What it takes to build one
A large AI data center campus is a power project with computers inside. Here is what one needs, in rough order of how hard each is to get, with numbers from the campuses tracked on this site and the sources behind the rest.
01 · Power
The long pole. A 1 GW campus draws about what a large nuclear reactor makes.
- Data centers used about 4.4% of US electricity in 2023 (176 TWh). Berkeley Lab expects 6.7% to 12% by 2028 (325 to 580 TWh).
- Utilities can't connect that much new load quickly. In Texas alone, about 410 GW of large loads were asking ERCOT to connect in March 2026, and about 87% of it was data centers.
- So builders bring their own: 11 of the 58 big campuses tracked here plan on-site or dedicated gas, covering 29.9 GW of their 58.7 GW planned.
- Gas turbines are now scarce too: orders placed in late 2025 were mostly slotted for 2028 to 2030, and Siemens Energy said over 60% of its US gas turbine orders were tied to AI data centers.
Sources: Berkeley Lab, 2024 US Data Center Energy Usage Report · ERCOT large load update, Apr 2026 · Tom's Hardware on turbine wait times, Oct 2025
02 · Grid equipment
Transformers and switchgear, ordered years ahead.
- Large power transformers can take up to four years to arrive (PwC analysts, May 2026).
- Demand for the step-up transformers that connect power plants rose 274% from 2019 to 2025, and for substation transformers 116% (Wood Mackenzie).
- A campus needs its own substation and often new high-voltage lines, so the equipment queue and the utility's connection queue run side by side.
03 · Money
About $29 billion per gigawatt, most of it chips.
- Across 49 US campuses running 100 MW or more, Epoch AI's estimates put the typical all-in cost at about $29 billion per GW (2025 dollars).
- About $20 billion of that is the computers, which wear out in a few years; about $9 billion is the buildings, power and cooling, which last decades.
- That is why owners lean on long leases, project debt and partners: developers build and own the shell, and the chip owner rents it.
04 · Time
About 26 months from ground to first power.
- For 40 US campuses where Epoch AI logged the start of work, the typical gap to first power was about 26 months.
- The fastest, xAI's first Colossus in Memphis, took under 4 months by moving into an existing factory and bringing its own turbines.
- Of the 42 tracked campuses still being built or planned, 19 aim to finish in 2027 and 16 in 2028. Most come online in phases, a building at a time.
05 · Cooling and water
AI racks need liquid cooling; water use depends on the design.
- A rack of current AI chips makes far more heat than air can carry away, so new campuses pipe liquid to the chips and reject the heat outside.
- Evaporative cooling uses a lot of water; closed-loop systems and dry coolers use little but need more power. Local water supply often decides which a site can use.
06 · Permits and neighbors
Air permits for backup and on-site power, zoning, and local votes.
- Backup diesel generators and on-site turbines need air permits. Virginia, the largest US market, now requires Tier 4-equivalent emission controls on new data center generator permits filed from July 1, 2026.
- Turbines run before permits have drawn challenges, as with xAI's units in Memphis, and some towns have said no: Lordstown, Ohio, banned new data centers.
Sources: Williams Mullen on Virginia's Tier 4 rule, Nov 2026 · Epoch AI on Stargate sites, Apr 2026 · Tom's Hardware on xAI's Memphis turbines
07 · Incentives
Most states cut sales tax on the equipment.
- 42 states offer full or partial sales tax exemptions for data centers. A 2025 CNBC review found 16 of them gave up about $6 billion over five years.
- Some states are now adding conditions or rethinking the breaks as data centers push up demand and power bills.
Sources: CNBC review via CO/AI, Jun 2025 · Construction Owners: who's still giving, who's taking back
08 · Chips and networking
The reason for the building, and its biggest power draw.
- Nvidia's data center revenue was $193.7 billion in its fiscal 2026. Its chips set how much power, cooling and floor space a campus needs.
- Moving data between chips takes a growing share of the power, which is why optical links (Ayar Labs, Lightmatter, Xscape Photonics) are drawing money.
Sources: Nvidia fiscal 2026 results
Campus counts update with the tracker. Cost and timing medians are from Epoch AI's estimates for US campuses (CC BY 4.0) as of November 2026 and will move as more campuses finish.