Every digital action you take — streaming a video, sending an email, asking an AI a question, making a credit card payment — routes through a data center. The physical building housing the servers is commercial real estate. And for the first time in history, that real estate is the most capital-intensive, most constrained, and most institutionally sought-after asset class in the world.

This lesson explains what data centers are, why the investment thesis is unlike anything else in CRE, and why AI has turned a strong secular growth story into one of the most urgent investment moments in modern real estate history.

What is a data center

A data center is a specialized building that houses computing infrastructure — servers, networking equipment, storage arrays, and the power and cooling systems required to keep that equipment running 24 hours a day, 365 days a year, without interruption.

At its core, a data center is an industrial building with extraordinary mechanical and electrical systems. The building itself is often unremarkable — a large concrete box. The value is in what's inside and what powers it.

The physical building

A modern data center campus might include:

White space — the raised-floor areas where servers and networking equipment are installed
Mechanical rooms — cooling systems (chillers, CRAC units, cooling towers, liquid cooling loops)
Electrical infrastructure — utility feed, transformers, switchgear, UPS (uninterruptible power supply) systems, backup generators and fuel tanks
Network rooms — fiber entry points and interconnection equipment
Administrative space — security, operations, and management areas
Exterior infrastructure — cooling towers, fuel tank farms, utility transformers, security perimeter

The ratio of "critical" power actually delivered to servers (IT load) to total power consumed by the entire facility is measured by PUE (Power Usage Effectiveness). A perfect PUE is 1.0 (all power goes to servers). Modern hyperscale facilities achieve PUE of 1.2-1.4. Older colocation facilities often run 1.5-2.0.

Three types of occupancy

Not all data centers are the same. Three primary occupancy models drive the market:

Hyperscale — One massive tenant occupies an entire campus or building. The hyperscale tenants are the largest technology companies in the world: Amazon Web Services (AWS), Microsoft Azure, Google Cloud, Meta, Apple, Oracle, and a handful of others. Hyperscale campuses can run 50-500+ megawatts (MW) of power. Leases are 15-20 years, NNN. The tenant designs the building to their specifications (often a powered shell where the tenant builds out the interior). This is the closest analog to a build-to-suit industrial development but at extraordinary scale.

Colocation ("colo") — Multi-tenant facilities where multiple companies rent rack space, cage space, or private suites. Equinix and Digital Realty are the dominant publicly traded colo operators. Enterprises, cloud companies, and service providers rent space from months to years. Pricing is typically quoted per kilowatt (kW) of power capacity. Think of colocation as a multi-tenant industrial building — but instead of square footage, you're renting power.

Enterprise private — A single company builds or owns a data center exclusively for its own operations. Banks, healthcare systems, government agencies, and large manufacturers often operate private data centers. These are rarely investment properties but do trade occasionally.

For CRE investors, the relevant markets are hyperscale (institutional development) and colocation (institutional ownership). Enterprise private is largely self-contained.

The numbers behind the category

Data center real estate has moved from niche alternative to dominant institutional asset class. The numbers explain why.

Investment pipeline

Global data center construction investment exceeded $200 billion in 2024 and is projected to exceed $300 billion annually by 2027. For comparison, total US commercial real estate construction in a typical year runs $100-150 billion across all property types combined. Data centers are now consuming more investment capital than every other CRE category added together — globally.

Vacancy

Vacancy in top US data center markets has fallen to historic lows:

| Market | Vacancy Rate | |---|---| | Northern Virginia (primary) | < 1% | | Chicago | < 3% | | Dallas-Fort Worth | < 4% | | Phoenix | < 4% | | Atlanta | < 5% | | Silicon Valley | < 3% |

Northern Virginia — specifically Loudoun County, known as "Data Center Alley" — is the densest concentration of data center capacity on earth. Vacancy is effectively zero. Tenants are signing leases on buildings that don't yet have power, years before delivery.

Cap rates and REIT performance

Institutional data center product trades at 5.0-6.5% cap rates for stabilized assets with long-term leases to investment-grade tenants. Development yields (yield-on-cost) for new hyperscale campuses run 6.5-8.5% before stabilization, often higher for merchant developers who lease to a single hyperscale tenant.

The publicly traded data center REITs have outperformed all major CRE sectors over the past decade:

| REIT | Focus | Market Cap (approx.) | |---|---|---| | Equinix (EQIX) | Global colocation, interconnection | ~$75B | | Digital Realty (DLR) | Hyperscale + colo, global | ~$55B | | Iron Mountain (IRM) | Colo, records management | ~$25B | | QTS Realty | Hyperscale + colo (now private, KKR) | Acquired | | CyrusOne | Hyperscale (now private, KKR/GIP) | Acquired |

The takeaway: the largest institutional investors in the world — Blackstone, KKR, Brookfield, GIP, Carlyle — have been actively acquiring and developing data center assets at scale. This is not a speculative asset class. It is the preferred institutional bet on digital infrastructure.

Why AI is a structural inflection point

The data center market was already strong before AI. AI has transformed it from a growth story into a structural scarcity event.

Power density explosion

Traditional enterprise data centers designed for 2-5 kilowatts (kW) per server rack.

AI training clusters designed for 10-30+ kW per rack — and next-generation GPU clusters from NVIDIA (H100, B200, GB200 systems) are pushing 40-100+ kW per rack.

This is not a modest upgrade. A single rack of NVIDIA B200 GPUs consumes as much power as 10-20 racks of conventional servers. The same building footprint can only fit a fraction of the AI infrastructure that would fit in a traditional data center without significant power and cooling upgrades.

The practical result: almost every existing data center in the world is functionally undersized for AI workloads. New construction designed for AI power densities is the only solution.

AI training campus scale

A single large-scale AI training campus — one facility for one AI company training one major model — can consume:

50-200+ MW of power
Equivalent to the peak electricity consumption of a city of 50,000-200,000 people
Hundreds of millions of dollars in construction cost

ChatGPT, Gemini, Grok, Claude, and every other major AI system in the world runs on purpose-built data centers. Every major enterprise is now building AI inference and training infrastructure. Microsoft announced $80 billion in data center investment for 2025 alone. Amazon, Google, and Meta are each planning similar scale.

This demand is not cyclical. It is not tied to interest rates. It is tied to the fundamental shift of global industry toward AI-enabled operations — and that shift is accelerating.

The multiplier effect

For every AI training cluster, there is an inference cluster (the infrastructure that runs the model after it's trained). Inference demand is often 5-10x training demand by compute hours. Every enterprise that deploys an AI application generates continuous inference workload.

AI also drives secondary data center demand:

Cloud capacity to serve AI APIs
Data storage for training datasets
Networking for GPU interconnects
Edge data centers for low-latency AI inference

The AI multiplier runs through the entire data center stack.

Why data centers are different from other CRE

Data centers share some characteristics with other CRE asset classes — especially industrial — but several factors make them fundamentally distinct.

Power, not square footage

The primary unit of data center capacity is megawatts (MW) of critical IT power, not square feet. A developer might lease a 200,000 SF data center with 20 MW of critical power, or a 200,000 SF data center with 80 MW of critical power. The 80 MW campus is worth 3-4x more — same building footprint.

Investors, lenders, and appraisers underwrite data centers on a cost per MW and revenue per kW basis:

Annual rent per kW = Monthly rate × 12
(Typical hyperscale: $80-120/kW/month → $960-1,440/kW/year)

NOI = Critical power (kW) × Annual rent per kW × Occupancy
Cap rate = NOI ÷ Value

A 20 MW colocation facility at 80% occupancy, $100/kW/month:

20,000 kW × 80% × $100/kW/month × 12 months = $19.2M NOI
At 6% cap rate → $320M value

A traditional industrial building of the same square footage would trade at $40-80M. The power infrastructure creates a 4-8x value premium.

Location determined by power, not demographics

Industrial real estate is located near highways, labor, and customers. Multifamily is located near jobs and amenities. Data centers are located where power is available — specifically, where large utility substations can provide 20-500 MW of capacity at competitive rates.

The secondary location factors:

Fiber connectivity — multiple diverse fiber paths to major internet exchanges
Water — for cooling (data centers are major water consumers)
Tax incentives — many states offer significant tax abatements for data center investment
Seismic risk — away from major fault zones
Natural disaster risk — flood plains, tornado corridors, hurricane paths all matter
Grid reliability — markets with stable, redundant utility infrastructure

This is why Northern Virginia dominates. Affordable, abundant power. Excellent fiber. Favorable tax treatment (Virginia data center tax exemption). One of the strongest utility grids in the country.

Tenant quality and lease structure

Data center leases are among the most favorable of any CRE asset class:

Lease terms: 15-20 years for hyperscale, 5-10 years for colocation suites
Tenant credit: AWS (Amazon), Microsoft Azure, Google Cloud, Meta, Oracle — these are some of the strongest corporate credits in the world
NNN structure: Tenants pay power costs directly (often the largest operating expense)
Renewal probability: Tenants who build out a data center campus almost never leave — the cost to migrate petabytes of data and re-equip a new facility is enormous
Escalations: Typically 2-3% annual CPI-linked increases

A 15-year hyperscale lease to AWS is arguably safer long-term income than a 15-year NNN lease to a major pharmacy chain. The credit quality and stickiness of hyperscale tenants is unmatched in commercial real estate.

Supply is structurally constrained

New data center supply faces constraints unlike any other CRE category:

Utility lead times: A new 100 MW data center campus requires a new utility substation or major grid upgrade. Utility construction timelines in most US markets run 2-5 years from application to energization. This is not a permitting issue — it is a physical infrastructure issue. Building a new transformer, upgrading transmission lines, and provisioning generation capacity takes years.

Equipment shortages: Specialized electrical equipment — large transformers, switchgear, UPS systems — has faced lead times of 2-3 years due to supply chain constraints and explosive demand. Generator manufacturers and cooling equipment suppliers face similar backlogs.

Land: Data center campuses require large, flat sites with direct utility access, multiple fiber paths, and water. Such sites near major markets are increasingly scarce and expensive.

Regulatory approvals: Local zoning, utility approvals, environmental review, and water permits all add time. Some markets have imposed moratoriums on new data center development due to grid capacity concerns (Northern Virginia, parts of the Netherlands and Ireland).

The result: even with unlimited capital, a developer cannot build a data center campus in less than 2-4 years from site identification to energization. This structural supply constraint is the fundamental driver of the pricing power enjoyed by existing data center owners.

The three investment theses

Data center investors pursue three primary strategies, each with different risk-return profiles.

Thesis 1: Own stabilized hyperscale assets

What it is: Acquire existing data center campuses already leased to hyperscale tenants on long-term NNN leases.

Profile: 5.0-6.5% going-in cap rate, investment-grade tenants, 10-20 year WALT, minimal management. Similar in risk profile to a sale-leaseback NNN deal — but the tenant is AWS rather than a restaurant chain.

Who does it: Large institutional investors (sovereign wealth funds, pension funds, REITs). Typical transaction size: $500M-$5B+.

Pros: Maximum safety, passive income, strong credit, minimal management intensity.

Cons: Low initial yield, limited upside beyond inflation escalations, requires massive capital.

Best for: Ultra-institutional capital seeking safe, long-duration income.

Thesis 2: Develop powered shell for hyperscale tenants

What it is: Acquire a site with utility access and fiber, permit and develop a "powered shell" (the structure, electrical infrastructure, and cooling rough-in), and pre-lease to a hyperscale tenant before or during construction.

Profile: 7.0-9.0% yield-on-cost, development risk during construction, near-immediate stabilization once leased. Typical development cost: $8-15M per MW all-in.

Who does it: Specialized data center developers (Compass Datacenters, aligned Data Centers, Stack Infrastructure), increasingly mainstream CRE developers partnering with power infrastructure firms.

Pros: Best risk-reward ratio in data center investing. Once the lease is signed, it behaves like a stabilized hyperscale asset. Short path to exit via sale to institutional buyer at compressed cap rate.

Cons: Requires access to land with power (the hardest part), specialized development expertise, capital intensity, pre-leasing risk.

Best for: Experienced developers willing to invest in the utility relationship and site pipeline.

Thesis 3: Operate colocation

What it is: Own and operate a multi-tenant colocation facility, leasing rack space, cage space, and private suites to enterprises, cloud companies, and service providers.

Profile: 6.5-8.5% stabilized yield, higher management intensity, revenue management, tenant diversification, recurring cross-connect revenue.

Who does it: Equinix, Digital Realty, QTS, and hundreds of regional colo operators. Active investors can enter at the regional level.

Pros: Higher yield than stabilized hyperscale. Revenue diversification across many tenants. Cross-connect and managed services revenue. Strong recurring cash flows once stabilized.

Cons: Operating intensity (24/7 staffed operations, sales team, customer support, technical expertise required). Longer lease-up than industrial. Competition from hyperscale cloud (AWS, Azure, Google) commoditizing enterprise workloads.

Best for: Operators with data center expertise, or investors partnering with an experienced operating platform.

What to take away

A data center is a specialized commercial building housing computing infrastructure — the physical backbone of every digital service
Three occupancy types: hyperscale (single massive tenant), colocation (multi-tenant rack rental), enterprise private (single company)
Global data center investment exceeds $200B annually and is growing rapidly — more capital than all other US CRE construction combined
Vacancy in top markets is below 5% and falling; Northern Virginia is effectively 0%
Institutional cap rates compress to 5-6% for stabilized product; REITs have outperformed all major CRE sectors
AI is not a cyclical driver — it is a structural inflection point. AI clusters require 10-30x more power per rack than traditional servers. A single AI campus can consume 50-200 MW, equivalent to a small city
Data centers trade on MW of power, not square feet — a 200,000 SF campus can be worth $80M or $400M depending on its power capacity
Tenant quality is unmatched: AWS, Microsoft, Google, Meta sign 15-20 year NNN leases and almost never leave
Supply is constrained by utility lead times (2-5 years), equipment shortages, and scarce powered land — not by capital or permitting
Three investment theses: stabilized hyperscale (low yield, max safety), powered shell development (best risk-reward), colocation operation (highest yield, most complex)

Next lesson: Data center fundamentals — power, cooling, and infrastructure. Understanding how data centers actually work makes you a better investor and a better counterpart to the engineers and operators who run them.

Why Data Centers — The Infrastructure Bet on Everything Digital