The rack that trips a branch, or sits dark waiting for the right cord, was almost always a planning gap — not a hardware failure. Here is the math and the checklist to do it once.
The single biggest mistake in AI rack power planning is budgeting from "typical" or "idle" power. GPU nodes ramp hard during training or inference and can approach their PSU limit for long stretches. Plan every circuit, cord and PDU branch from the PSU input rating on the nameplate, then add margin.
| Node config (illustrative) | Per-input current | Cord family |
|---|---|---|
| Standard compute node, single 800W-class PSU @ 240V | ~3.3A | C13 / C14 |
| Compute node, dual 1.2kW-class PSUs @ 240V | ~5A per input | C13 / C14 |
| Accelerator node, 3kW-class PSUs @ 240V | ~12.5A per input | C19 / C20 |
Figures are for illustration — always use the actual PSU nameplate in your quote. The rule of thumb: under 10A per input stays on C13/C14; 10A and above moves to the C19/C20 family. For the shapes, ratings and plug standards, see our guide to C13, C19 & C20 data center power cables.
Amps = watts ÷ volts, so the same node draws half the current at 240V that it draws at 120V. That is why high-density AI racks are usually fed at 208V or 240V: it keeps loads inside C13/C19 ratings and PDU branch limits. If your facility only offers 120V, expect an accelerator rack to need far more circuits — and check every cord and PDU rating for the voltage you actually run.
Related reading: server rack cable management for ops teams.
Stock the rack from our C13 power cords, C19 power cords and C20 power cords, bundle with a rack starter combo and 1U shelves, or send your power plan — we will quote cords, plugs and certifications against your actual load sheet.
Read the PSU nameplate: if each power input is rated 10A or less at your line voltage, C13/C14 covers it; above 10A per input, plan on the C19/C20 family. Standard compute nodes usually stay on C13/C14 — accelerator nodes with high-wattage PSUs are where C19/C20 appears. The two inlet shapes are not interchangeable, so match the cord to the actual inlet.
For high-power AI racks, 208V or 240V is usually the practical choice: doubling the voltage halves the current for the same load, which keeps you inside cord and PDU branch ratings and reduces copper losses. A 10A-rated C13 circuit at 240V can carry roughly twice the power it can at 120V. Confirm what your facility and PDU actually supply before ordering cords.
Add up the nameplate current of every device on each PDU branch, then keep the branch below about 80% of its rating. Budget per-node power from PSU ratings rather than idle draw, and remember that GPU nodes can pull near their PSU limit during training or inference. Leave spare capacity for service and reconfiguration — racks get denser over time.