What Is a Colocation Data Center? What the Facility Provides and What the GPU Owner Runs
A colocation data center provides rented space, power, cooling, physical security and network connectivity for organizations that own and operate their own servers. For GPU owners, colocation offers control over their hardware without the responsibility of building and maintaining an entire data center.
However, not every colocation facility can support high-density AI infrastructure. A GB200 NVL72 rack requires roughly 120 kW of power, far exceeding the capacity of many traditional facilities. GPU owners must confirm a facility's power and cooling capabilities before evaluating pricing or contract terms.
What Does Colocation Mean for a GPU Owner?
Colocation sits between owning a private data center and renting cloud capacity. Unlike cloud services, where the provider generally owns and manages the underlying hardware, colocation allows organizations to retain ownership and operational control of their GPUs and servers without managing the physical data center, as Equinix explains.
What Does the Facility Provide, and What Does the GPU Owner Run?
The division of responsibilities extends beyond hardware ownership. The facility manages building infrastructure, power delivery, cooling, physical security and telecommunications access, as HPE explains. The tenant manages its computing hardware, software and workloads.
| Responsibility | Primary owner |
|---|---|
| Building and floor space | Facility |
| Power delivery and distribution | Facility |
| Cooling infrastructure | Facility |
| Physical security | Facility |
| Telecommunications/WAN access | Facility |
| Servers, GPUs and storage | Tenant |
| Operating systems, software and workloads | Tenant |
| Router and switch configuration | Tenant |
| Reboots and power cycling | Tenant, with optional facility assistance |
| Hardware inventory and component replacement | Tenant, with optional facility assistance |
For a GPU fleet, this division determines who handles infrastructure failures. The facility maintains power delivery and cooling, while the tenant is responsible for GPU failures, switch configuration and server reboots. Some colocation providers offer remote-hands services, allowing on-site technicians to perform hardware maintenance or power cycling on the tenant's behalf.
Should a GPU Deployment Use Retail or Wholesale Colocation?
Retail and wholesale colocation differ in the amount of space and power contracted, pricing structure and available services, as described by Fortinet and RackSolutions.
| Criterion | Retail colocation | Wholesale colocation |
|---|---|---|
| Footprint | Individual racks, cabinets or cages | Large dedicated spaces or data halls |
| Per-unit pricing | Generally higher | Often lower at sufficient scale |
| Support | Standardized services with optional assistance | Services and responsibilities negotiated by contract |
| Deployment scale | Smaller deployments | Large-scale infrastructure deployments |
Retail colocation generally suits organizations deploying a limited number of GPU racks, while wholesale colocation may be more appropriate for large AI clusters requiring dedicated space and substantial power capacity.
However, rack count alone does not determine suitability. A small number of high-density GPU racks may exceed the power and cooling capabilities of a retail facility. GPU owners should therefore evaluate both options based on available power, cooling infrastructure and deployment requirements.
Can the Facility Power and Cool the GPU Racks?
Rack power density is a critical requirement for deploying GB200-class AI hardware. A GB200 NVL72 rack requires approximately 120 kW, with SemiAnalysis estimating 123.6 kW including conversion losses. By comparison, a typical general-purpose CPU rack supports up to approximately 12 kW.
The GB200 NVL36x2 configuration distributes approximately 132 kW across two interconnected racks, or roughly 66 kW per rack. This distinction matters because a facility capable of supporting two 66 kW racks may not support a single 120 kW rack.
Facilities designed for higher densities, often 20 kW per rack or more, with advanced cooling are better positioned for AI infrastructure, as Netrality explains. However, supporting 20 kW per rack does not establish compatibility with GB200-class hardware.
DataBank classifies rack power density into four categories:
| Density category (DataBank) | Power per rack | Typical cooling approach |
|---|---|---|
| Traditional | 5–8 kW | Air cooling |
| Standard AI | 15–25 kW | Enhanced air cooling |
| High-density AI | 30–50 kW | Advanced air or liquid cooling |
| Extreme density | 60–100+ kW | Liquid cooling |
A GB200 NVL72 falls into the extreme-density category, requiring high-capacity electrical distribution and compatible liquid cooling. A facility that cannot deliver the required power and cooling per rack is unsuitable for that GPU configuration, regardless of pricing or contract terms.
Before requesting pricing, GPU owners should confirm:
- Rack power: Calculate the planned draw for the specific GPU configuration, including conversion losses.
- Electrical capacity: Verify the facility can deliver the required kW per rack with appropriate power distribution and redundancy.
- Cooling: Confirm compatibility with the GPU system's liquid-cooling requirements.
- Physical requirements: Check floor loading, rack dimensions, maintenance access and expansion capacity.
Facilities that cannot meet these requirements should be excluded from consideration.
What Comes After Confirming Facility Compatibility?
Once a facility meets the deployment's power, cooling and physical requirements, GPU owners can compare colocation costs with renting cloud GPU capacity. The comparison should include hardware acquisition, facility fees, power, networking and maintenance.
Organizations unable to find suitable colocation capacity may consider purpose-built AI infrastructure. Zettabyte's TITAN service line supports AI data center design, integration and operational readiness.