Start Here · 2-minute read
What's an AI data center? (Plain English.)
No jargon. No slides. No assumed knowledge. Just the shape of the thing, why it matters, and where to go next depending on why you are here.
⏱️ 30-second definition
A data center is a building that keeps thousands of computers running 24/7. An AI data center is the same idea but scaled up ~10×: each rack uses 100+ kW instead of 10, needs liquid cooling instead of air, talks over InfiniBand instead of Ethernet, and the whole campus can consume the electricity of a medium city.
A data center has 4 physical parts. That's it.
⚡
Power
Electricity from the grid (or on-site turbines), backed up by batteries + diesel generators.
❄️
Cooling
Takes the heat OUT. Old DCs used fans + air. AI DCs pump liquid directly through the chips.
🌐
Networking
Fiber optic cables connecting servers to each other and to the internet. 400–800 Gbps per port.
💻
Compute (the IT)
The actual servers + GPUs + storage. Everything else exists to keep these boxes alive.
Why AI changes everything
For 25 years, server rooms drew ~5–10 kW per rack, used air cooling, and ran on standard networks. Boring. Predictable.
Then GPU training arrived. A single NVIDIA GB200 NVL72 rack draws ~120 kW. That's more than the entire corporate server room it replaced. You can't cool it with air. The existing power feed can't handle it. Standard Ethernet is too slow to stitch 1,000+ GPUs into one training job.
Everything downstream — building design, utility contracts, site selection, staffing, financing — had to be redesigned around that single fact. That's what "AI data center" really means: a whole industry re-engineered for 10× power density.
Why should I care?
- 💰 Money: The four US hyperscalers will spend $320B+ on capex in 2025 alone, mostly AI data centers. Cumulative 2024-2027: over $1 trillion.
- ⚡ Power: A single campus (Meta's Hyperion) will add 15% to an entire state's electricity demand. This is reshaping US utilities.
- 🌍 Emissions: Firm clean power can't scale fast enough. Gas, nuclear restarts, geothermal deals are all being done right now.
- 🏆 Competition: Whoever builds the most training compute first determines which AI lab leads. That's a $T+ geopolitical stake.
- 📈 Jobs: US DC technician roles start at $80k in hot markets. Engineer roles $130–220k. This field is hiring.
OK, where should I read next?
Pick the path that fits why you're here. Each path is 20–45 minutes of reading.
👨💻
I'm a software engineer curious about infra
Your advantage is the software stack. These lessons connect what you already know (K8s, CUDA, NCCL) to the physical reality underneath.
📊
I'm an investor / analyst
The money math: capex dominated by silicon, opex dominated by power. Case studies show real deal structures (Blue Owl, Crusoe, Oracle lease).
🏗️
I'm an engineer or facilities pro
You know electrical and HVAC. We map your domain to AI-era density, load profiles, and the tier decisions that matter.
🏛️
I'm a policymaker / journalist / citizen
The questions your readers/constituents have: who pays, what's the emissions story, why 10 new gas plants for one data center?
Need a vocabulary shortcut?
The glossary (205+ terms) and the FAQ are there whenever you hit a term you don't know.