The easiest way to think about data center operating models is as a stack. The operator can stop at the building layer, or keep moving upward into power, cooling, network, servers, GPUs, software, and managed compute.
Currently, I have a public model for data center colocation strategies. Over the next few months I'm planning on building a few more versions that fit different strategies.
At the bottom, the customer brings almost everything. At the top, the operator is basically selling an AI/cloud product.
1. Powered Shell / Powered BuildingThis is the most “real estate-like” version.
The operator develops or controls the site, building shell, utility access, fiber access, permits, and sometimes base electrical infrastructure. The customer handles much of the interior fit-out: UPS, generators, cooling configuration, racks, cabling, and IT deployment.
What the operator sells: a data-center-ready building or space.
What the customer brings: most of the technical interior and all IT equipment.
Operational burden: lower than full colocation because the tenant does more.
Typical customer: hyperscaler, large enterprise, cloud provider, AI infrastructure company, or another data center operator.
Digital Realty describes this style as fit-out-ready buildings on power- and fiber-provisioned sites, while H5 describes powered shells as completed exterior facilities with available power and connectivity but raw interior space left for customer fit-out.
Modeling implication: revenue is often lease-based, with lower service revenue but less operating complexity. The big risks are development cost, power delivery timing, tenant credit, and lease-up.
2. Build-to-SuitBuild-to-suit is similar, but more customized. The operator or developer builds a facility around one customer’s exact power density, cooling, redundancy, security, and location requirements.
What the operator sells: a custom facility.
What the customer gets: more control and a design tailored to its workload.
Operational burden: depends on the contract; the operator may only develop/lease the facility, or may also operate it.
Typical customer: hyperscaler, major enterprise, government, cloud provider, or large AI tenant.
A build-to-suit data center is generally designed around a single tenant’s requirements, while powered shell leaves more of the interior infrastructure to the customer.
Modeling implication: this is usually a large, long-cycle development case. It can have strong contracted revenue, but the risk is concentrated in one tenant and one custom design.
3. Wholesale ColocationThis is where the operator provides a finished data center environment, but leases large blocks of capacity to a small number of large customers.
The customer usually takes a full hall, suite, floor, or multi-megawatt block. The operator provides the building, power, cooling, redundancy, physical security, and sometimes connectivity options. The customer brings and operates the servers, storage, network equipment, and applications.
What the operator sells: large-scale space, power, cooling, and uptime.
What the customer brings: IT hardware and workload operations.
Operational burden: higher than powered shell, lower than managed cloud.
Typical customer: hyperscaler, SaaS company, enterprise, AI cloud provider, financial institution.
Modeling implication: revenue is often driven by committed power, square footage, lease duration, escalators, power pass-throughs, and expansion options. It is more infrastructure than software.
4. Retail ColocationRetail colocation is the classic “rent racks/cages/cabinets” model.
The operator runs the facility and sells smaller units of capacity to many customers. Customers place their own servers and equipment in the operator’s facility. The provider handles physical space, power, cooling, security, network access, and often remote hands. Equinix defines colocation as an organization placing its own servers or hardware in rented space within a third-party physical data center.
What the operator sells: cabinets, racks, cages, cross-connects, power, cooling, security, bandwidth options, and support.
What the customer brings: servers and software.
Operational burden: higher than wholesale because there are many smaller customers.
Typical customer: small/medium enterprises, SaaS companies, IT service providers, financial firms, healthcare, backup/disaster recovery users.
Modeling implication: more customers, more churn assumptions, more sales activity, more support labor, and more ancillary revenue. You can model revenue by rack, cabinet, power draw, bandwidth, cross-connects, remote hands, and managed add-ons.
5. Carrier Hotel / Interconnection HubThis is a specialized form of colocation where the real value is not just space and power. It is network density.
The operator attracts carriers, cloud on-ramps, internet exchanges, content networks, SaaS platforms, and enterprise customers who want to connect privately with one another. Equinix emphasizes colocation and interconnection as access points to cloud providers, AI providers, networks, SaaS platforms, and enterprise partners.
What the operator sells: proximity, cross-connects, low-latency access, network ecosystems, cloud connectivity.
What the customer brings: IT gear, network gear, and traffic.
Operational burden: high on network operations, meet-me-room management, security, provisioning, and customer coordination.
Typical customer: telecom carriers, cloud providers, financial exchanges, content delivery networks, enterprises needing hybrid cloud access.
Modeling implication: cross-connects and interconnection services can become a major revenue line. This is less about raw square footage and more about ecosystem density.
6. Managed ColocationManaged colocation moves one layer up.
The customer may still own the hardware, but the operator provides more hands-on support: remote hands, smart hands, monitoring, hardware swaps, troubleshooting, backups, firewall management, operating system support, or other managed IT services.
Remote hands usually means the customer can delegate physical maintenance or IT tasks inside the facility to technicians hired by the provider.
What the operator sells: colocation plus operational support.
What the customer brings: usually the hardware and business applications.
Operational burden: meaningfully higher because the operator now needs technical staff, ticketing, SLAs, monitoring, and customer support processes.
Typical customer: companies that do not want to send their own IT staff to the facility.
Modeling implication: labor becomes more important. Revenue can include recurring support plans, hourly remote hands, monitoring, backup services, firewall services, and managed network services.
7. Managed Hosting / Dedicated Servers / Bare MetalNow the operator owns the servers.
Instead of the customer bringing hardware, the operator buys and installs servers, then rents them to customers. The customer gets dedicated physical machines or configured hosting environments. IBM describes bare metal servers as cloud services where the user rents a physical machine that is not shared with other tenants.
What the operator sells: dedicated physical compute.
What the customer brings: applications, workloads, sometimes operating system configuration.
Operational burden: much higher than colo because the operator owns hardware procurement, maintenance, spares, imaging, provisioning, monitoring, replacement cycles, and support.
Typical customer: SaaS companies, gaming companies, security-sensitive users, database-heavy workloads, high-performance workloads.
Modeling implication: you now need server capex, useful life, refresh cycles, hardware utilization, provisioning time, support labor, bandwidth, storage attach, and gross margin by server type.
8. Private Cloud / Bare Metal CloudThis is a more productized version of bare metal or dedicated infrastructure.
The operator provides dedicated servers, storage, network, virtualization or container orchestration, APIs, portals, and usage-based provisioning. Google Cloud’s Bare Metal Solution, for example, provides purpose-built bare-metal servers connected to Google Cloud through managed low-latency networking.
What the operator sells: dedicated infrastructure with cloud-like deployment and management.
What the customer brings: workloads, applications, data, and configuration.
Operational burden: high. The operator needs cloud orchestration, automation, monitoring, identity/security controls, billing systems, and support.
Typical customer: enterprises that want cloud-like operations but need dedicated hardware, compliance, predictable performance, or hybrid cloud integration.
Modeling implication: the revenue model starts looking like cloud: monthly recurring commitments, usage-based billing, reserved capacity, storage, bandwidth, support tiers, and utilization-driven margins.
9. GPU-Ready ColocationThis is colocation designed for high-density AI hardware, but the customer still owns the GPU servers.
This is a big category right now. The operator provides high-density power, advanced cooling, reinforced floor/rack designs, liquid cooling support, high-capacity network pathways, and operational procedures for GPU clusters. The customer brings GPU servers, networking, storage, and AI software.
What the operator sells: AI-capable facility capacity.
What the customer brings: GPU hardware and AI stack.
Operational burden: higher than conventional colo because GPU racks can require much more power density, different cooling designs, more careful commissioning, and tighter facility controls.
Typical customer: AI labs, GPU cloud companies, enterprises building private AI clusters, research institutions, HPC users.
This is not just marketing. CBRE reported that AI and GPU workloads are creating demand for advanced infrastructure, and operators with AI-optimized facilities, including liquid cooling and high-power-density racks, can capture rent premiums over conventional colocation. Uptime Institute also notes that operators are facing rising costs, power constraints, and the need to modernize for AI-related density requirements.
Modeling implication: the drivers change from “racks rented” to “kW/MW committed,” density premiums, cooling capex, utility constraints, phased energization, and power availability.
10. GPU-as-a-Service / AI Cloud / NeocloudThis is where the operator owns the GPUs and sells access to them.
Instead of leasing space, the company rents GPU instances, clusters, reserved GPU capacity, inference capacity, or training clusters. Examples in the market include AI cloud providers selling on-demand GPUs, private clusters, dedicated GPU capacity, Kubernetes-native clusters, and high-performance storage/networking around those GPUs. CoreWeave describes itself as purpose-built for GPU performance for demanding AI workloads, while Lambda markets cloud GPUs, on-demand clusters, private cloud, and AI training/inference infrastructure.
What the operator sells: compute capacity, not just facility capacity.
What the customer brings: models, data, training jobs, inference workloads.
Operational burden: very high. The operator needs GPU procurement, cluster architecture, networking, storage, orchestration, job scheduling, security, monitoring, customer portals, support, and utilization management.
Typical customer: AI startups, model labs, enterprises, research teams, rendering users, HPC customers.
Modeling implication: this is a completely different model from colocation. You need GPU capex, depreciation, hardware refresh risk, hourly/monthly utilization, reserved contracts, spot/on-demand pricing, power cost per GPU-hour, customer concentration, networking/storage attach, and obsolescence risk.
11. Full-Stack AI Infrastructure PlatformThis is the farthest up the stack.
The operator does not just rent GPUs. It provides a full AI environment: GPU clusters, high-speed networking, storage, Kubernetes or Slurm, model development tools, inference serving, security, monitoring, support, and sometimes prebuilt AI software. NVIDIA’s DGX Cloud was launched as an AI supercomputing service giving enterprises access to infrastructure and software for training advanced models. NVIDIA also describes DGX SuperPOD architecture as a combination of DGX systems, InfiniBand/Ethernet networking, management nodes, and storage, which shows how much more complex AI infrastructure is than standard rack colocation.
What the operator sells: an AI platform.
What the customer brings: data, models, use cases, engineering teams.
Operational burden: highest. This starts to resemble running a specialized cloud company, not just a data center.
Typical customer: enterprises building AI products, model developers, AI labs, government/research groups, companies needing private AI compute.
Modeling implication: software and service margin can be higher, but so are capex, technical hiring, hardware obsolescence, customer support, uptime requirements, and utilization risk.
The Practical SpectrumHere is the clean way to summarize the options:
| Operating path | Operator owns/controls | Customer owns/controls | Main revenue driver |
|---|---|---|---|
| Powered shell | Site, shell, power/fiber readiness | Interior fit-out, IT stack | Long-term lease |
| Build-to-suit | Custom facility development | Tenant-specific requirements, often IT | Long-term lease / development return |
| Wholesale colo | Facility, power, cooling, security | Servers, network, workloads | MW / power commitment |
| Retail colo | Facility, racks/cages, power, cooling, support | Servers and software | Cabinets, kW, cross-connects |
Interconnection hub | Colo facility plus network ecosystem | Network/customer equipment | Cross-connects, cloud/network access |
| Managed colo | Colo plus support services | Usually hardware and apps | Colo fees plus managed services |
| Bare metal / hosting | Facility plus servers | Apps/workloads | Server rental / monthly compute |
| Private cloud | Servers, network, storage, orchestration | Workloads/data | Reserved/usage-based infrastructure |
| GPU-ready colo | High-density AI-capable facility | GPU hardware and AI stack | High-density kW/MW rent |
| GPU-as-a-Service | GPUs, clusters, storage, network | AI workloads | GPU-hour / reserved clusters |
| Full-stack AI platform | Facility, GPUs, software platform, support | Data/models/use cases | Compute + software/platform revenue |
The Big Strategic Choice
A data center operator has to decide where it wants to sit:
Low-stack operator: sells space, power, cooling, and reliability. This is closer to real estate and infrastructure. It can be easier to finance, easier to understand, and more contract-driven.
Mid-stack operator: adds interconnection, remote hands, managed services, and customer support. This creates more revenue lines but also more staffing and operational complexity.
High-stack operator: owns servers, GPUs, cloud orchestration, storage, software, and customer workloads. This can generate far more revenue per MW, but it brings hardware risk, utilization risk, technical support risk, and fast obsolescence.
The GPU/full-stack path is the most interesting but also the most dangerous to underwrite casually. A conventional colocation project can be modeled around leased power and occupancy. A GPU cloud project has to be modeled around hardware deployment schedules, GPU utilization, customer commitments, power cost per GPU-hour, networking/storage cost, hardware refresh cycles, and whether the operator can actually run a reliable compute platform.
You may find this AI B2C Financial Model of Interest.
If you are looking into a new project and need custom financial modeling, I'm available for hire here.
Article found in General Industry.
