Build, Lease or Colocate? How to Source Data Center Capacity

Every data center user, from a regional bank to a hyperscale cloud provider, chooses some mix of four sourcing models. They differ in who owns the land and building, who carries the power and construction risk, and how long the commitment runs.

The lines blur in practice. A hyperscaler may self-build in one market, sign a build-to-suit in another and take wholesale suites in a third. Synergy Research counts this mixed footprint as “hyperscale capacity” whether the buildings are owned or leased, and estimates that almost 60% of it sits in buildings the operators built and own.⁶

For site selection, the model matters because it decides who is looking for land and what they need from it. A self-builder needs a parcel with a credible power path. A colocation tenant needs a market with available space. A cloud customer needs neither, only a region with the right latency and services.

Enterprise workloads have been moving off premises for years, though more slowly than many forecasts predicted. Uptime Institute’s 2025 survey estimated that about 45% of corporate workloads ran in on-premises data centers, down from 58% five years earlier, and noted that the share did not fall between 2024 and 2025.⁷ Uptime cautions that these figures are approximate trend indicators.⁷ Its 2026 survey found third-party venues at 46% of workloads and enterprise-owned facilities at 44%, with the gap projected to widen to 48% versus 42% by 2028.²

Measured by capacity rather than workloads, the shift is larger. Synergy Research put hyperscale operators at 48% of global data center capacity at the end of 2025, non-hyperscale colocation at 20% and on-premises facilities at 32%, compared with 56% on-premises in 2018.⁶ It forecasts the hyperscale share reaching 67% by 2031, with on-premises falling to 19%.⁶

The mix varies by company, which shows there is no single right answer even at the top of the market. DC Byte found that Google, AWS, Meta and Microsoft each keep less than half of their capacity in colocation, while Oracle, ByteDance, Alibaba, CoreWeave and Tencent rely on it almost entirely.³ Our guide to how hyperscalers choose sites covers what drives their location decisions.

Building and owning gives the user full control over location, design, redundancy, security and expansion. It also hands the user every hard part of the project: land, utility service, entitlement, equipment procurement and construction. For a large campus, that is a multiyear program run by a dedicated team.

The capital is substantial. Cushman & Wakefield’s 2026 cost guide, as summarized by CRE Daily, put all-in greenfield costs in the U.S. and Canada at $8.9–23.3 million per MW, excluding chips and GPUs, after an average 21% rise since late 2024.⁴ Power infrastructure was the largest single category at about 21% of a greenfield build, ahead of core and shell plus sitework at 17%, contingencies at 16% and land at about 7%.⁴ Our guide to construction costs and site drivers breaks these down.

Time is the other cost. JLL reported average grid connection lead times exceeding four years in primary markets,⁸ and pad-mounted transformers alone carried lead times of 68–113 weeks in the same cost guide.⁴ A self-builder therefore needs to control land with a realistic utility path years before the capacity is needed, usually through an option or purchase, and often stages the campus as described in phased vs. full build-out.

Self-build fits best when the load is large, long-lived and strategic: tens or hundreds of megawatts, a horizon of 15 years or more, and requirements (security, density, custom cooling) that a shared building would struggle to meet.

In a build-to-suit, a developer acquires the land, secures power, builds to the tenant’s specification and leases the finished facility to that single tenant. GLP describes the rent as typically set on a yield-on-cost basis: a pre-agreed return over the developer’s total development cost.⁹ The model has grown because hyperscale demand outpaced what the operators could deliver through their own self-build programs.⁹

A powered shell is a lighter version: the developer delivers the building, utility service and often the electrical backbone, and the tenant fits out the data halls and mechanical systems. The developer’s core contribution is the site and the power, which is why powered land has become a distinct asset class.

Both structures move capital off the tenant’s balance sheet but not the long-term obligation. Credit rating agency KBRA reports that most hyperscale data center leases are net leases, and that absolute triple-net structures shift operational control and lifecycle capital costs to the tenant.⁵ KBRA also flags that termination, contraction and assignment rights can reduce cash flow visibility, which is why investors negotiate them closely.⁵ Our guide to how data center deals are structured covers the lease mechanics.

Colocation providers own and operate multi-tenant buildings and lease space, power and cooling. Wholesale colocation leases whole data halls or multi-megawatt suites, often to cloud and large enterprise tenants. Retail colocation leases racks, cages and small suites, usually with interconnection to carriers and cloud on-ramps as part of the value.

The economics have tightened. CBRE reported that average asking rates for a 250–500 kW requirement rose 2.5% between the second half of 2024 and the first half of 2025, while larger deployments of 10 MW or more rose by up to 19%.¹⁰ CBRE attributed the increases to scarce contiguous power blocks, higher build-out costs and competition from cloud and AI tenants, and expected pricing to stay elevated at levels comparable to 2011–2012, above $200 per kW per month.¹⁰ Its quoted rates assume N+1 or Tier III-type configurations.¹⁰

Colocation still wins on speed when space exists, and for loads under a few megawatts it is usually the only practical option short of cloud. The trade-off is less control over location, design and expansion. Our guides to colocation site selection and vacancy, absorption and pricing cover the market side.

Cloud removes the facility decision entirely. The user rents compute, storage and services by usage and leaves land, power and buildings to the provider. For variable, short-lived or experimental workloads, that flexibility is hard to match.

The case weakens for large, steady workloads. A widely cited 2021 Andreessen Horowitz analysis pointed to Dropbox, whose IPO filing showed about $75 million in cumulative savings over two years, mostly from moving workloads off public cloud, and argued that repatriation can bring costs to one-third to one-half of equivalent cloud spend.¹¹ The same essay noted that committed-use discounts narrow the gap.¹¹ Critics countered that Dropbox spent roughly $200 million in capital around its IPO that did not show up in cost of goods sold, so the savings were less clear-cut than presented.¹²

The lesson for a sourcing decision is not that one answer is cheaper. It is that the comparison has to include capital, staff, power and the cost of time on the owned side, and egress, lock-in and discount commitments on the cloud side. The 2021 figures are also dated; prices and hardware have changed since.

Most sourcing decisions come down to four questions. Answer them in order; each narrows the options.

1. 01How big is the load? Below a few megawatts, retail colocation or cloud usually wins. At tens of megawatts, wholesale colocation and build-to-suit come into play. At hundreds of megawatts, self-build and build-to-suit dominate.
2. 02How long will it last? A load expected to run 15 years or more justifies ownership or a long lease. A load that may change within five years argues for colocation or cloud.
3. 03How much control is required? Custom density, liquid cooling, security or regulatory requirements can rule out shared buildings.
4. 04When is it needed? If capacity is needed within two years, leased space that is already under construction may be the only option. If the need is five years out, a self-build or build-to-suit on controlled land is feasible.

Industry views on the long-term mix differ. L.E.K. Consulting has argued that most hyperscaler capacity will be leased from third-party operators over the next 10–15 years,¹³ while Synergy’s data show the largest operators still owning most of their hyperscale capacity today.⁶ A hybrid approach is the common outcome.

The sourcing model decides what kind of site search follows. A self-builder or build-to-suit developer needs land that can actually be powered: transmission access, substation headroom, buildable acres, a zoning path, water and fiber. A colocation tenant compares markets and buildings, not parcels. A cloud customer compares regions.

• If self-building, start the land and utility conversations early. Grid lead times of four years or more in primary markets mean the site decision often precedes the capacity need by years.⁸
• If leasing, confirm the provider’s power is contracted and energized on your schedule, not just planned. With most new supply preleased, delivery dates matter more than asking rents.¹
• If using build-to-suit, review the developer’s site as if you were buying it. The lease shifts capital, not site risk.
• In every case, model the full term: rent escalation, power cost and the cost of moving later.

BlackForge screens land for power, water, fiber, terrain and zoning, which is the starting point for a self-build or build-to-suit decision. Our methodology explains the screen, and you can get a site reviewed when you have a parcel in mind.