approximately 51% (see Sources, [2]).

That does not make concentration inherently bad. Large customers can support long leases and major developments. It means the operator’s results may become more dependent on a smaller number of contract negotiations and customer investment plans.

Financing is another part of the model. A developer may spend money years before rent begins. Higher interest rates or construction delays increase the cost carried during that period. If demand weakens before completion, the operator can be left with expensive capacity that takes longer to lease.

Joint ventures and private-capital partnerships can distribute the required investment, but they also divide ownership, cash flow, and control.

The important question is not simply whether AI demand is growing. It is whether the operator can build the correct capacity, secure power, lease it at an adequate price, and finance it without destroying the eventual return.

What the data-center operator is really selling

A modern data center is not merely a building full of servers.

The operator is selling the right to place computing equipment inside a controlled environment with reserved power, cooling, connectivity, security, and reliability. The customer pays repeatedly because recreating that complete system independently can be expensive, slow, and operationally difficult.

Retail colocation turns cabinets and cages into recurring contracts. Hyperscale leasing turns megawatts and entire data halls into long-duration rent. Interconnection turns proximity to other customers and networks into an additional service. Efficient operations protect the margin between revenue and the cost of maintaining the environment.

Power determines how much computing the facility can support. Utilization determines how much constructed capacity is currently earning. Contract pricing and financing determine whether the revenue produces an acceptable return.

That is the hidden business model.

Data centers make money by transforming scarce physical infrastructure—especially power and connectivity—into dependable, long-term digital capacity.