Most home-lab power advice on the internet starts from the wrong number. It takes the breaker's label — 20 amps, 120 volts, 2,400 watts — and works backward from there. The National Electrical Code (NEC) works forward from the load instead, and for loads that never stop, which is exactly what a server is, the code subtracts a fifth. The result is a hard, checkable ceiling on what a residential rack may draw: 1,920 watts continuous on a 20-amp circuit. This piece is the arithmetic, the code sections behind it, and the three decisions — load tally, receptacle count, voltage — that actually determine whether your lab trips at 2 a.m. This is educational content, not electrical design advice: local amendments apply, and licensed electricians exist for a reason.

What does the NEC actually say?

Two sections carry the whole rule. NEC 210.20(A), Overcurrent Protection: "Where a branch circuit supplies continuous loads or any combination of continuous and noncontinuous loads, the rating of the overcurrent device shall not be less than the noncontinuous load plus 125 percent of the continuous load." NEC 210.19(A)(1), General applies the same 125% factor to conductor sizing — so the wire and the breaker are both sized for the same inflated number, and the 80% figure is simply the inverse view: a breaker that must be rated at 125% of your continuous load can only be loaded to 80% of its rating, continuously. The definition that decides which column you are in sits in Article 100: a continuous load is one "where the maximum current is expected to continue for three hours or more." Space heaters, EV charging and fixed lighting are the canonical examples. So is anything in your rack that does not sleep.

Apply the formula. A 20 A / 120 V circuit has a nameplate ceiling of 20 × 120 = 2,400 W. Continuous ceiling: 16 A × 120 V = 1,920 W. The 15 A circuit next to it: 15 × 120 = 1,800 W nameplate, 1,440 W continuous. These four numbers — 2,400/1,920 and 1,800/1,440 — are quoted consistently across independent electrical-engineering references and match the code formula exactly, which is why this piece can state them as facts rather than estimates.

For mixed loads — the realistic case, where a rack runs 24/7 but you also plug a vacuum into the same room — the code gives the formula directly: required breaker rating ≥ noncontinuous load + (1.25 × continuous load). Run the example the other way: a 1,500 W space heater (12.5 A, and by definition continuous) plus a 200 W of noncontinuous gear needs a breaker rated at 12.5 × 1.25 + 1.7 ≈ 17.3 A — it fits a 20 A breaker with almost nothing left, which is why the heater owns the circuit and your lab should not share it.

Why does the code derate a breaker that "handles" 20 amps?

Because the breaker's amperage is not the limiting surface — heat is. A thermal-magnetic breaker carrying near its rated current continuously generates heat at its bimetal strip and, more importantly, at every termination point: breaker lugs, wire splices, receptacle contacts. Terminations age with sustained temperature, and a breaker run at 100% for hours can trip from accumulated heat even though the instantaneous draw is below its stamped rating — the nuisance-trip that everyone blames on "a bad breaker" is usually a load that was never continuous-rated for that circuit. The 80% rule exists so that a circuit carrying its continuous load for three hours plus never reaches the thermal region where connections degrade. That is also why the rule is not negotiable by swapping hardware: the conductor, the receptacles and the panel bus are all part of the same thermal system, and NEC 110.14(C) further requires that for circuits 100 A and under, ampacity be taken from the 60 °C column unless terminals are rated higher — so the "12 AWG is rated 30 A at 90 °C" argument fails twice.

How many things can you actually plug in?

The NEC also caps receptacle planning at 180 volt-amperes per receptacle yoke (NEC 220.14(I) for the commercial/general case). Dividing a 20 A circuit's 2,400 VA by 180 VA gives about 13 yokes for general, non-continuous use; for a continuous-load circuit the continuous ceiling of 1,920 W gives about 10. But the yoke count is a planning baseline, not the real limit — the real limit is your device tally. The version that matters for a home lab:

DeviceTypical drawNotes
Dual-socket rack server, idle→full load~200 W → 650–700 WThe variable that decides everything; check the PSU label and measure
4-bay NAS (spinning disks)~60–100 WStartup spin-up adds briefly; UPS transfer rides through
Switch (8–24 port)~15–60 WPoE injectors change this fast
Workstation~150–450 WGaming/ML workloads are bursty, not continuous
UPS (online, loaded)~5–15% of throughputConversion loss counts against the circuit too
20 A / 120 V continuous ceiling1,920 W — the sum of everything running 3 h+

Read the table against the ceiling and the shape of the constraint appears: one full-load server (700 W) plus NAS plus switch plus conversion losses lands near 900 W — comfortable. Two full-load servers plus storage runs past 1,600 W — inside the limit with little margin, and one spin-up event or one PoE switch later you are testing 1,920. Three full-load servers do not belong on one 120 V / 20 A branch at all. This is the arithmetic behind the folk wisdom that a rack "needs its own circuit": at continuous duty, it usually does (ours).

What hardware must match, beyond the breaker?

Four things, and mismatching any of them is the actual fire risk. Wire gauge: 12 AWG copper for 20 A circuits, 14 AWG for 15 A — a 20 A breaker on 14 AWG is an upgrade of the breaker onto wire it was never protecting, not an upgrade of the circuit. Receptacle: a true 20 A circuit uses a NEMA 5-20R (T-slot) receptacle for 125 V; the slot shape is the code's own enforcement mechanism, preventing a 20 A plug from landing on a 15 A circuit. Terminals: conductors must land on terminals rated for the size and temperature as noted above. Extension cords and power strips: they are not branch-circuit wiring, they add termination points, and a server rack drawing 1,000 W continuously through a consumer power strip is the single most common home-lab fire-adjacent configuration; rack PDUs with listed branch-circuit cords exist for this. If you are putting a rack on a circuit, the circuit is infrastructure, not furniture.

How do you double the watts without new copper?

Change the volts. Watts are volts × amps, and the code caps amps — so the same 12 AWG copper on a 240 V / 20 A two-pole circuit (NEMA L6-20) is 20 × 240 = 4,800 W nameplate and 3,840 W continuous. Virtually every rack server PSU is auto-ranging 100–240 V and ships more efficiently at 240 V; a licensed electrician can often pull the same run to a different breaker and receptacle for a fraction of the cost of a service upgrade. This is the standard homelab escalation path, and it stacks cleanly with the numbers in this piece: 3,840 W continuous is two to four full-load servers with headroom. The 120 V / 15 A starter lab (1,440 W continuous) remains a perfectly good origin — the point is to know which ceiling you are standing under before you add the third node, not after the 2 a.m. trip.

One more number belongs in your planning file, because it connects the circuit to the battery: your UPS's output in watts, not VA. A 1,500 VA UPS is commonly 900–1,000 W of actual capacity, and if the wall behind it is only good for 1,920 W continuous, sizing the UPS above the circuit protects nothing — the breaker opens first. We cover that arithmetic — VA vs watts, power factor, runtime curves — in the companion piece below.

What are the caveats?

Jurisdiction and vintage. The NEC is revised on a three-year cycle and adopted with local amendments; the sections quoted here (210.19(A)(1), 210.20(A), 220.14(I), 110.14(C), Article 100) are stable across recent cycles, but your AHJ's adopted edition and local rules govern. Definition nuance. "Expected to continue for three hours or more" is a design expectation, not a meter reading — a lab that runs 24/7 is continuous by any reading, but a bursty workstation is not, and the mixed-load formula exists precisely for rooms with both. Measured vs nameplate. Device wattages in the table are typical ranges, not data: a $15 plug-in power meter settles the question for your hardware, and PSU efficiency (80 Plus ratings) means DC load ≠ wall watts. And the 80% rule is about thermal safety, not about performance tuning. It costs you a fifth of your nameplate watts in exchange for terminations that do not age into failure — on the list of things a home lab can cheaply buy, that is the cheapest insurance there is. Nothing here is electrical design advice; when in doubt, the licensed electrician is the product.