Data Center Business Models: Colocation, Wholesale, Hyperscale and Edge Explained

Data center business models determine who carries the construction risk, who pays for power and who captures the margin in the digital infrastructure economy. This guide, part of the Kurums Data Center Economy hub, explains how colocation, wholesale, hyperscale, edge and GPU cloud operators actually earn their revenue, what their margins look like in 2026 results, and why the market is tilting toward giant leased campuses. For the cost side of the equation, see our breakdown of data center construction cost per MW, and for the companies themselves, our ranking of the largest colocation providers.

What are the main data center business models?

There are five main models: retail colocation, wholesale or build-to-suit leasing, hyperscale self-build, edge data centers and GPU-as-a-service. They differ in customer size, contract length, who supplies the IT equipment and how much capital the operator puts at risk per megawatt.

A useful way to picture the industry is as a spectrum running from many small customers to a handful of giant ones. At one end, a retail colocation provider leases a few cabinets to a bank, a SaaS company or a telecom carrier, charges for space, power and cross-connects, and earns a premium for its ecosystem. At the other end, a hyperscaler owns the land, the building, the power contracts and every server inside. Between the two, wholesale developers build entire halls or campuses and lease them to one tenant for ten or fifteen years.

The newest layer, the neoclouds, sits on top of the stack. Companies such as CoreWeave, Nebius, Lambda and Crusoe lease or build data center capacity, fill it with Nvidia GPUs and rent the computing power to AI labs and enterprises by the hour or under multi-year contracts. Their customers buy compute, not real estate. We profile them in our guide to {L(‘neoclouds-coreweave-nebius-lambda-crusoe’, ‘CoreWeave, Nebius, Lambda and Crusoe’)}.

Source: Company Q2–Q3 2026 results releases (Equinix, Digital Realty, CoreWeave); JLL 2026 Global Data Center Outlook.

How does retail colocation make money?

Retail colocation providers rent cabinets, cages and private suites in shared buildings, charging monthly for space and power, plus recurring fees for cross-connects that link customers to each other. Equinix, the largest retail operator, earned about 95% of its Q3 2026 revenue from recurring sources.

Retail colocation is a recurring-revenue business. Customers sign contracts of one to five years and pay a monthly charge per cabinet or per kilowatt of committed power, often with metered energy passed through on top. The real moat is the ecosystem: once networks, clouds, payment processors and exchanges are in the same building, every new customer wants to join them. Equinix says it operates more than 260 data centers in 33 countries and over 500,000 interconnections, the most in the industry.

Equinix’s third-quarter 2026 results, released on 28 October, show the model at scale: revenue of $2.571 billion, up 11% year on year; adjusted EBITDA of $1.321 billion, a 51% margin; and record gross bookings of $533 million in annualised value, up 35%. The company raised its full-year 2026 revenue guidance to $10.325–10.365 billion and its adjusted EBITDA guidance to $5.29–5.33 billion. Recurring capital spending, the maintenance capex needed to keep existing sites running, is guided at only $290–310 million, about 3% of revenue.

Why interconnection is the most profitable product

A cross-connect is a physical cable between two customers inside the same facility. It costs little to install, generates a monthly fee for years, and makes customers reluctant to leave. Equinix earned $465 million from interconnection in Q3 2026, up about 10% year on year, and added 6,900 net interconnections. Digital Realty, which is better known for wholesale, booked $22.2 million of new interconnection revenue in the same quarter. Interconnection revenue is the reason retail colocation trades at premium valuations.

How do wholesale and build-to-suit leases work?

Wholesale developers build large halls or entire campuses and lease them to a single tenant, usually a hyperscaler, under multi-year contracts measured in megawatts. In build-to-suit deals, the facility is designed to the tenant’s specification and pre-leased before or during construction, which lowers the developer’s risk.

Wholesale is a yield business. The developer finances land, power and construction, signs a lease of typically ten to fifteen years or more, and earns a stabilised return on cost. Rent is usually quoted per kilowatt per month, and power is passed through to the tenant. Because a single hyperscaler lease can fill hundreds of megawatts, wholesale developers grow in large steps, and their order books matter more than quarterly revenue.

Digital Realty illustrates the scale. In Q3 2026 it signed $626.6 million of annualised bookings at its share, of which $474.2 million, or 76%, came from deals larger than 1 MW, and $152 million from smaller colocation and interconnection deals. Its signed-not-yet-commenced backlog reached a record $2.5 billion at 100% share. The company guided 2026 revenue to $7.05–7.125 billion and adjusted EBITDA to $3.8–3.85 billion. In its second-quarter materials it reported about 1.4 GW under construction at a total cost of roughly $20 billion, about 63% pre-leased, at an expected stabilised yield of around 11.5%.

The private wholesale developers, including Vantage, QTS, Aligned, STACK and CyrusOne, follow the same model, usually backed by private equity or infrastructure funds. We cover their owners in our guide to private equity data center platforms.

Why do hyperscalers build some data centers and lease others?

Hyperscalers self-build where they have land, power and time, because ownership is cheaper over the long run, and lease where they need speed or capacity in markets where they lack sites. JLL expects about 41 GW of owner-occupied hyperscale capacity and about 62 GW leased between 2026 and 2030.

Owning gives a hyperscaler control over design, efficiency and long-term cost. Google, Microsoft, Amazon and Meta all operate custom campuses designed around their own servers and cooling systems. Leasing gives speed and flexibility: the developer takes the construction and financing risk, and the hyperscaler keeps capital free for servers and GPUs, which now absorb the majority of AI capex. JLL’s forecast that leased hyperscale capacity will exceed owner-occupied capacity over 2026–2030 shows how much the giants are relying on third-party developers.

The split has an accounting dimension too. Microsoft said in 2026 that more of its future data center leases would be classified as operating rather than finance leases, which lowers reported capex while the spending continues, a point we explain in our analysis of the AI capex supercycle. Synergy Research Group estimates that hyperscale operators accounted for 48% of all data center capacity worldwide at the end of 2025, against 20% for non-hyperscale colocation and 32% for on-premise enterprise facilities, and forecasts the hyperscale share reaching 67% by 2031.

What is edge data center colocation?

Edge data centers are smaller facilities placed close to end users, mobile networks or industrial sites to cut latency and backhaul costs. They earn revenue like retail colocation, from space, power and connectivity, but at much smaller scale per site and with a heavier dependence on telecom partners.

The edge model was hyped in the late 2010s as 5G, connected vehicles and streaming promised a need for thousands of micro data centers. In practice, most demand has concentrated in large metro hubs and giant campuses, because AI training does not need to sit near users and cloud regions already cover major cities. Edge facilities remain important for content delivery networks, telecom operators, gaming and some industrial and retail workloads, but they represent a small share of global capacity and investment.

AI inference may change the picture. JLL expects inference to overtake training as the dominant AI requirement around 2027. Many inference workloads still run in large regional data centers, but latency-sensitive applications could push some capacity closer to users. For operators, edge is a portfolio business: margins depend on filling many small sites, which is harder than filling one large hall. Site economics are covered in our site selection guide.

How do GPU-as-a-service neoclouds make money?

Neoclouds buy or lease Nvidia GPUs, install them in leased or self-built data centers and sell computing time, mostly under multi-year take-or-pay contracts with AI labs and hyperscalers. Revenue grows quickly, but heavy depreciation and interest costs leave thin operating margins.

CoreWeave, the largest listed neocloud, reported second-quarter 2026 revenue of $2.575 billion, up about 112% from $1.212 billion a year earlier, and a revenue backlog of about $104 billion, excluding more than $25 billion of new commitments signed early in Q3. Adjusted EBITDA was $1.51 billion, a 59% margin, but adjusted operating income was only $128 million, a 5% margin, because GPUs are depreciated over a few years. Net interest expense was $640 million and the net loss was $626 million. CoreWeave had 1.5 GW of active power and about 3.7 GW contracted at the end of June.

The model is closer to equipment leasing than to real estate. Revenue is secured by contracts with large customers, and lenders underwrite those contracts: a CoreWeave executive described lenders as effectively underwriting Microsoft’s credit rather than CoreWeave’s balance sheet. Concentration is the trade-off; Microsoft accounted for about 67% of CoreWeave’s 2025 revenue. We discuss how these structures are financed in our guide to data center financing.

How do margins compare across data center business models?

Retail colocation and wholesale leasing earn adjusted EBITDA margins of roughly 50–55%, with modest maintenance capex. GPU clouds can report similar EBITDA margins but carry far heavier depreciation, so operating margins are much thinner. Hyperscalers’ data center returns are embedded in their cloud margins.

Equinix’s 51% adjusted EBITDA margin and Digital Realty’s guidance, which implies roughly 54% on the midpoints of its revenue and EBITDA ranges, reflect long-lived buildings that are depreciated over decades and leases that pass power costs through. CoreWeave’s 59% adjusted EBITDA margin looks higher, but the 5% adjusted operating margin shows the cost of assets that must be replaced every few years. Hyperscalers do not disclose data center margins separately, but Amazon reported a 38.7% operating margin for AWS in Q3 2026, and management warned that the mix shift toward AI and higher component costs, including memory, add pressure.

Pricing is currently on the landlord’s side. JLL reports global occupancy approaching 97%, forecasts rents rising about 5% a year through 2030, and says AI-optimised space commands lease rates up to 60% higher than traditional space. Those conditions favour owners of powered capacity in any model. They also invite new supply, which is why investors watch preleasing rates and power availability as closely as current rents.

Which data center business model will win in the AI era?

No single model wins outright. Hyperscale and leased wholesale capacity will capture most new megawatts, while retail colocation keeps the richest margins through interconnection. GPU clouds offer the fastest growth but depend on a few large customers and continuing AI demand.

The trend lines point in clear directions. Synergy Research Group expects hyperscale operators to own or lease two-thirds of global capacity by 2031, and on-premise enterprise data centers to fall from 32% to 19%. Colocation capacity keeps growing in absolute terms even as its share slips. Equinix raised its 2027–2029 outlook to 10–13% annual revenue growth and 9–12% annual AFFO per share growth, suggesting that enterprise AI and hybrid cloud are reviving retail demand rather than displacing it.

The biggest uncertainty is whether AI demand stays strong enough to fill the capacity under construction. Bulls point to backlogs: CoreWeave’s $104 billion, Digital Realty’s record $2.5 billion and the hundreds of billions in cloud backlogs at Microsoft, Amazon and Oracle. Bears point to concentration, depreciation and circular financing between chipmakers, AI labs and cloud providers. We weigh both sides in Is there an AI data center bubble?.

Frequently Asked Questions