The Data Center Primer.

Ten facts that frame the data center industry.

The essentials a decision-maker needs before any conversation about data centers.

A factory that turns electricity into computing.

A secure building that houses computers and keeps them powered, cooled and connected around the clock.

Nearly every digital service, from email and banking to streaming and AI, runs on servers inside a data center.

Scale ranges from a single server room to campuses drawing over a gigawatt, more power than a mid-size city.

Three services: compute, store and connect.

Every digital interaction combines processing, stored data and network movement, delivered from a data center.

From computer rooms to AI factories in eighty years.

Each era raised the scale, density and power of the facilities that house computing.

The constant: every generation of computing has needed more power per square foot than the last.

What happens in the half-second after you tap an app.

A request crosses several networks to reach a data center, then the answer travels back, usually in milliseconds.

1. Device

   Phone or laptop sends the request

2. Access network

   Wi-Fi, 5G or home broadband

3. Internet provider

   Routes traffic across its backbone

4. Exchange point

   Networks hand off traffic at a carrier hotel

5. Data center

   Load balancer picks a server

6. Servers and storage

   Compute the answer, read the data

The response returns along the same path.

Carrier hotels such as One Wilshire in Los Angeles are where networks meet; they anchor nearby data center clusters.

Four essentials: power, cooling, connectivity and land.

A site that lacks any one of these cannot host a data center, however good the others are.

Plus people: electricians, engineers and technicians to build it and keep it running 24 hours a day.

From one server to a gigawatt campus.

Data centers are built in repeating blocks. Each level multiplies the one below it.

Typical power at each level

Power is the unit of measure.

A "100 MW data center" usually means 100 MW of IT capacity; the building draws more to run cooling and losses.

What runs inside determines how it is built.

Different workloads need different density, latency, location and resilience.

The shift from cloud to AI workloads is the single biggest change in data center design in two decades.

Record supply, record-low vacancy.

North America primary markets, first half of 2026. New capacity is absorbed as fast as it is completed.

Four forces are driving demand.

AI is the newest and largest driver, layered on top of a cloud migration that is still under way.

Inference, serving AI to billions of users, is expected to become the larger and more durable share of AI demand.

The largest capital program in corporate history.

2026 capital expenditure guidance, $ billions. Most is spent on data centers, chips, servers and power.

Capex is now outrunning cash flow: hyperscalers are raising debt and equity to fund the build.

About 20% of total data center spend goes to buildings, power and cooling; about 80% to chips, servers and networks.

Data centers are becoming a major electricity user.

U.S. consumption could double or triple by 2028; global use is expected to roughly double by 2030.

How the industry fits together.

Capital funds powered land, developers build, operators run, tenants fill the space and serve end users.

1. Capital

   REITs, PE and infrastructure funds, hyperscaler balance sheets

2. Powered land

   Sites with secured utility power and fiber

3. Development

   Design, permits, financing, construction management

4. Construction

   General contractors, MEP trades, equipment makers

5. Operations

   Facility management, security, maintenance

6. Tenants

   Cloud, AI, enterprise, carriers, content

7. End users

   Consumers, businesses, governments

Hyperscalers span the chain: they self-build, lease from developers, and sell cloud capacity to end users.

Six ways to sell data center capacity.

Models differ by who owns the building, who owns the IT, contract length and margin.

The same building can serve several models; operators blend wholesale scale with retail and interconnection margin.

Hyperscalers build, operators lease, funds own.

Hyperscalers are both the largest builders and the largest tenants of third-party capacity.

Private equity and infrastructure funds own much of the third-party sector; hyperscalers increasingly co-fund builds.

Who buys capacity, and what they buy.

Pricing is usually quoted per kW of committed power per month; primary-market rents are rising 4–8% a year.

Growth goes wherever power can be delivered fastest.

Northern Virginia is still the largest market; Atlanta now leads on construction.

CBRE inventory counts leased third-party capacity; other trackers include self-built and planned sites.

Five constraints decide what gets built.

Capital and demand are plentiful. Power, equipment, people and permission are not.

30–50% of planned 2026 U.S. capacity could slip or be cancelled because electrical gear cannot be delivered on time.

California: strategic, expensive and power-constrained.

Tracked facilities and projects, August 2026. Power figures are reported or announced and use mixed measures.

Five California clusters and the utilities that serve them.

The utility determines power cost, interconnection timeline and the rules a project must follow.

Local pushback is rising: Monterey Park voters banned data centers citywide in June 2026; Coachella imposed a moratorium.

Where California capacity is being built.

Largest published power figures. Measures differ (IT load, campus power, backup generation), so treat as indicative.

The companies that control California capacity.

Operators, developers and hyperscalers, with their California footprint and market position.

Utilities matter as much as operators: PG&E, Silicon Valley Power, SMUD, LADWP, SCE, Vernon PU and IID.

California now makes data centers pay their own way.

SB 886 and AB 2383, the California Technology Innovation and Ratepayer Protection Act, signed September 21, 2026.

Higher grid costs raise the value of anything that reduces a site’s peak demand or required interconnection capacity.

Seven facility types, defined by owner, user and power.

The fastest growth is in hyperscale and AI campuses; enterprise sites are consolidating.

Colocation is 85% of tracked California sites by count; hyperscale is only 4% by count but a growing share of power.

How a colocation building is organized.

One building, many customers: shared power, cooling and security, with private space sold by the cabinet, cage or suite.

Unit sold	What it is
Cabinet	One rack, typically 5–15 kW, in a shared, secured row
Cage	A fenced, locked area holding a group of racks
Suite	A private room with its own access control
Hall / building	Wholesale: an entire hall or building for one tenant
Cross-connect	A cable linking two customers, carriers or clouds

The meet-me room is where carriers and customers connect. Dense interconnection is what makes retail sites valuable.

An AI factory is a different kind of building.

AI clusters concentrate far more power and heat in each rack, which changes almost every design choice.

Ten factors decide whether a site works.

Power and fiber come first; the rest determine cost, speed and long-term risk.

From a long list of markets to one secured site.

A disciplined funnel narrows options quickly and spends diligence money only on the finalists.

Developers increasingly buy "powered land": sites where the utility has already committed capacity.

Speed to power now decides where capacity is built.

Grid connections can take years, so developers combine several paths to energize sites sooner.

California’s new law credits zero-emission resources installed behind the meter against generation charges.

Five systems keep the computing running.

Power flows in from the utility to the racks; heat flows out from the racks to the atmosphere.

Roughly half of a site is white space for IT; the rest is grey space for power, cooling and support.

A typical data center campus, from above.

Illustrative plan of a two-building campus with room for a third phase.

Element	Role
Substation	Steps down transmission voltage for the campus
Generators	Backup power; typically N+1, 24–72 h fuel
Buildings	Data halls plus electrical and mechanical rooms
Cooling plant	Rejects heat from all buildings
Fiber A / B	Diverse entries so one cut cannot isolate the site
Expansion pad	Land and power reserved for later phases
Perimeter	Fence, gates, cameras, guard house

An industrial building designed around power.

The structure must carry heavy equipment, resist hazards and leave room for power and cooling distribution.

Hot aisles, cold aisles: how air moves through a data hall.

Racks face each other across a cold aisle; servers pull cold air in front and exhaust hot air out the back.

Key concepts

• Containment: physically separating hot and cold air saves fan and chiller energy.
• Inlet temperature: ASHRAE recommends about 18–27°C at the server intake.
• Blanking panels seal empty rack slots so hot air cannot recirculate.
• Many new halls skip the raised floor: slab floors with overhead air, busway and cable trays.
• Above ~30–40 kW per rack, air alone cannot remove the heat; liquid takes over.

Airflow in a data hall, without and with containment.

Cold air rises through the floor into the cold aisle, passes through the servers and leaves as hot exhaust.

Containment is one of the cheapest efficiency upgrades: it lets operators raise supply temperatures and slow the fans.

The power chain: from grid to chip, without interruption.

Electrical systems are the largest cost block, typically 45–54% of construction cost.

1. Utility and substation

   Grid feed at transmission or distribution voltage; on-site substation for large sites; dual feeds preferred.

2. Switchgear and transformers

   Medium-voltage switchgear and transformers step power down to building voltage; long-lead items.

3. UPS and batteries

   Uninterruptible power bridges any disturbance instantly; lithium-ion is replacing lead-acid.

4. Distribution

   Power distribution units, panels and overhead busway carry power to each row and rack.

5. Rack and IT

   Rack PDUs feed dual server power supplies; AI racks draw 40–130+ kW.

A and B: two independent paths to every server.

A simplified "2N" one-line diagram. Either path alone can carry the full load if the other fails.

How to read it

• Each path has its own utility feed, switchgear, transformer, UPS and distribution.
• If utility power fails, UPS batteries carry the load until generators start.
• If an entire path fails or is under maintenance, the other carries 100% of the load.
• Lower-cost designs share some components (N+1) instead of fully duplicating them.

When the grid fails: the first 30 seconds.

Batteries carry the load instantly, generators take over within seconds, and IT never notices.

Timings are typical, not universal.

N, N+1, 2N: how much spare capacity is built in.

"N" is the capacity needed to carry the load. Each level adds spares, and cost, for higher reliability.

Cost uplifts are indicative relative to an N design for the affected systems; actual premiums depend on design and scale.

The cooling chain: from chip to atmosphere.

Heat is captured at the chip or in the room, carried by air or water, and rejected outdoors.

1. Chip and server

   Server fans push air across heat sinks; AI chips use cold plates with liquid coolant.

2. Rack and row

   Hot / cold aisle containment, rear-door heat exchangers, coolant distribution units.

3. Data hall

   Computer room air handlers (CRAH / CRAC) or fan walls move conditioned air.

4. Central plant

   Chillers make chilled water; pumps and heat exchangers; optional thermal storage.

5. Heat rejection

   Cooling towers (evaporative), dry coolers or adiabatic coolers release heat outdoors.

Two water loops carry heat out of the building.

The chilled water loop collects heat from the halls; the condenser loop carries it to the cooling towers.

How a chiller moves heat out of the water loop.

A refrigeration cycle takes heat from the chilled water and hands it to the condenser water and cooling towers.

1. Evaporator

   Cold refrigerant absorbs heat from the building's chilled water, which leaves at roughly 7–18 °C.

2. Compressor

   Raises the refrigerant's pressure and temperature; this is where most of the chiller's electricity goes.

3. Condenser

   Hot refrigerant gives up its heat to the condenser water, which carries it to the cooling towers.

4. Expansion device

   Drops the pressure so the refrigerant turns cold again and the cycle repeats.

Free cooling bypasses the compressor when outside air is cold enough.

Cooling approaches, from air to immersion.

Most sites combine several: air for general IT, liquid for high-density racks, economizers when weather allows.

Densities indicative.

About one in five data centers used liquid cooling by early 2026; adoption is accelerating with each GPU generation.

Liquid cooling: bringing water to the chip.

Water carries heat thousands of times more effectively than air, which is why AI racks depend on it.

Three water loops connect the cooling tower to the chip.

A coolant distribution unit (CDU) separates the clean coolant in the racks from the building's facility water.

1. Condenser water (CWS)

   Cooling tower to chiller: rejects the heat to the atmosphere.

2. Facility water (FWS)

   Chiller to the CDUs: the building's main cooling loop.

3. Technology cooling (TCS)

   CDU to the cold plates on each chip: a small, sealed loop of treated coolant.

Warm-water designs run the facility loop hot enough to skip the chiller for much of the year, using only dry coolers.

Chip power decides when air cooling runs out.

As processors pass a few hundred watts each, air can no longer carry the heat away and liquid takes over.

Reading the chart

• The horizontal bar is power per processor, from 75 W to 600 W and beyond.
• Air cooling (top) works for chips up to roughly 300–400 W, depending on server size.
• Liquid cooling (bottom) covers the whole range and is the only option at the top end.
• Today's AI GPUs draw around 1,000 W or more each, far past the air-cooled limit.

Water use and heat reuse are now design decisions.

Cooling choices trade energy against water; waste heat is increasingly treated as a resource.

The electronics drive every other design decision.

The IT inside often costs more than the building, and is replaced every 4–6 years.

Each GPU generation raises power and heat per rack, so facilities must be designed for the next cycle, not the current one.

Spine and leaf: how data moves inside the building.

Traffic enters through the meet-me room and fans out through core, spine and leaf switches to every server.

Key points

• Every leaf connects to every spine, so traffic between any two racks crosses the same number of hops.
• Diverse carriers and fiber routes avoid a single point of failure.
• Cloud on-ramps give private, direct links to AWS, Azure and Google.
• AI clusters add a separate, very high-bandwidth fabric just for GPU-to-GPU traffic.

From one GPU to a 100,000-GPU cluster.

AI systems are built in tightly linked blocks; the rack has become the unit of design.

Thousands of sensors, one picture of the facility.

Layered control systems turn raw measurements into alarms, automation and decisions.

Layers of protection, from the fence to the cabinet.

Each layer adds a check; reaching a server means passing every one of them.

The standards that define a well-built data center.

Owners, tenants, insurers and auditors rely on a common set of design, safety and operating standards.

Four tiers of reliability, from basic to fault tolerant.

The Uptime Institute Tier system is the common language for how resilient a facility is designed to be.

From a requirements brief to construction drawings.

Design moves through defined stages; each ends with a cost check and an owner sign-off.

1. Owner requirements

   Capacity, density, tier, budget and schedule agreed with the owner; designers respond with a basis of design.

2. Concept

   Site plan, building massing, power and cooling strategy, early cost estimate.

3. Schematic

   Room layouts, one-line diagrams, equipment sizing, utility coordination.

4. Design development

   Detailed systems, coordination between trades, BIM model, refined cost.

5. Construction documents

   Drawings and specifications for permits, bids and construction.

6. Procurement

   Long-lead equipment ordered early; issued-for-construction set released.

How owners contract for design and construction.

The delivery model sets who carries design, cost and schedule risk, and how fast the project can move.

Hyperscalers increasingly use repeatable designs, framework contracts and prefabrication to build in parallel at many sites.

Nine approvals stand between a site and a building.

Typical U.S. permitting takes 6–18 months; contested or re-zoned sites can take 2+ years. California examples shown.

Air permits and grid interconnection usually set the schedule, not construction.

From site to live capacity: typically two to four years.

Indicative greenfield schedule in months. Phases overlap; power and long-lead equipment drive the critical path.

How a data center goes up, phase by phase.

Phases overlap heavily; electrical and mechanical work is the largest and longest part of the build.

1. Site work

   Clearing, grading, drainage, roads, underground utilities and duct banks.

2. Foundations

   Piles or footings and heavy slabs; thousands of anchors drilled.

3. Structure and shell

   Steel or precast frame, roof, walls; building made weathertight.

4. MEP rough-in

   Conduit, cable tray, piping, busway and ductwork installed.

5. Equipment setting

   Switchgear, UPS, generators, chillers and CDUs placed and connected.

6. Fit-out and test

   Containment, racks, controls, then commissioning and handover.

Transformers and switchgear are under 10% of cost but decide whether a facility opens on time.

Equipment lead times now exceed construction times.

2026 lead times for critical electrical equipment, versus pre-2020 norms. Orders are placed before design is final.

Five levels of testing before a single server goes live.

Commissioning proves every system works alone and together, including under failure, before handover.

An independent commissioning agent usually leads the process for the owner.

1. L1 Factory witness testing

   Equipment tested at the factory before shipping

2. L2 Site acceptance

   Delivered and installed correctly, undamaged

3. L3 Start-up

   Each component energized and started individually

4. L4 Functional testing

   Each system tested in every operating mode

5. L5 Integrated systems test

   Whole facility under load banks; failures simulated

Level 5, the "pull-the-plug" test, cuts utility power and fails equipment on purpose to prove the design works.

A digital twin is a living virtual copy of the facility.

A 3D, physics-based model kept in sync with real data, used to simulate, predict and optimize the physical site.

Insights and decisions flow back: what to change, where to add capacity, what will fail next

Build it virtually first, then build it for real.

Digital twins cut risk at every stage, from design choices to day-to-day cooling control.

Vendor-reported: AI-driven cooling control in a digital twin can yield about 10% more compute from the same power.

Example cited in NVIDIA’s Omniverse DSX blueprint (Phaidra, March 2026). Results vary by site.

One model that grows from design to operations.

The same twin is handed from designers to builders to operators, gaining data at every step.

The value compounds: a twin started at design is far cheaper and more accurate than one built after the facility opens.

Five layers make a working digital twin.

Data flows up from the facility; simulations and AI turn it into decisions people can act on.

Eight ways operators use a digital twin.

Each use case answers a question that is expensive, slow or dangerous to test on the real facility.

Six steps to plan and implement a digital twin.

Start with the decisions the twin must support, then build only the data and models those decisions need.

1. Define goals

   Pick the decisions and use cases: design options, capacity, cooling, maintenance.

2. Inventory data

   BIM and CAD, equipment lists, sensors, BMS, power and DCIM data, IT telemetry.

3. Choose platform

   Simulation tools and a data model; open formats and APIs avoid lock-in.

4. Build the model

   3D geometry and accurate equipment models; vendor "SimReady" assets help.

5. Connect live data

   Stream BMS, power and IT data; set update rates and data quality rules.

6. Validate, expand

   Calibrate simulations against measurements, then add use cases.

Five levels of digital twin maturity.

Most facilities today sit at levels 1–2; leading AI factories are moving toward levels 4–5.

1. Level 1: Descriptive

   A 3D model with static asset data

2. Level 2: Informative

   Live monitoring data shown in the model

3. Level 3: Predictive

   Simulation forecasts what will happen

4. Level 4: Prescriptive

   The twin recommends what to do

5. Level 5: Autonomous

   AI acts on recommendations within set limits

The industry is standardizing on digital twins.

NVIDIA released the Omniverse DSX blueprint for AI factory digital twins on March 16, 2026, with broad partner support.

Digital twins are becoming a standard deliverable for large AI campuses, from first design through daily operations.

Running a facility that can never switch off.

A control room at the center coordinates engineering, IT, security and customer teams around the clock.

Who runs a data center: a typical site organization.

Illustrative structure for a large operating site; headcount scales with size and service model.

Site / data center director: Safety, uptime, customers, budget.

Large campuses add energy, capacity and reliability engineers; smaller sites combine roles or outsource them.

The people who keep a data center running.

Operations teams are small relative to construction crews, but highly skilled and on call around the clock.

Who builds a data center, and when they are on site.

A large campus can peak at several thousand workers; electrical work is roughly half of all data center labor.

People are as scarce as transformers.

The build-out competes for the same electricians and engineers as utilities, factories and other construction.

Maintenance is how uptime is earned.

Equipment is serviced on a strict calendar, with written procedures, so failures are found before they happen.

Most outages start with power, or with people.

Outages are becoming less frequent but more expensive; disciplined change control prevents most of them.

1. Request

   Describe the change, the reason and the systems affected.

2. Risk review

   Assess impact on redundancy and customers.

3. Write the MOP

   Step-by-step method, checks and rollback plan.

4. Approve

   Change board signs off; customers notified.

5. Execute

   Two-person verification at each step.

6. Verify and close

   Confirm normal state; record lessons.

80% of operators believe better management and processes would have prevented their most recent outage.

UPS failures are a leading cause of power outages; failure to follow procedures is a rising cause of human-error incidents.

Energy, water and carbon are now measured and reported.

Customers, regulators and communities expect data centers to show efficiency and a path to clean power.

Efficiency is also economics: every point of PUE saved is power that can run revenue-generating IT instead.

Ten numbers every operator watches.

Operating KPIs measure reliability, efficiency and how much of the built capacity is earning money.

Robots are arriving on data center sites.

Labor shortages, repetitive precision work and schedule pressure are pulling robots onto construction sites and into operations.

ONE PILOT, BY THE NUMBERS: AUTONOMOUS DRILLING AT A LEADING HYPERSCALER

Robots take the repetitive, precise and dangerous tasks; skilled people supervise, handle exceptions and do the complex work.

Seven construction tasks robots already perform.

Most construction robots handle one task very well; they are guided by the BIM model and supervised by crews.

The next frontier is cable pulling and installation, the largest labor block in a data center build.

Robots in the running data center.

Inspection and patrol robots are in use today; hands-on hardware tasks are emerging.

Four deployments that show where this is going.

Real projects on data center sites, from construction drilling to humanoids in operations.

From human crews to hybrid human-robot teams.

Robots join the work task by task; people move toward supervision, exceptions and complex work.

The goal is not fewer skilled people, but more capacity built and run with the skilled people available.

An indicative roadmap for robotics adoption.

Adoption moves from narrow, proven construction tasks toward broader autonomy in operations.

Customers rent power, not square feet.

The core charge is dollars per kilowatt of committed capacity per month, on leases of 1–15 years.

Rents are rising in every size band.

Asking rents in U.S. dollars per kW per month. Larger deployments pay less per kW but sign longer.

Rent growth, H1 2026 year over year (CBRE)

About $11 million per megawatt, before land and IT.

Costs are quoted per MW of critical IT capacity. AI-ready builds cost far more once liquid cooling and denser power are added.

From the building to a fully loaded AI megawatt.

Each layer adds cost: the building, the AI-ready infrastructure, then the computing equipment itself.

Electricity is the biggest running cost, and cooling drives it.

Annual cost per MW of IT load. At California power prices, electricity costs more than paying off the building.

Cooling overhead is the controllable share: at PUE 1.5, one-third of every power dollar goes to non-IT load.

A 20 MW colocation facility, in round numbers.

Illustrative model built from the market benchmarks in this primer. Not a forecast for any specific site.

A PUE reduction of 0.1 saves ~11.6 GWh a year: ~$1.4M at 12¢/kWh, ~$2.5M at ~21.5¢/kWh, captured by whoever pays the power bill.

From powered land to a stabilized, valuable asset.

Revenue starts only when capacity is delivered, so speed to power is speed to cash.

1. Secure land and power

   Site control and a utility commitment: the scarcest asset today.

2. Pre-lease

   Leases signed before or during construction; 80%+ preleased.

3. Lease terms

   10–15 years hyperscale, 1–5 years retail; escalators; power floors.

4. Build and deliver

   Halls delivered in phased blocks as each is commissioned.

5. Rent commencement

   Rent starts on delivery; tenant IT load ramps over months.

6. Monthly billing

   Recurring rent plus metered power, interconnection and services.

7. Renew and expand

   Renewals re-price upward (+6.5–8.5%); tenants expand in place.

8. Stabilize and recycle

   Refinance, sell or form a joint venture at ~6% cap rate.

The bottleneck has shifted from finding tenants to delivering power and capacity on schedule.

The financial scorecard of a data center business.

The metrics investors, lenders and operators use to judge development and operating performance.

Returns come from the spread, and the spread needs cheap capital.

Developers build at ~10–11% yields, borrow at ~6–7% and sell or value stabilized assets at ~6% cap rates.

Financing has shifted: hyperscaler debt rose from 9% of capex in FY24 to about 32% by mid-2026.

Build at cost, lease up, and value rises by about 80%.

Using the illustrative 20 MW facility: value at a 6% cap rate versus the cost to build it.

AI is rewriting the data center rulebook.

Six shifts that separate AI factories from the cloud data centers of the last decade.

Facilities designed today must anticipate at least two more generations of AI hardware over their life.

Powering the build-out: the energy options.

No single source will meet demand; operators are assembling portfolios of grid, on-site and contracted power.

Storage and demand flexibility reduce the peak power a site must draw from the grid, which can speed interconnection.

Six trends shaping the next generation of facilities.

Technology, construction and operating models are all changing at once.

Demand stays ahead of supply through 2027.

The leaders will deliver power and capacity fastest, and run it most efficiently.

The industry’s challenge is no longer finding demand; it is delivering power, equipment and people on time.

Ten things to remember about data centers.

The essentials of the primer on one page.

Key numbers and rules of thumb.

Indicative figures for fast reference in discussion.

The organizations that shape the industry.

Standards bodies, research groups and associations that define practice and represent the sector.