Hyperscale vs colocation: power density planning

The density gap

Data center power density is measured per rack and per square foot of white space. The numbers have changed dramatically over the last decade. A traditional enterprise data center in 2015 ran 3–5kW per rack averaged across the floor. A colocation data center built in 2020 typically designed for 6–10kW per rack with provisions for 15–20kW high-density zones. A hyperscale AI cluster built today operates at 50–100kW+ per rack with liquid cooling as the baseline assumption.

These aren’t small differences. A 50kW rack draws 10x the current of a 5kW rack at the same voltage. Branch circuit ampacity, busway sizing, rack PDU ratings, cooling capacity, and floor loading all change at that scale. Electrical infrastructure designed for one density assumption can’t serve a different one without substantial rework.

What hyperscale operators actually build

Hyperscale operators (AWS, Microsoft Azure, Google Cloud, Meta) design their own data center architectures. Each operator’s internal standards differ in details, but several broad patterns hold across the industry:

• Per-rack power. 30–100kW+ per rack for current AI training workloads. Liquid cooling (rear-door heat exchangers, direct-to-chip, or full immersion) as a baseline rather than an exception. Air-cooled racks at lower densities only for specific applications.
• Distribution. Heavy busway distribution overhead, typically 800A–1,600A per zone, with substantial overcapacity provisioned at construction for density growth.
• UPS topology. 2N or 2(N+1) UPS at building scale, often with operator-specific custom configurations. Battery rooms sized for the full critical load with cycle-life provisions for high-frequency utility events.
• Generator plants. 30MW–200MW+ aggregate plants per building, fueled by diesel or natural gas. Paralleling switchgear sized for operator-specified concurrent maintenance scenarios.
• Operator-specific topology. Each operator has internal design standards that exceed generic Tier IV on some axes and trade off on others. Electrical scope follows operator design, not generic specifications.

What colocation operators design for

Colocation operators (Equinix, Digital Realty, QTS, Stack, CoreSite, DataBank) build facilities that will be leased to dozens of tenants over their operating life. The design has to support an unknown future tenant mix:

• Base building density. Typically designed for 6–15kW average rack density across the white space, with provisions for higher-density zones (often 20–30kW) in specifically designated areas.
• Distribution flex. Overhead busway sized for substantial future load growth, with plug-in capacity that allows tenant-level density variation without base building rework. Plug-and-play tenant fit-out scope.
• Tier III concurrent maintainable. Generally Tier III as the baseline operator standard. Multiple distribution paths with one active at a time, allowing component-level maintenance without IT load disruption.
• Tenant scope demarcation. Clear electrical scope demarcation between operator and tenant work. Operator scope ends at the tenant’s PDU input or RPP main breaker; tenant scope picks up from there.

Why this matters during design

The current confusion in the market is around AI workloads moving into existing colocation facilities. A colocation building designed in 2020 for 8kW average rack density doesn’t have base building electrical infrastructure to support a 50kW AI rack deployment in any meaningful volume. Tenants asking for AI-capable colocation space in older buildings frequently find:

• Available busway capacity exhausted by the first few high-density racks
• Tenant PDU ratings inadequate for the requested branch circuit count
• Cooling capacity in the specific tenant zone insufficient regardless of electrical capacity
• Operator-side service entrance margin that’s already committed to other tenants

For tenants planning AI workloads, the relevant evaluation isn’t "does this colocation operator have Tier III certification" — it’s "does this specific building have the base infrastructure to support 30–50kW racks in the zone we’re looking at." That requires base building review, not tier certification review.

What it means for owners and developers

If you’re building a new colocation data center in 2026, you’re probably designing for substantially higher density than you would have in 2022. The base building infrastructure decisions made during the next 18 months will determine which tenant types you can serve over the next 10–15 years. Under-designed buildings will struggle with tenant retention as workload patterns shift; over-designed buildings will carry capital cost for capacity that may take years to lease.

The electrical scope decisions to weigh include base building service entrance sizing, overhead busway capacity (and physical headroom for future upsizing), tenant demarcation provisioning, and cooling infrastructure capability. None of these are decisions that should be made during construction documents — they belong in early design development with the operator’s long-term tenancy plan as the input.

What it means for contractors

Electrical contractors working in this space need to be fluent in both hyperscale and colocation conventions, because the same operator might run both. They need to coordinate with mechanical contractors on cooling capacity at densities that didn’t exist five years ago. They need to be familiar with the rapidly-evolving rack-level distribution products (intelligent rack PDUs, busway plug-in systems, liquid cooling power infrastructure).

Contractors who built data centers in 2015–2020 patterns and haven’t updated their playbook will struggle on current projects. The pace of density growth has outrun a lot of institutional knowledge that used to be foundational.