In conventional commercial real estate, you underwrite rent per square foot. In data center real estate, you underwrite megawatts of critical power. Everything else — the building, the land, the location — is infrastructure in service of delivering power reliably to servers.

This single shift in perspective is the most important thing to understand about the sector. A 200,000 SF building might be worth $50M as a distribution warehouse. The same shell, repositioned as a 100 MW hyperscale data center with a utility interconnection and robust cooling, is worth $500M or more. The power delivery capability created that value.

This lesson covers the power terminology you need to speak fluently with operators and brokers, how to evaluate facility efficiency, cooling infrastructure, backup systems, and the utility interconnection dynamic that controls new supply across the entire sector.

Why power is everything

Data centers are purpose-built to do one thing: keep servers running at full capacity, continuously, with no unplanned outages. Every system in the facility — cooling, backup generators, UPS, electrical distribution — exists to protect the power supply to IT equipment.

From an investment standpoint, the critical power rating of a facility is its productive capacity. Colocation tenants lease in terms of kilowatts (kW) per cabinet or megawatts (MW) for large deployments. Hyperscalers sign pre-lease agreements for hundreds of megawatts before a building is even fully constructed. The "product" being sold is reliable power at a given density.

Consider two deals in the same market:

Deal A: Generic industrial building

200,000 SF
Standard 3,000 amp electrical service
No cooling infrastructure
Value: $40–60M (standard industrial pricing)

Deal B: Purpose-built data center, same footprint

200,000 SF
80 MW of critical IT load
N+1 cooling, 2N UPS, 48-hour generator fuel storage
20-year lease with an investment-grade hyperscaler
Value: $400–600M

Same dirt. Same structure, roughly. The operational infrastructure and the contracted power delivery capability created an order-of-magnitude difference in value.

This is why power infrastructure is the central due diligence item in data center real estate — not the building, not the location, not the lease terms.

Power terminology

The industry uses several overlapping units and concepts. Confusing them is a common mistake for investors coming from conventional real estate.

MW, kW, and kVA

| Term | Definition | Practical use | |---|---|---| | Watt (W) | Base unit of electrical power | Too small for facility-level discussion | | Kilowatt (kW) | 1,000 watts | Cabinet-level power density (e.g., 10 kW per rack) | | Megawatt (MW) | 1,000,000 watts (1,000 kW) | Facility-level capacity (e.g., 100 MW campus) | | kVA | Kilovolt-ampere — apparent power | Electrical engineering spec; always larger than kW because of power factor |

For real estate conversations, MW and kW are the primary units. kVA appears in electrical specs (UPS ratings, transformer ratings) and is always somewhat larger than the kW figure due to power factor losses.

Rough conversion: kW = kVA × power factor. A modern data center typically targets a power factor of 0.95–1.0, so kW and kVA are nearly equivalent in practice.

Critical power vs total facility power

This distinction matters enormously for evaluating efficiency.

| Concept | Definition | Example | |---|---|---| | IT Load (Critical Load) | Power consumed by servers, storage, and networking equipment — the actual computing work | 80 MW | | Total Facility Power | IT load plus everything else: cooling, lighting, security, office space | 120 MW | | Overhead Power | Total facility minus IT load | 40 MW |

When a data center is described as "100 MW," the specification should always clarify: is that critical power (IT load) or total facility power? Critical power is the commercially meaningful number — it is the power actually delivered to tenants and generating revenue.

Power density — kW per rack

Power density measures how much power is delivered per server rack (also called a cabinet). It drives infrastructure requirements more than total facility size.

Power density = IT load (kW) / number of racks

| Era / Workload | Typical density | |---|---| | Legacy enterprise servers (2010s) | 3–5 kW per rack | | Modern cloud servers | 10–20 kW per rack | | AI/GPU training clusters | 40–100+ kW per rack | | Liquid-cooled AI pods | 100–200+ kW per rack |

A 10,000-square-foot data center hall designed for 5 kW per rack requires completely different cooling, electrical, and structural infrastructure than the same hall retrofitted for 80 kW per rack AI workloads. Legacy air-cooled facilities simply cannot serve high-density AI deployments — the cooling is inadequate.

PUE — Power Usage Effectiveness

PUE is the standard efficiency metric for data centers and directly affects tenant economics.

PUE = Total Facility Power / IT Equipment Power

A PUE of 1.0 is theoretically perfect — every watt consumed goes directly to computing. Every fraction above 1.0 is overhead: cooling, lighting, electrical losses.

| PUE | Classification | Who achieves it | |---|---|---| | 1.0 | Theoretical perfect | Unachievable | | 1.1–1.2 | Best-in-class | Google, Meta, Microsoft hyperscale campuses; modern liquid-cooled facilities | | 1.3–1.4 | Efficient | Well-designed newer facilities with economizer cooling | | 1.5–1.6 | Industry average | Most existing colocation facilities | | 1.7–2.0+ | Inefficient | Older facilities, poorly designed cooling | | 2.0+ | Poor | Legacy facilities without optimization |

The 2024 industry average PUE reported by the Uptime Institute was approximately 1.58.

Why PUE affects tenant economics

Colocation tenants typically pay for power consumption in one of two ways:

A flat monthly rate per kW of reserved power capacity
Metered power at a rate per kWh consumed

In either case, a lower-PUE facility is more attractive to tenants because their effective energy cost per unit of computing is lower. If a tenant's servers draw 10 MW and the facility PUE is 1.2, the facility consumes 12 MW total. At PUE 1.6, the same tenant drives 16 MW of total facility consumption — and in a metered arrangement, the tenant pays for some or all of that overhead.

For hyperscalers like Google, Amazon, and Microsoft that run millions of servers continuously, a 0.1 improvement in PUE across a 100 MW campus saves roughly $1–2M per year in power costs at typical US commercial electricity rates.

This is why hyperscalers build their own campuses with state-of-the-art cooling rather than co-locating: they can design for 1.1–1.2 PUE and capture the efficiency gains at scale.

How PUE affects facility value

From an investor perspective, a facility with a certified PUE of 1.3 or better commands:

Higher rental rates per kW (tenants pay premium for efficiency)
Longer lease terms (tenants are reluctant to move after optimizing workloads for a facility)
Stronger tenant credit quality (hyperscalers and well-funded operators pursue efficient facilities)
Greater expansion demand (once a hyperscaler is in, they want more capacity in the same facility)

A legacy facility with PUE 1.8+ is functionally obsolescent for most modern cloud and AI workloads and would require significant capital investment to compete.

Cooling systems — the biggest variable cost after power

Cooling is the primary overhead that separates PUE from 1.0. A data center generating 100 MW of IT heat must remove that heat as efficiently as possible to protect equipment and keep PUE low.

Air-based cooling (legacy and current mid-density)

CRAC (Computer Room Air Conditioning) The oldest approach. Standalone air conditioning units positioned on the raised floor blow cold air under the floor, through perforated tiles, and into server intakes. Problems: inefficient, cold air mixes with hot exhaust air before reaching servers, limited to low densities.

CRAH (Computer Room Air Handler) An improvement over CRAC. CRAHs use chilled water from a central chiller plant rather than direct-expansion refrigerant. More efficient, easier to scale. Still limited in density versus liquid cooling.

Economizer (Free) Cooling Uses outside air or evaporative cooling when ambient temperatures are low enough to remove heat without mechanical refrigeration. This is the key technology behind low-PUE hyperscale facilities.

Google's data centers in Oregon, Finland, and other cold-climate locations achieve PUE around 1.1 precisely because they can use outside air for cooling most of the year. The same facility in Phoenix would have PUE closer to 1.4–1.5 because mechanical cooling is required year-round.

Economizer-capable facilities are significantly more valuable — lower operating costs, better PUE, preferred by large tenants.

Liquid cooling (AI-driven demand shift)

AI GPU clusters — particularly Nvidia H100 and H200 pods — generate heat densities that air cooling cannot efficiently manage at 40–200 kW per rack.

Direct Liquid Cooling (DLC) Coolant circulates through cold plates attached directly to processors. The liquid absorbs heat at the chip level, far more efficiently than air. Requires purpose-built server infrastructure compatible with DLC.

Immersion Cooling Servers are submerged in dielectric (non-conductive) fluid. Two-phase immersion uses a fluid that boils at a low temperature, carrying heat away via phase change. Extremely efficient — can achieve PUE as low as 1.02–1.05.

Why this matters for real estate

Legacy air-cooled data centers were designed for 5–20 kW per rack. AI workloads require 80–200 kW per rack. The structural, electrical, and cooling infrastructure required is fundamentally different:

Heavier floor loads (liquid cooling equipment is dense)
Higher electrical capacity per cabinet position
New plumbing infrastructure for coolant distribution
Different fire suppression systems

This infrastructure gap is obsoleting a meaningful percentage of the existing US data center supply and creating enormous upgrade and new-build demand. Operators and investors who can deliver liquid-cooled capacity are capturing a premium — lease rates for AI-ready space (high-density, liquid-cooled) are commanding 20–40% premiums over standard colocation.

Backup power — UPS and generators

Data centers are marketed on uptime. The major classification system (Uptime Institute Tiers I–IV) is largely a measure of redundancy in power delivery. Mission-critical tenants (financial services, healthcare, government, hyperscalers) require Tier III or IV facilities with N+1 or 2N redundancy.

UPS — Uninterruptible Power Supply

When utility power fails, there is a gap of milliseconds to seconds before backup generators reach full load. UPS systems bridge that gap.

A UPS is a large battery (or flywheel) system that:

Continuously conditions incoming utility power (removes fluctuations, harmonics)
Instantly provides power during a utility outage
Sustains server operation while generators start (typically 5–15 minutes of runtime)

UPS ratings are expressed in kVA/kW and runtime. A 10 MW data center might have 12 MW of UPS capacity (N+1 configuration) providing 10 minutes of runtime.

Redundancy configurations:

N: Exactly enough capacity with no backup
N+1: One additional UPS module beyond what's needed; if one fails, no outage
2N: Fully duplicated UPS systems on separate electrical paths; maximum redundancy

During due diligence, verify UPS age (expected service life 10–15 years), maintenance history, and whether redundancy configuration matches what's claimed.

Generator backup

Generators provide extended runtime beyond UPS. Modern data centers run diesel generators that can sustain full IT load indefinitely as long as fuel is supplied.

Key specs to verify:

| Spec | Typical range | Notes | |---|---|---| | Generator capacity | Slightly exceeds total facility load | Must cover IT load plus cooling plus facility systems | | Fuel storage | 24–72 hours | On-site diesel tank; 48 hours is common for Tier III | | Startup time | 10–30 seconds | UPS bridges the gap | | Redundancy | N+1 minimum | Critical facilities use N+2 or 2N | | Load bank testing | Annual at minimum | Full-load testing under simulated conditions |

Generator due diligence is non-negotiable. Facilities with poor generator maintenance records, undersized fuel storage, or untested redundancy represent significant operational and liability risk. A single unplanned outage can trigger tenant SLA credits worth hundreds of thousands of dollars and permanently damage the facility's reputation.

Request the last three years of generator test logs, maintenance records, and any incident reports during acquisition due diligence.

Utility interconnection — the supply constraint

Everything described above can be engineered and built. Cooling systems, UPS, generators — these are capital deployment problems. The single constraint that cannot simply be bought or engineered is the utility interconnection.

What interconnection means

A data center connects to the electrical grid through a utility interconnection — a formal agreement with the local utility company (Duke Energy, Florida Power & Light, Pacific Gas & Electric, etc.) granting the facility the right to draw a specified amount of power from the grid.

A 100 MW data center needs a 100 MW+ interconnection agreement. The utility must:

Confirm it has available capacity at a grid node near the facility
Upgrade substations, transmission lines, and switching infrastructure if needed
Review, approve, and contract the connection formally

This process takes 18 months to 5+ years in high-demand markets — and the utility may simply say no if the grid is constrained.

The interconnection queue

Utilities manage a queue of interconnection requests from all power users — data centers, manufacturers, EV charging networks, etc. In Northern Virginia (the world's largest data center market), the interconnection queue at Dominion Energy stretched to 3–5 years by 2023–2024. Dominion at one point imposed a moratorium on new large-load interconnection requests.

Similar constraints have emerged in:

Phoenix (Salt River Project and APS at or near capacity in data center corridors)
Chicago (ComEd queue backlog)
Silicon Valley (PG&E grid constraints)
Dallas-Fort Worth (Oncor capacity pressure in key submarkets)

This is not a temporary phenomenon. AI infrastructure demand has grown faster than utilities planned for. A utility that planned for 200 MW of new industrial load per year is suddenly fielding requests for 5,000 MW. The grid needs time and capital to catch up.

Why interconnection position is the most valuable asset

For an existing data center with a signed interconnection agreement covering 100+ MW of capacity, that interconnection is often the most durable competitive advantage:

It cannot be replicated quickly by competitors
It can be expanded (existing interconnections are often easier to upsize than new)
It creates a natural barrier to entry in the immediate submarket
It is a prerequisite for every tenant deployment

When hyperscalers pre-lease data center capacity 3–5 years in advance, they are primarily securing the power — the interconnection position. The building they will worry about later.

This dynamic explains why large data center platforms (Equinix, Digital Realty, Iron Mountain, NXT Data Centers) aggressively acquire interconnection capacity even before they have tenants. The interconnection is the option value.

Expansion power — the most important negotiated term

Beyond existing capacity, an interconnection agreement may include expansion rights: the right to draw additional megawatts at a future date. Facilities with 20 MW of current capacity and expansion rights to 100 MW are significantly more valuable than facilities with 20 MW and no expansion path.

During due diligence, always determine:

Current interconnection capacity (contracted and available)
Expansion rights (MW, timeline, any utility conditions attached)
Whether the utility has indicated willingness to serve expansion
Any competing large-load requests in the same utility zone that could consume capacity

What investors need to verify

Before acquiring or lending against a data center, verify the following power infrastructure items:

Interconnection and utility

Copy of interconnection agreement and any amendments
Contracted capacity vs currently committed capacity vs expansion rights
Utility name and relationship history
Any pending utility rate changes or tariff reclassifications
Local utility capacity constraints — is expansion realistic?

Electrical infrastructure

Total facility power vs critical IT load capacity
UPS capacity, configuration (N+1 / 2N), age, and maintenance records
Generator count, total capacity, fuel storage volume, and load bank test records
Electrical one-line diagram reviewed by a qualified data center engineer

Cooling

Cooling type (air, water-cooled, economizer, liquid)
Rated PUE (and actual measured PUE from utility bills)
Maximum supportable rack density in existing configuration
What density upgrades would require and at what cost

Redundancy classification

Claimed Tier level (if any) — verify against Uptime Institute documentation
Actual N+1 / 2N / 2N+1 configuration by system

Capacity committed vs available

Total contracted IT load under current leases
Uncommitted capacity available for new leasing
Whether existing tenant agreements include expansion options

Gaps in any of these areas are either a negotiating lever or a reason to pass.

What to take away

In data center real estate, you underwrite MW of critical power, not rent per SF — power capacity is the product being sold
Critical power (IT load) is the commercially meaningful figure; total facility power includes cooling and facility overhead
PUE = Total Facility Power / IT Load; best-in-class is 1.1–1.2, industry average is ~1.58; lower PUE directly reduces tenant operating costs and increases facility value
AI GPU clusters require 40–200+ kW per rack, making legacy air-cooled facilities functionally obsolescent for the fastest-growing workload category
Liquid cooling (direct liquid cooling and immersion cooling) is the required infrastructure for AI-dense deployments — it is becoming a prerequisite for competing for hyperscale and AI tenants
UPS provides seconds-to-minutes of bridge power during utility transitions; generators provide extended backup — both require strong maintenance and testing records
Utility interconnection is the primary supply constraint in the sector; it cannot be manufactured quickly, it takes 2–5 years in constrained markets, and existing interconnection positions are the most durable competitive moats in data center real estate
During due diligence, verify the interconnection agreement, expansion rights, utility capacity, UPS and generator records, PUE history, and cooling redundancy — these are the items that determine whether the facility can deliver what it promises

Next lesson: site selection and market fundamentals — why Northern Virginia, Phoenix, and a handful of other markets dominate, and how to evaluate secondary markets emerging around power availability and land cost.

Power Infrastructure — The Most Important Spec in Data Center Real Estate