An AI data center is a purpose-built facility that packs thousands of specialized chips into tightly networked racks to train or run AI models — and because those chips draw far more electricity and generate far more heat than the servers in an ordinary data center, the buildings themselves are put together differently, down to the power plants some companies now build just to feed them.
What it is
A conventional data center is a building full of general-purpose servers running websites, email, and business software; a single rack of that equipment typically draws somewhere in the range of 3 to 10 kilowatts. An AI data center is built around graphics processing units (GPUs) instead — chips originally designed for rendering video game graphics that turned out to be extremely good at the parallel math behind AI. A single high-end GPU, such as Nvidia’s full-specification Blackwell B200, can draw up to 1,200 watts on its own, up from 700 watts on the previous generation. Packed 72 to a rack in systems like Nvidia’s GB200 NVL72, that adds up to roughly 120 kilowatts per rack — and Nvidia’s next-generation Rubin platform is expected to push that toward 500 kilowatts. Companies including OpenAI (Stargate, in Abilene, Texas), xAI (Colossus 2, in Memphis), Microsoft (Fairwater, in Wisconsin), and Meta (Prometheus, in Ohio, plus a new campus in Sturgeon County, Alberta) are now building campuses sized not in racks but in gigawatts — enough power for a small city.
How it works
Where a normal data center runs thousands of unrelated jobs side by side, an AI data center behaves more like one giant machine: tens of thousands of GPUs wired together over dedicated high-speed networking so they can train a single model, or serve one pool of requests, in parallel. That density is also why cooling has to change. Ordinary air conditioning can only move so much air past a rack before it stops keeping up — engineers generally run out of headroom for air cooling at around one or two high-end GPU servers per rack. Beyond that, operators switch to liquid cooling: coolant piped through cold plates mounted directly on the hottest chips, carrying heat away far more efficiently than blown air and letting a single rack hold four, eight, or more GPU servers instead.
Why it matters
That much power and heat turns a data center into an energy project as much as a computing one. Meta’s new Alberta campus — its first in Canada, at roughly C$13 billion (US$9.17 billion) — is built for 1 gigawatt of capacity, scalable to 1.8 gigawatts, and it’s paired with a dedicated new natural-gas plant, Pembina Pipeline’s Greenlight Electricity Centre, built specifically to supply it. That pattern — a tech company financing or co-locating its own power generation instead of relying solely on the existing grid — is becoming common at gigawatt scale, and it’s why these campuses draw scrutiny: Greenpeace Canada, for one, has called for a moratorium on new mega data centers pending stronger environmental rules.
Water is the other flashpoint. Microsoft says its newest Fairwater campus recirculates the same coolant in a closed loop, using little enough fresh water annually to compare to a single restaurant. That’s real progress on-site, but analysts estimate on-site cooling accounts for only about 4% of the additional water the AI buildout will demand through 2050 — far larger shares come from generating the electricity these facilities consume (roughly 54%) and manufacturing the chips inside them. Closed-loop cooling fixes one part of the water problem, not the whole thing.
None of this is only about efficiency for its own sake: companies increasingly build their own AI data centers rather than renting GPU capacity from a cloud provider — see our explainer on neoclouds — because demand for AI compute has been outrunning supply, and owning the building, the chips, and increasingly the power plant is how a company guarantees it gets the capacity it needs.
Why it matters for Georgia
Georgia draws roughly 79% of its electricity from hydropower, a genuine low-carbon advantage on paper. But the country also imports a large share of its winter electricity — imports rose 109% year-over-year in one recent January — and its grid is far smaller and less diversified than those of regional competitors such as Turkey or Azerbaijan. Today’s largest facility in the country, Bitfury’s roughly 40-megawatt site in Tbilisi, is a small fraction of the gigawatt-scale campuses described above. Industry analysts have concluded that AI data centers becoming a central pillar of Georgia’s economy is unlikely in the near term, though a smaller, complementary facility — serving regional clients or supporting domestic “digital sovereignty” goals — remains plausible.
In the news
Meta’s Alberta campus is exactly this kind of project: see our report on the $13 billion Sturgeon County data center.
FAQ
Why can’t companies just add more air conditioning? Air can only carry away so much heat before the required airflow becomes impractical; beyond roughly one or two high-end GPU servers per rack, operators switch to liquid cooling instead.
How much electricity does one of these campuses use? The newest campuses are built for a gigawatt or more of capacity — comparable to the output of a large power plant, and enough to supply a small city.
Does closed-loop cooling solve AI’s water problem? Only partly. It cuts the water used on-site, but most of AI’s water footprint comes from generating electricity and manufacturing chips, not from cooling the building itself.
Why do companies build their own data centers instead of renting cloud capacity? Owning the facility, the chips, and sometimes the power supply guarantees a company gets the compute it needs, rather than competing for scarce rented capacity.
Sources: Reporting from Yahoo Finance, BNN Bloomberg, Tom’s Hardware, TechCrunch, and BTU AI.