An AI data center differs from a traditional data center in what it is built for: it trains and runs AI models on dense racks of accelerators, mostly GPUs, rather than hosting general-purpose servers for websites, storage and business software. What separates it from a traditional data center is power density per rack, liquid cooling, a high-bandwidth network that lets thousands of chips work as one cluster, and electricity secured at the scale of hundreds of megawatts or gigawatts. The companies building them, including Meta, Microsoft, OpenAI and Crusoe, describe these differences in their own announcements, and the examples below come from those pages as of October 2026.

On this page

Accelerator racks instead of general-purpose servers

A conventional data hall is filled with CPU servers that each handle separate jobs. An AI data hall is filled with accelerator systems designed to work together on one model.

Nvidia’s GB200 NVL72 is a common building block. Nvidia’s product page describes a single liquid-cooled rack with 72 Blackwell GPUs and 36 Grace CPUs, 13.4 TB of HBM3E GPU memory and NVLink links that give 130 TB/s of GPU-to-GPU communication inside the rack. The rack behaves like one large accelerator rather than 72 separate ones.

Server-sized systems are the other format. Nvidia’s DGX B200 user guide lists eight B200 GPUs with 1,440 GB of total GPU memory in a 10U air-cooled chassis, with maximum system power of 14.3 kW for that one box. The /ai-chips catalog lists the accelerators that go into these racks, with figures from each vendor’s datasheet.

Power density per rack

Microsoft gave unusually specific numbers for its Fairwater AI data centers. In a November 12, 2025 post on the Atlanta site, the company said the design runs at about 140 kW per rack and 1,360 kW per row, with up to 72 Nvidia Blackwell GPUs in a rack. Loads at that level shape the electrical distribution, the cooling plant and the building itself; Microsoft built the Wisconsin and Atlanta sites on two floors to keep racks close together.

Density also explains why AI campuses are announced in megawatts or gigawatts rather than square feet. The /ai-data-centers tracker records each project’s capacity in the unit its owner used.

Liquid cooling

The Nvidia GB200 NVL72 rack is liquid-cooled by design, and the new AI campuses announced by Microsoft, Meta, Crusoe and OpenAI all describe closed-loop systems, several of them presented as a way to cut water use:

  • Microsoft Fairwater, Wisconsin: Microsoft said more than 90% of the site’s capacity uses a closed-loop liquid cooling system with “zero water waste,” served by what it called the second-largest water-cooled chiller plant in the world. For Atlanta, it said the initial fill of the closed loop uses about as much water as 20 homes consume in a year.
  • Meta Sturgeon County, Alberta: Meta said its 1 GW AI-optimized data center will use “a water efficient closed-loop, liquid-cooled system with dry cooling.”
  • Crusoe and Lancium, Childress, Texas: the 1.0 GW campus will use “closed-loop, non-evaporative liquid cooling.”
  • OpenAI PORTS-Pike, Ohio: OpenAI said the site will use “closed-loop, air-cooled cooling systems that recirculate water.”
  • Stargate Norway: OpenAI described “closed-loop, direct-to-chip liquid cooling.”

Networking: one cluster, not many servers

Training a frontier model splits the work across thousands of GPUs that exchange data constantly, so the network is part of the computer. AI data centers use two layers of interconnect.

Inside a rack, NVLink connects the GPUs. Microsoft said each Fairwater rack has 1.8 TB/s of GPU-to-GPU bandwidth over NVLink. Between racks, the operators use InfiniBand or high-speed Ethernet. For Wisconsin, Microsoft described 800 Gbps InfiniBand and Ethernet in a full fat-tree, non-blocking design, and a two-story building that shortens cable runs between racks. For Atlanta, it described an Ethernet-based back-end network with 800 Gbps GPU-to-GPU connectivity running the SONiC network operating system, plus a dedicated “AI WAN” optical network that links Fairwater sites to each other.

Nvidia’s September 9, 2026 announcement with Australian partners shows the same pattern: Sharon AI’s deployment of up to 68,000 Nvidia GPUs is connected with Nvidia Quantum InfiniBand and Spectrum-X Ethernet.

Power procurement

A conventional data center usually takes power from the grid like any large building. An AI campus often has to arrange new generation before it can be built, and the power deal can be as large a project as the data center.

Meta’s July 13, 2026 update on Richland Parish, Louisiana, which it is expanding to 5 GW of compute capacity, said its agreement with Entergy “will fund seven new natural gas-fueled generating plants, three grid-scale batteries, nuclear uprates, and other purchased power.” At the Department of Energy’s Paducah Site in Kentucky, NextEra Energy and Brookfield plan over 1.2 GW of compute capacity alongside up to 2 GW of natural gas and up to 2.6 GW of battery storage; NextEra said the project is subject to definitive documentation. OpenAI said its Project Camellia campus in Effingham County, Georgia will draw 3.2 GW of power delivered by Georgia Power in phases between 2028 and 2032.

Long-term contracts with existing plants are the other route. On September 30, 2026, Constellation and Amazon announced a 20-year agreement for 690 MW from the Calvert Cliffs nuclear plant in Maryland; our news report covers the terms.

Examples from companies’ own pages

Project Owner Location Capacity as announced Other detail from the announcement
PORTS-Pike OpenAI with SB Energy, Nvidia, US DOE Pike County, Ohio About 8 GW-IT; first 800 MW in 2028 Exclusively Nvidia AI compute
Richland Parish Meta Louisiana 5 GW of compute capacity Entergy agreement funds seven new gas-fueled plants and three batteries
Sturgeon County Meta Alberta, Canada 1 GW Closed-loop liquid cooling with dry cooling
Childress campus Crusoe and Lancium Texas 1.0 GW Purpose-built for a hyperscale technology company
Fairwater Microsoft Mount Pleasant, Wisconsin; Atlanta About 140 kW per rack (Atlanta) Nvidia GB200 and GB300 GPUs

The capacity column mixes units on purpose: OpenAI states IT load for PORTS-Pike, Meta states compute capacity, and Microsoft gives rack density rather than a site total. These figures measure different things and should not be added up or ranked against each other without that caveat.

Sources