What Is a Colocation Data Center? Costs and How It Works

What Is a Colocation Data Center?

A colocation data center, abbreviated as "colo," is a third-party facility where multiple businesses house their own servers, storage systems, and networking equipment in a shared building. The colocation provider builds and operates the facility: the physical structure, power systems, cooling infrastructure, fire suppression, and physical security. The tenant owns and manages all hardware inside.

Colocation sits between two more familiar options. On-premises computing means owning and running your own facility with all the capital expense, staffing, and maintenance that entails. Public cloud (AWS, Azure, Google Cloud) means renting virtual compute from machines you never physically touch. Colocation is the middle ground: you retain physical ownership of your hardware and complete administrative control over your servers, while offloading the complexity and cost of operating a data center building to a specialist provider.

The model achieves its value through economies of scale. A colocation facility serving 300 tenants spreads the costs of power redundancy systems, cooling infrastructure, security staffing, and network connectivity across all of them. Each tenant pays a fraction of what equivalent redundancy would cost in a private data center.

Colocation comes in three forms, based on deployment size:

Retail colocation uses per-rack or per-cabinet pricing, typically on monthly or annual terms. Wholesale colocation uses per-kilowatt pricing on 3- to 10-year contracts. For a broader view of the data center landscape, including the hyperscale facilities that run AI model training, see our overview of AI data centers.

What a Colocation Contract Covers

A standard colocation contract covers five categories of service. Understanding what is and is not included avoids the most common pricing surprises when comparing provider quotes.

• Physical space: The customer receives dedicated rack space (measured in U units, where 1U equals 1.75 inches of vertical rack height), a full 42U cabinet, a private cage (a locked section of the floor with multiple racks), or a private suite (an enclosed room for large deployments). The space allocation determines which provider personnel can physically access the customer's equipment.

• Power: Dedicated circuits measured in kilowatts or amps deliver power to the customer's racks. Contracts specify the committed power draw and any burst allowances. Redundant power feeds are standard at Tier 3 and Tier 4 facilities. An N+1 configuration has one backup unit for every active unit. A 2N configuration has a complete duplicate power chain, end to end.

• Cooling: The facility maintains ambient temperature and humidity in server rooms. High-density AI deployments increasingly require in-row liquid cooling, direct-to-chip water loops, or rear-door heat exchangers, which most providers offer as premium add-ons.

• Physical security: Biometric access controls, CCTV monitoring, mantraps at entry points, and 24/7 on-site staffing protect the facility. Tier 3 and Tier 4 sites maintain single-tenant cage access logs that satisfy SOC 2, HIPAA, and PCI DSS audit requirements.

• Network connectivity: Access to multiple telecommunications carriers without a separate contract with each one. Carrier-neutral facilities allow tenants to connect directly to any network provider on-site via a cross-connect, typically at $100 to $300 per month per connection.

What colocation does not include

The customer is responsible for all servers, switches, storage devices, operating systems, and software licensing. Hardware maintenance, including physical repairs and drive replacements, belongs to the customer. If a server requires a physical fix, the customer's team must travel to the facility or purchase remote hands services from the provider, typically at $100 to $250 per hour for on-site labor.

This ownership structure is what separates colocation from managed hosting or cloud. The colocation provider has no visibility into the workloads running on customer hardware unless the customer explicitly grants access.

Major Colocation Providers: Equinix, Digital Realty, and Others

Equinix and Digital Realty are the two largest colocation providers globally by revenue and footprint. Below them, a tier of large regional and specialized providers serves enterprise and government customers with specific geographic or compliance requirements.

Equinix's competitive advantage is its interconnection model. Its IBX (International Business Exchange) facilities act as neutral exchange points where internet service providers, cloud platforms, financial networks, and enterprises exchange traffic directly, bypassing the public internet. Equinix reported approximately $8.7 billion in revenue for full-year 2024 and serves more than 10,000 customers globally (Equinix FY2024 earnings).

"The value of a colocation facility is not just the space and power inside it. It is the fabric of 400 networks and 3,000 cloud and IT services you can reach from inside our IBX." (Equinix corporate overview, 2025)

Digital Realty focuses on large-scale, long-duration deployments for enterprise and cloud provider customers. Its PlatformDIGITAL strategy targets enterprises building hybrid multi-cloud architectures. Digital Realty reported approximately $5.7 billion in full-year 2024 revenue (Digital Realty FY2024 earnings).

For context on how hyperscalers such as AWS, Microsoft, and Google use colocation facilities alongside their owned infrastructure, see our explanation of what hyperscalers are and how they operate.

Colocation Pricing in 2026: What You Actually Pay

Colocation pricing uses two models depending on deployment scale: retail pricing (per rack or cabinet) and wholesale pricing (per kilowatt of committed power).

Retail colocation covers 1 to 50 racks and suits businesses that want predictable monthly costs without long-term power commitments. Typical monthly rates as of 2026:

• Standard rack (3-5 kW): $500 to $1,500 per month
• High-density rack (10-20 kW): $1,500 to $4,000 per month
• AI/GPU rack (20-40 kW): $3,900 to $8,000 per month
• Single 1U server slot: $79 to $599 per month

Wholesale colocation covers 250 kW and above, priced per kilowatt with multi-year contracts. The US wholesale market averaged $195.94 per kW per month for 250-500 kW deployments in H2 2025, up from $184/kW in H1 2025, a 6.5% year-over-year increase. According to Cushman and Wakefield's Data Center Power and Lease Pricing Outlook, power availability, not floor space, is the binding constraint on new colocation supply in every major US market in 2026.

Prices vary significantly by geography, driven by local power grid capacity and proximity to internet exchange points:

The Number Most Guides Don't Show

Most comparisons of colocation versus build-to-own stop at headline construction costs. Here is the complete break-even calculation.

Building a 5 MW private data center in the US in 2025:

• Construction cost: $53.5 million (5 MW x $10.7M per MW, per JLL's 2026 Global Data Center Outlook)
• Land and site preparation: $8 to $15 million
• Annual facility maintenance and operations staffing: $3 to $5 million
• Construction timeline: 18 to 24 months before a single server goes live (JLL, 2026)

Total capital outlay: $60 to $70 million, before a single server is installed.

Colocation for the same 5,000 kW:

• Monthly cost: $979,700 (5,000 kW x $195.94/kW per month, CBRE H2 2025)
• Annual cost: $11.76 million

After subtracting $3 to $5 million in annual maintenance costs that disappear when you own rather than rent, the net annual saving from owning is roughly $7 to $9 million. At that rate, recovering the $60 to $70 million capital outlay takes approximately 7 to 9 years.

Most enterprise hardware cycles run 5 to 7 years. By the time the break-even on a self-built facility arrives, the servers inside it are typically due for a complete replacement. This is the core reason why even large enterprises with stable, predictable workloads often continue choosing colocation over build-to-own.

Colocation vs. Cloud vs. On-Premises: How to Choose

The right infrastructure model depends on workload stability, compliance requirements, and the cost structure that fits your planning horizon.

When colocation is the better choice:

1. Compliance requirements mandate physical control of hardware, as in financial services, healthcare (HIPAA), and certain government contracts. 2. Latency-sensitive workloads need guaranteed sub-5ms response times that shared cloud infrastructure cannot reliably deliver. 3. High-bandwidth workloads where cloud egress fees (typically $0.08 to $0.15 per GB transferred out) exceed the cost of colocation bandwidth at scale. 4. Existing hardware investments are not fully depreciated, and moving to cloud would mean paying twice for the same compute capacity.

When cloud is the better choice:

1. Workloads are variable or unpredictable, such as development environments, batch processing, or seasonal traffic spikes. 2. The organization lacks engineering resources to manage physical infrastructure and hardware failures. 3. Global deployments across many regions are needed within weeks rather than months. 4. The team is pre-revenue and cannot justify the capital commitment of hardware ownership.

How AI Is Changing Colocation Requirements

AI workloads are forcing a fundamental redesign of what colocation providers need to deliver. The shift started in 2023 and accelerated through 2025.

Standard enterprise IT racks drew 3 to 5 kW of power. A single NVIDIA H100 GPU server (with 8x H100 SXM5 cards) draws approximately 10 to 11 kW under full load. A GPU cluster of eight such servers, configured in a dense full rack, needs 80 to 90 kW from the facility. Most colocation facilities built before 2020 were designed for 3 to 8 kW per rack. Serving AI compute in those facilities requires expensive power infrastructure retrofits and cooling system upgrades.

As a result, the colocation market has split into two tracks:

1. Standard colocation (3-10 kW per rack): Handles traditional enterprise compute, networking, and storage. This segment faces increasing price pressure as supply has grown relative to demand.

2. High-density AI colocation (20-100+ kW per rack): Purpose-built or retrofitted for GPU clusters. Requires liquid cooling systems (direct-to-chip or rear-door heat exchangers) and dedicated high-capacity power distribution. Monthly rack costs range from $4,000 to $8,000+, plus separate power circuit charges.

According to the JLL 2026 Global Data Center Outlook, AI is the dominant driver of new US data center construction. Construction starts reached $26.9 billion through July 2025, nearly triple 2024 levels. The supply response is concentrated in high-density AI-capable facilities.

Hyperscalers such as AWS, Microsoft Azure, and Google handle the majority of their AI training on owned infrastructure, but use colocation for network exchange points and edge deployments. Mid-market enterprises, AI inference startups, and companies building private AI models are the primary buyers of high-density AI colocation space in 2025 and 2026.

Power, not square footage, limits AI colocation growth in every major US market. In Northern Virginia, utility interconnection queues extend beyond three years. In Silicon Valley, available grid capacity has been effectively exhausted for new large-scale deployments. This constraint is pushing some AI companies toward purpose-built GPU cloud rental for inference workloads that do not require dedicated hardware ownership.

Three Things People Get Wrong About Colocation

Misconception 1: You lose control of your hardware in colocation

The opposite is true. Customers retain complete ownership and administrative control of all hardware in a colocation facility. The provider has no access to customer servers unless the customer explicitly grants it through a remote hands service agreement. Physical access to a customer's cage or cabinet requires customer-provided credentials. The provider logs power consumption data but cannot access, view, or touch customer hardware without authorization.

Misconception 2: Colocation is only for large enterprises

Retail colocation starts with a single 1U server slot from $79 per month and scales from there. Regional providers such as DataBank, Zayo, and dozens of smaller operators serve startups and mid-market businesses that want predictable infrastructure costs without cloud pricing variability. A 5-person engineering team can colocate a single rack for under $1,000 per month and gain enterprise-grade power redundancy, physical security, and network diversity that no private office building can match.

Misconception 3: Colocation SLAs guarantee 100% uptime

Standard Tier 3 colocation SLAs guarantee 99.982% uptime, which allows 1.6 hours of unplanned downtime per year. Tier 4 facilities guarantee 99.995% uptime, roughly 26 minutes per year. No commercially available colocation product guarantees 100% uptime.

Tier classifications are defined by the Uptime Institute. Tier 3 uses N+1 redundancy: one backup for every active power and cooling component. Tier 4 uses 2N+1: a complete duplicate of every system, plus one additional spare. The distinction matters for AI workloads. A GPU training run lasting 72 hours can be invalidated by a 90-minute power event. Companies running long AI training jobs in colocation typically use Tier 4 facilities, or implement model checkpointing to cloud storage to resume from the last saved state after any unexpected interruption.

Frequently Asked Questions

Related Articles