AI Data Centers: What Changes When Traditional IT Infrastructure Meets AI
Author : NPOD Data Center Solutions | Published On : 05 Oct 2026
AI data centers look like normal server rooms from the door, but the numbers inside are far bigger. A client once asked me to add "just four GPU servers" to a room built for office file storage. The room could not carry them. Below I cover what fails first and what I'd check before spending money.
Power Use Jumps Up
A normal office rack draws 5 to 8 kW. One NVIDIA DGX H100 server with eight GPUs can pull up to 10.2 kW by itself, so four in one cabinet means about 40 kW. That is five to eight times the original plan.
I look at the electrical feed before anything else. A single-phase 32 A circuit at 230 V gives roughly 7 kW, which is less than one of these machines needs. Three-phase wiring is the usual fix. Battery backup (UPS) also needs a bigger size because training jobs make power use swing within seconds. Most AI data centers get designed around the power supply before anything else.
Cooling Hits a Wall Near 20 to 30 kW
Fans can cool a cabinet up to about 20 to 30 kW. Past that point they cannot push enough air and hot spots form. Each H100 chip is rated at 700 W, and every watt turns into heat that has to leave the cabinet.
Liquid cooling comes in three forms. A rear door unit bolts onto the back of an existing cabinet, so I'd start with it. Cold plates run liquid straight to the chips and suit new builds. Dunking the whole server in special fluid works well but changes how staff repair hardware, so I'd save it for a planned project. New AI data centers often pick liquid at the design stage.
Why AI Data Centers Need Stronger Floors and Faster Cables
One DGX H100 weighs about 130 kg, so four of them top 500 kg before cables. Older buildings hit floor limits fast. Ask for the floor rating and measure the doors for the trolley.
Speed matters just as much. Each server has eight 400 Gb/s links for chip-to-chip traffic. Connect one to an old 10 Gb/s switch and the GPUs sit idle while they wait for data. Copper cables at 400 Gb/s reach only 2 to 3 m, so I plan fiber for longer runs.
Training and Running a Model Need Different Sites
Training a large model needs a huge hall, and most firms rent one from a cloud or hosting provider. I wouldn't build it in-house unless AI is your main product. Running a finished model is lighter: a small one that reads camera feeds on a factory line or scans hospital images often draws 1 to 2 kW. A compact cabinet on site handles that, and nothing travels to a far-off cloud. Small AI data centers like this make a good starting project.
Micro data center makers such as NPod pitch AI data centers at this size. Its largest standard model gives 94U of space and 12 kW for IT gear, which is about one DGX-class server plus a little extra. Custom builds go up to 90 kW of cooling, and the company lists liquid cooling products of all three kinds. For work near the factory floor that looks like a fit. For training I'd ask for power-per-rack figures in writing before buying.
Questions to Ask Before You Build AI Data Centers
- What is the real kW per server on the vendor datasheet?
- How many servers go in one rack? Multiply, then add 20% for growth.
- Does the quote state cooling capacity per rack?
- Can the floor, the doors and the network handle the load? These get forgotten but can stall a project.
Conclusion
My advice is simple: find out your kW per rack. Cooling, wiring and floor choices all follow from that number. For a few small AI jobs I'd upgrade the room or buy a compact unit. For GPU training I'd call a hosting provider, since AI data centers built for that work are a different project from a server room upgrade.
FAQs
How much power does a rack in AI data centers use?
A rack of eight-GPU servers often draws 30 to 40 kW. The newest full-rack systems pass 100 kW. A normal office rack uses 5 to 8 kW.
Can I run GPU servers in my old server room?
A few low-power cards usually fit. A full eight-GPU server needs stronger wiring and better cooling in most rooms I've seen, which is why full AI data centers get built from scratch.
Do AI data centers need liquid cooling?
Not always. Below roughly 20 kW per rack good air cooling can work. Above 30 kW I'd go with liquid.
