Enterprise Data Center vs Colocation vs Cloud: Total Cost and Control

Building your own data center costs roughly $11.3 million per MW of critical IT load in 2026 and only pays off past 10 MW of scale and a decade-plus horizon. Colocation removes that capital outlay for $130-215+ per kW per month in primary US markets while keeping the enterprise in control of its own hardware. Public cloud removes hardware ownership entirely, at the highest unit cost but zero capex and fastest deployment. On-premises capacity has fallen from 56% of worldwide IT infrastructure capacity in 2018 to 32% in 2026 as most enterprises decide the control an owned facility provides is no longer worth its cost.

Key takeaways

• On-prem share has nearly halved in eight years. 56% of worldwide IT capacity in 2018 to 32% in 2026, per Synergy Research Group, while colocation capacity keeps growing roughly 19% a year.
• Building a facility costs ~$11.3M/MW globally in 2026 (Turner & Townsend), a number that only amortizes favorably above roughly 10 MW and a 10+ year hold — well beyond what most single enterprises need.
• Colocation removes the capex line entirely. Primary US wholesale rates run $130-215+/kW/month; retail racks run $500-1,500/month — see our colocation pricing guide.
• Cloud has the lowest capex and the highest unit cost at scale. It wins on deployment speed (minutes, not months) and loses on steady-state cost — full comparison in our colocation vs cloud TCO guide.
• Control does not disappear in colocation. The enterprise still owns, configures, and secures its own servers; what it gives up is site-selection sovereignty over the building itself.
• The remaining reasons to own outright are narrowing but real: data sovereignty law, workloads that must sit next to specific physical infrastructure, and legacy systems too costly to relocate.

Benchmark current colocation rates in our price index, browse facility options in the data center catalog, and get a quote through our quote service.

The three models, side by side

Dimension	Enterprise (owned)	Colocation	Public cloud
Upfront capital	~$11.3M/MW facility build	None (hardware only)	None
Who owns the servers	The enterprise	The enterprise	The provider
Who owns the building	The enterprise	The operator	The provider
Typical deployment time	18-30 months (new build)	Weeks to a few months	Minutes to hours
Unit cost at high, steady utilization	Lowest per unit once amortized	Low-to-moderate	Highest
Unit cost at low or bursty utilization	Very high (fixed assets, idle)	Moderate (fixed capacity, idle)	Lowest (pay only for use)
Operational burden	Full — power, cooling, physical security, facilities staff	Shared — enterprise runs servers, operator runs the building	Minimal — provider manages nearly everything
Site-selection control	Full	Limited to operator’s footprint	None
Typical minimum viable scale	~10 MW to justify build economics	1 rack and up	Single instance and up

What “control” actually buys

Control is the variable every pure cost comparison flattens out, and it is the real reason on-premises has not gone to zero. Three things scale with how much of the stack an enterprise owns:

1. Data sovereignty and legal jurisdiction. Sovereignty is not the same as residency — it is about which government can legally compel access to the data, not merely where the servers sit. A US-incorporated cloud or colocation provider remains subject to the US CLOUD Act regardless of where its facility is located, which is why regulated sectors (defense, healthcare, financial trading, government) frequently keep specific systems on infrastructure they own outright, per Duality Tech’s 2026 sovereignty analysis.
2. Customization depth. An owned facility can be engineered around one workload — custom power density, custom cooling loop, custom physical layout — without negotiating around another tenant’s requirements. Colocation supports high customization at the rack and cage level but not at the building-systems level.
3. Regulatory and audit posture. The EU AI Act’s high-risk enforcement (effective August 2026) and sector rules in finance and healthcare increasingly require organizations to demonstrate exactly who can physically and administratively access specific systems — easier to document end-to-end on infrastructure the enterprise fully controls.

None of this is free. Every increment of control an enterprise buys back from a shared facility or a cloud provider shows up as capex, operating headcount, or both.

Cost comparison: what each model actually charges

Cost driver	Enterprise build	Colocation	Cloud (on-demand)
Facility capex	~$11.3M/MW (global 2026 benchmark); $15-20M+/MW for AI-density builds	None	None
Recurring facility fee	Internal opex (power, cooling, staff, maintenance)	$130-215+/kW/month (US primary markets); $500-1,500/month retail rack	Instance-hour rate, no facility fee
Hardware capex	Enterprise-owned, same either way if self-managed	Enterprise-owned	Provider-owned (bundled into hourly rate)
Refresh cycle	Enterprise’s own budget and timeline, 4-6 years	Enterprise’s own budget and timeline, 4-6 years	Continuous, provider-managed
Deployment lead time	18-30 months for a new facility	Weeks once a contract is signed	Minutes to hours
Where the premium sits	Capital committed regardless of utilization	Small facility markup over enterprise’s raw operating cost, no capex	Provider’s full margin, priced into every hour

Facility build costs from our data center construction cost guide; colocation and cloud figures from CBRE H2 2025 primary-market data and current provider rate cards.

When ownership still wins

The list of good reasons to own a facility outright has gotten shorter, but it has not gone to zero:

• Scale that clears the capex bar. At 10+ MW and a 10+ year hold, amortized owned infrastructure can undercut both colocation and cloud on a per-unit basis — this is why hyperscalers still build.
• Hard data-sovereignty requirements. Government, defense, and some financial-services workloads must sit on infrastructure the organization legally controls end to end, not merely infrastructure located in the right country.
• Physical co-location with non-IT infrastructure. Industrial control systems, trading floors, and research facilities where compute must sit meters from the equipment it serves.
• Legacy systems that resist migration. Mainframes and specialized hardware where the cost of relocation or re-platforming exceeds years of colocation fees.

When colocation wins — the default for most enterprises in 2026

Colocation has absorbed the bulk of the capacity enterprises used to run themselves. The mechanism is straightforward: build costs and power-grid queues have made new enterprise-owned facilities harder to justify at any scale below hyperscale, while colocation removes the capex line and the multi-year construction timeline without giving up server-level control. Enterprises that migrated workloads to cloud and colocation through the late 2010s and early 2020s are now reaching the point where their remaining on-premises footprint is small enough to retire entirely, accelerating the on-prem decline further. See our due diligence checklist before signing.

Colocation is the right default when an enterprise needs capacity inside 6-12 months, lacks in-house facility design and construction expertise, can accept operating inside a shared building’s physical constraints, and does not have a hard sovereignty requirement that rules out shared infrastructure.

When cloud wins

Cloud remains the right choice for bursty, seasonal, or experimental workloads, and for any deployment shorter than the 6-18 month window it typically takes owned-and-colocated hardware to break even against equivalent rented instances. The full mechanics — egress fees, reserved-instance discounting, GPU buy-vs-rent math — are covered in our colocation vs cloud TCO guide; the short version is that cloud wins on flexibility and loses on cost for anything steady-state and long-lived.

Worked example: 1 MW of steady enterprise compute, 5-year view

Path	Year 1 cost	5-year total	Notes
Build and own (1 MW facility)	~$11.3M capex + ~$1.5-2M opex	~$19-20M	Only rational if this is one MW of a much larger, longer-lived campus; a standalone 1 MW build is uneconomic
Colocate (1,000 kW wholesale-lite, $170/kW/month)	~$2.0M	~$10.2M (plus hardware capex, held constant across paths)	No facility capex; enterprise still owns and refreshes its own servers
Public cloud (equivalent steady compute, on-demand)	Varies by workload; typically 2-4x colocation for constant, hardware-specific load	Typically the highest of the three for genuinely steady-state use	Reserved/Savings Plan discounts of 37-72% narrow but rarely close the gap for constant load

A standalone 1 MW build never clears the economics that make ownership rational at hyperscale — it is included here to show why virtually no single-tenant enterprise builds a facility this size from scratch in 2026.

What to do

1. Segment infrastructure by driver, not by habit. Ask which workloads genuinely require sovereignty or physical co-location, which are steady-state and hardware-specific, and which are bursty or experimental — each answer points to a different model.
2. Default to colocation for anything that clears the cloud breakeven window but does not require standalone facility ownership; benchmark current rates against our price index before signing.
3. Reserve on-premises ownership for the narrow cases — regulatory sovereignty, physical co-location with non-IT plant, or scale genuinely above 10 MW with a decade-plus horizon.
4. Model total cost over at least 5 years, including facility capex or colocation fees, hardware refresh cycles, and operating headcount — not the first-year number alone.
5. Get a colocation quote to compare directly against the cost of maintaining or expanding an owned footprint, through our quote service.