🏛️ Issues and Advocacy Training
Courses › Data Centers: Types, Purposes and How They Work

Lesson 3 of 9 · about 15 minutes

Who Builds Data Centers, and Why Now

You'll be able to

  • Name the main players in data center development and how each makes money
  • Explain the utility's role and why grid connection decides projects
  • Put 2026 hyperscaler spending plans and market vacancy figures in context
  • Trace who actually stands behind an LLC named on a local filing

A data center project usually has five or six parties behind it, and the one whose name is on the zoning application is often the least important. Knowing the cast tells you who can answer which question, and who has money riding on the answer.

The players

PlayerWhat they wantHow they make moneyExamples
HyperscalersCapacity for their own cloud, AI and consumer servicesSelling cloud services, ads, software and AI subscriptions that run in the buildingsAmazon (AWS), Microsoft, Alphabet (Google), Meta, Oracle, Apple
AI labs and AI cloud firmsHuge GPU clusters for training and serving modelsSelling AI products or renting GPU timeOpenAI (with Oracle and others on Stargate), xAI, Anthropic, CoreWeave
DevelopersLand with power, entitled and ready to buildSelling or leasing finished, powered sites, often pre-leased to a hyperscalerPrivate developers, often formed per project
REITs and landlordsLong leases with creditworthy tenantsRent for space, power and connectivityEquinix, Digital Realty, Iron Mountain
Private equity and infrastructure fundsSteady, long-term returnsOwning developers and buildings, then selling or refinancingLarge infrastructure funds and pension investors
UtilitiesNew load that pays its way without hurting other customersRegulated returns on new substations, lines and plants, approved by state commissionsInvestor-owned utilities, co-ops, public power
State and local governmentsTax base, construction jobs, prestigeProperty and sales taxes, though incentives may reduce themEconomic development agencies, counties, cities

Hyperscalers and the AI money

The scale of spending since 2023 is hard to picture. After first quarter 2026 earnings at the end of April, the four largest U.S. hyperscalers' guidance for 2026 capital spending looked like this:

Company2026 capex guidance (as of April 30, 2026)
AmazonAbout $200 billion
MicrosoftAbout $190 billion for calendar 2026
Alphabet$180 to $190 billion
Meta$125 to $145 billion
TotalMore than $700 billion (reports cite about $725 billion)

Read those numbers with care. Capex covers more than data center buildings: it includes chips, servers, network gear, land and some non-data-center property. Chips and servers are a large share. Guidance also changes every quarter, and these companies have mostly revised it upward since 2023. Check the latest earnings release before you cite a number, and say which quarter it came from.

Why spend this much? The companies say demand for AI services exceeds the capacity they can bring online and that falling behind competitors is the bigger risk. Skeptics, including some investors and analysts, argue that revenue from AI has not yet caught up with spending and that some capacity may go underused if demand cools. Both sides are working from the same fact: spending is running ahead of proven AI revenue. They disagree on what happens next. Nobody at a county hearing needs to settle that bet. It does help to know that local projects are being planned on its outcome.

Developers, REITs and the LLC on the application

Big tech companies build some campuses themselves, but they also lease heavily from developers and REITs. A typical sequence goes like this. A developer options farmland near a high-voltage line, signs a load request with the utility, wins rezoning, and then leases or sells the powered site to a hyperscaler. Each project usually sits in its own LLC. That keeps financing separate, and it also keeps the tenant's name quiet while land is assembled and incentives are negotiated. Nondisclosure agreements with local officials are common, and residents often learn about them late.

REITs such as Equinix and Digital Realty own buildings and rent them out. Retail colocation REITs make money from many tenants and the network connections between them. Wholesale landlords sign long leases, often 10 to 15 years or more, with one big tenant per building.

The market is tight

CBRE's report for the first half of 2026, released September 8, 2026, found vacancy across eight primary North American markets at 1.4 percent, the same as at year-end 2025. Construction underway hit a record 7,481 MW, up 24.8 percent year over year, and more than 80 percent of it was already pre-leased. Northern Virginia, the largest market, had 4,496.5 MW of inventory and 0.2 percent vacancy. Stu Dyer of CBRE's Data Center Solutions Group said in the release that power is "the deciding factor" there.

Worked example: what record construction still leaves

Use the CBRE figures above.

  • Under construction: 7,481 MW
  • Share pre-leased: at least 80 percent, so at least 0.80 times 7,481 = 5,985 MW already spoken for
  • Left for new tenants: at most 7,481 minus 5,985 = about 1,496 MW, which CBRE described as roughly six months of supply at the current pace

So even a record building year leaves new customers scrambling. That pressure is what sends developers into rural counties looking for power.

Utilities: the gatekeepers

A building can go up in 12 to 24 months. Delivering several hundred megawatts to it may need a new substation, new transmission lines, and sometimes new power plants. Interconnection queues for new generation are long: Lawrence Berkeley National Laboratory's 2026 Queued Up report found about 2,060 GW of generation and storage waiting at the end of 2025 and median times from request to operation above five years for projects built in 2025. That is why "time to power" now drives site choice more than land price.

Utilities have responded in several ways. They build, and ask the state public utilities commission to approve rates that recover the cost. They write special large load tariffs with minimum bills and long contract terms, so that if a data center leaves early, other customers are not stuck with the cost. And some build new generation dedicated to a campus: for Meta's Hyperion project in Richland Parish, Louisiana, the utility Entergy planned three new natural gas plants, reported at about $3 billion. Who pays for that kind of buildout is the subject of the next module.

Two flagship AI projects

Stargate Abilene (Texas)Meta Hyperion (Louisiana)
WhoCrusoe builds, Oracle runs cloud capacity, OpenAI is the userMeta
Power1.2 GW plannedTotal not given in the source; Entergy planned three new gas plants, about $3 billion
Size8 buildings, about 4 million sq ftOver 4 million sq ft planned on 2,250 acres, plus about 1,400 acres added
Construction workforceAbout 7,000 workers on site daily (late 2025)More than 3,700, with plans for 5,000 (Feb. 2026)
Permanent jobsNot stated in the source500 full-time positions announced
TimelineConstruction began July 2024; first two buildings live by Sept. 2025Announced Dec. 2024

Common mix-ups

  • Claim: The company on the application is the company that will use the building. What the evidence shows: Often it is a developer's single-purpose LLC, and the tenant is disclosed later, if at all.
  • Claim: Every announced gigawatt will get built. What the evidence shows: Utility load requests routinely exceed what gets built. Developers file in several places and drop the slow ones. Treat announced capacity as a ceiling.

Reporting angle

  • Pull the LLC's filing from your secretary of state's business registry. Note the registered agent, organizer and address, and search those names against other projects.
  • Request any nondisclosure agreements signed by city, county or economic development staff under your state's public records law.
  • Search your public utilities commission docket for the utility's large load tariff, any special contract for the project, and load forecasts in its integrated resource plan.
  • Read the hyperscaler's latest quarterly earnings release and call transcript for capex guidance before quoting one.

Research angle

Check yourself

1. As reported at the end of April 2026, the four largest U.S. hyperscalers' combined 2026 capital spending plans were about how much?

2. What did CBRE report for vacancy across eight primary North American data center markets in the first half of 2026?

3. Why do developers often form a new LLC for each project?

4. What does a data center REIT mainly sell?

5. In most of the U.S., what is usually the slowest piece of a large data center project?

Terms in this lesson

Hyperscaler
One of the handful of companies that run cloud and internet services at the largest scale, such as Amazon, Microsoft, Google and Meta.
Capital expenditure (capex)
Money spent on long-lived assets such as buildings, land and equipment, as opposed to day-to-day costs.
REIT
Real estate investment trust: a company that owns income-producing property and must pay out most of its taxable income as dividends.
Pre-leasing
Signing tenants to a building before it is finished. High pre-leasing means demand is outrunning supply.
Large load tariff
A utility rate schedule with special terms, such as minimum bills or long contracts, for customers that need very large amounts of power.
Special purpose entity
A company formed for one project, often an LLC, used to hold land, financing and contracts.

Full glossary

Sources