AI is pushing electricity demand up worldwide. Learn the units, work out rack power, PUE, cooling load, yearly energy and cost with simple formulas.
Traditional data centers ran many modest servers. AI data centers pack in dense racks of GPUs that run flat out for hours. Reports on the sector commonly put a conventional rack at about 5 to 10 kW, against roughly 40 to 100 kW for racks built for AI. The International Energy Agency has also reported that global electricity demand grew faster than overall energy demand in 2025, with data centers among the fastest-growing contributors. That makes the basic electrical maths behind these sites worth understanding.
| Unit | Measures | Meaning |
|---|---|---|
| kW / MW | Power (rate) | 1 MW = 1,000 kW |
| kWh / MWh / GWh | Energy (amount) | 1 GWh = 1,000 MWh = 1,000,000 kWh |
Power is how much a site draws at one moment. Energy is power multiplied by time, which is what a bill charges for. Our electrical power guide covers the difference in detail.
One AI rack drawing 100 kW around the clock:
Per day: 100 × 24 = 2,400 kWh
Per year: 2,400 × 365 = 876,000 kWh (876 MWh)
If a home uses about 10 kWh a day (an assumed figure), that single rack uses as much as 240 homes.
Not all power reaches the computers. Cooling, power conversion and lighting take a share. Power Usage Effectiveness (PUE) measures that overhead.
IT equipment draws 10 MW and the whole site draws 13 MW:
PUE = 13 / 10 = 1.3
That means 3 MW goes to overhead. A PUE of 1.0 would be perfect, and lower is better.
1,000 racks at 40 kW each, with a PUE of 1.3:
IT load = 1,000 × 40 kW = 40 MW
Facility load = 40 × 1.3 = 52 MW
Yearly energy = 52 MW × 8,760 h ≈ 455,520 MWh (about 456 GWh)
A 10 MW IT load at PUE 1.3, using an example electricity price of $0.08 per kWh:
Facility load = 10 × 1.3 = 13 MW
Energy = 13,000 kW × 8,760 h = 113,880,000 kWh
Cost = 113,880,000 × 0.08 ≈ $9.1 million per year
Real prices vary a lot by country and contract, so treat this as an illustration.
Almost all the electricity a server uses ends up as heat, so a 100 kW rack needs about 100 kW of heat removed. Cooling capacity is often quoted in tons of refrigeration, where 1 ton equals about 3.517 kW.
100 kW / 3.517 ≈ 28.4 tons of cooling per rack
That is why dense AI racks often move from air to liquid cooling.
High power means high current. Using the three-phase formula and assuming a power factor close to 1:
A 100 kW rack on a 415 V three-phase supply:
I = 100,000 / (1.732 × 415) ≈ 139 A
Cables and breakers must be sized for that current, which is why large sites feed racks with busbars and heavy cables. Our Wire Gauge calculator helps with voltage drop.
If grid power drops, batteries carry the load until generators start. They can also smooth out sudden changes in AI demand. Sizing uses the same basic maths as in our BESS guide.
Bridging a 10 MW load for 15 minutes (0.25 hour):
Energy = 10 MW × 0.25 h = 2.5 MWh usable
Allow extra for depth of discharge and efficiency, so you would size a battery bigger than 2.5 MWh.
Try each one before you open the answer.
40 × 24 = 960 kWh.
6.5 / 5 = 1.3.
2 × 1.5 = 3 MW.
A conventional data center rack typically draws about 5 to 10 kW, while racks built for AI workloads can need roughly 40 to 100 kW.
Power Usage Effectiveness is total facility power divided by the power used by the IT equipment. A PUE of 1.3 means 30% extra power goes to cooling, lighting and losses.
Nearly all the electrical power a server uses ends up as heat, so a 100 kW rack must have about 100 kW of heat removed.
Batteries bridge the gap during a power cut until generators start, and they can also absorb sudden load changes so the grid stays stable.
Check how much current a load draws and how much voltage you lose in the cable.
Open the Wire Gauge Calculator