AI data centers: power, cooling, sites and the build-out

What makes an AI data center different

Density. A conventional enterprise rack has run at 10 to 30 kW for years, and the Uptime Institute's 2025 survey found few facilities above 30 kW. An NVIDIA GB200 NVL72 rack, the workhorse of 2025 and 2026 AI training clusters, draws about 120 kW. NVIDIA's roadmap goes further: Vera Rubin systems ship in the second half of 2026 at similar rack power, and the company has said its 2027 Rubin Ultra design will reach roughly 600 kW per rack, with power delivered over 800-volt DC.

Everything else follows from that number. A building designed for 10 kW racks cannot simply be refilled with 120 kW racks: the power feeds, the switchgear, the floor loading and, above all, the cooling all have to change.

Training and inference also pull in different directions. Training clusters tolerate distance from users and go where power is cheap and available, which is why the largest campuses are rising in West Texas and other frontier markets. Inference, serving answers to users, is moving closer to population centers as it grows.

Power: how much, and how to get it

The International Energy Agency estimates that data centers used about 415 TWh in 2024, roughly 1.5% of the world's electricity, and projects about 945 TWh by 2030. The United States accounted for 45% of the 2024 total. In April 2026 the IEA reported that data center demand grew 17% in 2025, with AI-focused sites growing faster.

US forecasts vary with their assumptions. Lawrence Berkeley National Laboratory, in its report for the Department of Energy, put US data centers at 176 TWh in 2023, or 4.4% of the country's electricity, rising to between 6.7% and 12% by 2028. EPRI's 2026 study projects between 9% and 17% by 2030.

The queue is the constraint

Money and land are rarely what holds a project back now. The grid is. Berkeley Lab counted about 2,061 GW of generation and storage waiting in US interconnection queues at the end of 2025, with a median of about five years from request to operation. On the demand side, JLL reports average waits of more than four years for a grid connection in primary data center markets.

Texas shows the scale. ERCOT's large-load queue reached 474.7 GW in June 2026, 90% of it data centers, more than five times the grid's record peak of 91 GW set that July. Texas Senate Bill 6, signed in 2025, now requires loads of 75 MW or more to disclose backup generation and allows new large loads to be curtailed in emergencies, and in August 2026 the governor ordered an audit of the data center queue.

Bringing your own power

Operators are answering with power of their own. The Stargate campus in Abilene, Texas, connected to ERCOT but also built about 360 MW of on-site gas turbines, with plans for more. Hyperscalers have turned to nuclear: Microsoft's 20-year agreement to restart Three Mile Island Unit 1 (about 835 MW), Meta's 20-year contract for the full 1.1 GW output of the Clinton plant in Illinois, and Amazon's 1.92 GW arrangement with Talen Energy at Susquehanna. The IEA counts 45 GW of conditional agreements between data centers and small modular reactor developers.

Cooling past the limit of air

Air cooling struggles somewhere around 100 to 120 kW per rack. Liquid carries far more heat per unit volume, so AI racks are cooled by liquid that reaches the chips directly. Cooling has historically consumed a large share of a data center's electricity, which makes the choice of system a cost decision as much as an engineering one.

Rear-door heat exchangers

A liquid-cooled coil on the back of the rack. Industry estimates put the practical range at roughly 20 to 50 kW per rack, so they suit mixed halls rather than dense AI clusters.

Direct-to-chip liquid cooling

Cold plates on the GPUs and CPUs carry heat to a coolant loop. This is the standard for NVIDIA's GB200 and later systems; industry estimates run from about 50 kW to well over 200 kW per rack.

Single-phase immersion

Servers sit in tanks of non-conductive fluid that is pumped through a heat exchanger. Widely used in Bitcoin mining, and increasingly tested for AI.

Two-phase immersion

The fluid boils on the chip and condenses on a coil above the tank. It removes heat very efficiently but needs specialised fluids and sealed tanks.

Efficiency is usually measured as power usage effectiveness (PUE): total facility energy divided by IT energy. The Uptime Institute reports an industry average near 1.56 that has barely moved in six years, while liquid-cooled AI designs target figures much closer to 1.1. Water is the other side of the ledger: Berkeley Lab estimated that US data centers used about 17 billion gallons directly in 2023, a figure that could double or quadruple by 2028. Closed-loop liquid systems use far less water than evaporative air cooling. For an operator's view of the problem, read designing for the AI cooling crisis and Bitmain's liquid cooling for AI data centers.

Sites, costs and markets

JLL puts the 2026 global average construction cost at about $11.3 million per megawatt, up from $7.7 million in 2020, with premium markets above $15 million. Cushman & Wakefield reports costs per megawatt up 21% since late 2024. Long-lead equipment sets the pace: large transformers can take about three years to arrive. Liquid cooling adds roughly $1 million to $2 million per megawatt over an air-cooled build.

Demand is outrunning supply. CBRE's mid-2026 count of eight primary North American markets showed 7.5 GW under construction, more than 80% of it preleased, and vacancy of 1.4%. JLL, counting every North American market, put more than 66 GW under construction and vacancy near 1% for a third year, with 77% of new construction in frontier markets led by West Texas. Many tenants are now signing for capacity delivered in 2028.

Northern Virginia

About 4,500 MW of inventory, 0.2% vacancy and 2,420 MW under construction (CBRE, H1 2026).

Atlanta

About 1,800 MW of inventory and roughly 2,900 MW under construction, the most in North America (CBRE).

Dallas–Fort Worth

About 1,430 MW of inventory and more than 765 MW under construction (CBRE).

Phoenix

About 1,070 MW of inventory (CBRE).

Chicago

About 910 MW of inventory (CBRE).

Why Bitcoin miners are becoming AI landlords

A Bitcoin mine is, physically, most of a data center: an energized grid connection, substations, land, cooling and crews used to running thousands of machines around the clock. What it lacks is the redundancy and fit-out that AI tenants require, and that can be added far faster than a new grid connection can be won. JPMorgan analysts estimated in 2024 that getting power three to four years sooner is worth $105 million to $140 million per megawatt to a hyperscaler. The deals that followed show how much of the market agreed:

• Core Scientific contracted about 590 MW across six sites to CoreWeave, worth about $10.2 billion in projected revenue over 12-year terms. CoreWeave's bid to buy the company outright was rejected by shareholders and ended in October 2025. Read how Core Scientific's operating system began.
• IREN signed a $9.7 billion, five-year contract with Microsoft for NVIDIA GB300 capacity at Childress, Texas, and delivered the first 50 MW phase in August 2026.
• Hut 8 leased 245 MW at River Bend, Louisiana, to Fluidstack for Anthropic, with Google backstopping the lease: about $7 billion over 15 years, and up to $17.7 billion if every option is taken.
• Cipher Mining signed about $5.5 billion over 15 years with AWS and about $3 billion over 10 years with Fluidstack, backed by Google.
• TeraWulf signed Fluidstack deals at Lake Mariner, New York, and Abernathy, Texas, with Google backstops and a Google equity stake.
• Galaxy contracted 526 MW of critical IT load at its Helios campus in West Texas to CoreWeave and delivered the first 133 MW in July 2026.
• HIVE runs GPUs alongside Bitcoin mining and is building a large AI site near Toronto. Read HIVE's dual-engine strategy.

Not every miner is converting. Most still mine, and many split sites between Bitcoin and AI. But the direction is clear enough that this conversion is the founding idea of Intelligence Disrupt: the people who own the power and the people who need it, on one floor.

Who is renting the capacity

Beyond the hyperscalers, a group of GPU clouds, often called neoclouds, rents capacity and resells compute. The largest, CoreWeave, reported second-quarter 2026 revenue of about $2.6 billion, a backlog near $104 billion, 1.5 GW of active power and about 3.7 GW contracted. Fluidstack appears as the tenant in several of the miner deals above, with Google standing behind the lease payments, a structure that lets a younger company sign long leases on a giant's credit.

Colocation rents reflect the shortage. JLL's figures put premium-market rents at roughly $140 to $160 per kW per month. For operators deciding between owning, leasing and renting GPU capacity, the trade-off is speed against control, and in 2026 speed is usually winning.

Dallas–Fort Worth

Dallas–Fort Worth is one of the largest and fastest-growing data center markets in the country. CBRE counted more than 1,400 MW of inventory in mid-2026, more than 765 MW under construction with 95% preleased, vacancy of 2.8% and another 3.7 GW planned. Oncor, the area's utility, reported 737 active large-load requests at the end of June 2026, about 282 GW of them from data centers, and plans more than $7 billion of transmission work along the southern DFW and I-35 corridor through 2034.

The draw is cost and speed: land that sells for a fraction of Northern Virginia prices, no state income tax, sales-tax exemptions on data center equipment and local property-tax abatements. The constraint, here as everywhere, is power. That is the subject of Intelligence Disrupt 2027, March 22–24 at the Irving Convention Center at Las Colinas, in the middle of the DFW market. See passes or exhibit.

Questions

How much power does an AI data center use?

The IEA puts a typical AI-focused data center at the electricity use of about 100,000 households, and the largest ones under construction at roughly 20 times that. A single NVIDIA GB200 NVL72 rack draws about 120 kW.

How much electricity do US data centers use?

About 176 TWh in 2023, or 4.4% of US electricity, according to Lawrence Berkeley National Laboratory, which projects 6.7% to 12% by 2028. EPRI projects 9% to 17% by 2030.

What is liquid cooling in a data center?

Coolant is brought to cold plates on the chips (direct-to-chip) or servers are submerged in a non-conductive fluid (immersion). Liquid carries far more heat than air, which is why racks above roughly 100 kW are liquid cooled.

How much does it cost to build an AI data center?

JLL puts the 2026 global average at about $11.3 million per megawatt, up from $7.7 million in 2020, with premium markets above $15 million. Liquid cooling adds roughly $1 million to $2 million per megawatt.

How long does it take to get grid power for a data center?

In primary US markets, more than four years on average, according to JLL. Newer markets can deliver power 12 to 24 months sooner, which is a large part of why West Texas and other frontier markets are growing.

Why are Bitcoin miners becoming AI data centers?

They already hold energized, interconnected sites. Hyperscalers and GPU clouds are paying for that head start: IREN, Hut 8, Cipher, TeraWulf, Galaxy and others have signed multi-billion-dollar AI hosting and lease contracts since 2024.

What is PUE?

Power usage effectiveness: total facility energy divided by the energy used by IT equipment. The industry average has held near 1.56 for several years, per the Uptime Institute. Liquid-cooled designs can run well below that.

Why is Dallas–Fort Worth a top data center market?

CBRE counts more than 1,400 MW of DFW inventory with 765 MW under construction, 95% preleased, at 2.8% vacancy in mid-2026. Land costs a fraction of Northern Virginia's, Texas has no state income tax, and ERCOT offers a large, if crowded, grid.

Sources

1. IEA, Energy and AI: executive summary · 2025
2. IEA, Data centre electricity use surged in 2025 · 2026-04-16
3. US DOE on the LBNL data center energy report · 2024-12-20
4. EPRI, Powering Intelligence 2026 · 2026-02-26
5. Berkeley Lab, Queued Up 2026 · 2026-06
6. NVIDIA, Blackwell platform and liquid cooling · 2025-04-22
7. Uptime Institute, Global Data Center Survey 2025 · 2025-07-29
8. POWER Magazine, Texas data center interconnection queue audit · 2026-08-04
9. ERCOT, Senate testimony on the grid · 2026-07-29
10. CBRE, North American data center trends H1 2026 · 2026-08-27
11. JLL, North America data center report, midyear 2026 · 2026-08-11
12. Datacentres.com on JLL construction costs · 2026-05-12
13. Dallas Innovates on CBRE DFW data · 2026-09-03
14. Oncor, second-quarter 2026 results · 2026-08-06
15. Core Scientific and CoreWeave partnership · 2025-02-26
16. Core Scientific ends CoreWeave merger agreement · 2025-10-30
17. The Block, IREN delivers first Microsoft AI cloud capacity · 2026-08-17
18. DCD, Hut 8 and Fluidstack 245 MW deal · 2025-12
19. The Energy Mag, TeraWulf and Fluidstack in Texas · 2025-10-28
20. The Block, Galaxy delivers Helios phase I to CoreWeave · 2026-07-07
21. CoreWeave, second-quarter 2026 results · 2026-08-11
22. DCD, Meta and Constellation Clinton PPA · 2025-06-03
23. Talen Energy and Amazon, Form 8-K · 2025-06-11