Breaker Sizing and Inrush Current for AI Servers

The complaint we hear most often after an AI server install is some variant of "the breaker keeps tripping." Roughly half the time the rack is fine on average — the GPUs run, the BMC stays up, the SNMP graphs look reasonable — and yet at unpredictable moments the panel flips and the room goes dark. Less often the trip is the residual-current device (RCD), which is even more frustrating because the load looks balanced and the equipment looks healthy.

In our experience neither symptom is usually an overload. About half the time it is the wrong trip curve on the MCB — typically a B-curve breaker installed by an electrician who assumed AI compute behaves like office IT. About 30% of the time it is the wrong RCD type, where a Type AC or Type A device is being asked to discriminate DC-component leakage from a stack of PFC-corrected server PSUs. The remaining 20% is a real overload, almost always at the moment several PSUs synchronously cold-start and the inrush envelope briefly exceeds anything the average draw would predict.

This article is the electrical-side companion to W04 (PSU sizing) and P02 (PDU types). It covers what trip curves actually mean under IEC 60898, the real inrush behaviour of the PSUs we ship, when the 80% continuous-load derating bites, and which RCD type to specify for a rack of switch-mode AI server PSUs. Worked examples for a 4-GPU 5090 build and an 8-GPU build at the end.

What an MCB actually measures — two trip mechanisms in one device

Every IEC 60898 miniature circuit breaker has two protection mechanisms wrapped in the same housing.

The thermal element is a bimetallic strip that heats up with current over many seconds. It models the I²t profile that determines whether a cable is overheating. Under IEC 60898, the conventional thermal points are: at 1.13 × the breaker's rated current (In) the device must not trip within the conventional time (typically > 1 hour for ratings ≤ 63 A), and at 1.45 × In it must trip within the conventional time. This is the part that protects the wire from a sustained overload.

The magnetic element is an electromagnetic coil that snaps the contacts open within a fraction of a half-cycle when the current crosses a fixed threshold. This is the part that protects against short circuits — and the part that decides whether your AI server's inrush is "fine, brief surge" or "trip, see you tomorrow."

The trip curve letter (B, C, D) tells you the magnetic threshold band. The thermal characteristic is the same across curves at the same In; only the magnetic point differs.

For a 16 A breaker the magnetic thresholds are roughly: B16 trips above 48–80 A; C16 trips above 80–160 A; D16 trips above 160–320 A. The exact point inside each band depends on the manufacturer and the temperature.

Most EU residential and small-commercial buildings ship with B-curve breakers as the default. That choice is correct for office sockets, lighting, and resistive loads. It is the wrong default for an AI server circuit for reasons we are about to walk through.

Real PSU inrush behaviour — the part data sheets understate

When a server PSU is plugged in and switched on cold, the first job of the input stage is to charge the bulk capacitors on the DC side. These are typically 470 µF to 2200 µF rated 400 V or 450 V, sitting after the bridge rectifier. From a flat start they look like a near-short to the AC line for a few hundred microseconds, until the voltage across the capacitor approaches the peak line voltage.

The mechanism in most decent PSUs is an NTC (negative temperature coefficient) thermistor in series with the input. Cold, it presents 5–20 Ω; the inrush current self-heats it within milliseconds, the resistance drops to a few percent of cold value, and the thermistor goes effectively transparent for steady-state operation. The downside is that an NTC needs to cool back down after a power cycle — a PSU that just ran and was switched off for ten seconds may inrush almost as if there were no limiter at all, because the thermistor is still hot and low-resistance.

Server-grade CRPS modules often add an active inrush limiter: a relay or MOSFET that bypasses a current-limiting resistor once the bulk caps are charged. This combines the cold-start protection of an NTC with the cool-down independence of a hard switch. CRPS modules in the 1.6–3 kW range are typically better-behaved on inrush than ATX units of the same wattage, partly because the active circuit is standard, partly because the source impedance assumption (datacenter 230 V) is tighter.

Two consequences for AI rack design:

1. The inrush peak in the data sheet is at nominal voltage and "typical" source impedance. At 245 V (high end of EU tolerance) on a short, low-impedance feed from a nearby panel, the peak can be 1.5–2× the typed number. We have measured 180 A peaks on a 1500 W ATX unit that was specced at 100 A.
2. Multiple PSUs starting simultaneously add their inrush peaks. The line cycle is the same for all of them; if four PSUs are plugged into the same circuit and all close their input switches at the same instant, the breaker sees roughly 4× the per-PSU peak. A 4× 5090 server with two 2 kW ATX PSUs (~150 A peak each on a worst case) starts with ~300 A peak briefly. A B16 breaker (magnetic threshold 48–80 A) trips. A C16 breaker (80–160 A) trips on the worst case. A C32 breaker (160–320 A) holds.

This is the mechanism behind the most common "my new AI server keeps tripping the breaker" complaint. The sustained load is fine. The inrush envelope on cold start is what trips.

Continuous-load derating — the 80% rule that catches AI workloads

EU practice (and IEC 60364 in spirit) is to derate the continuous-current rating of a breaker to roughly 80% of nameplate for loads that draw sustained current for more than three hours. North America codifies this in NEC 210.20(A); Europe applies it as best practice rather than a single hard rule, but the underlying physics — bimetal heating in the breaker, conductor heating in the cable — is identical.

AI inference and training are explicitly continuous loads. A vLLM endpoint serving a 70B model under steady request load draws 80–95% of nameplate GPU TDP for hours. A training run is even more uniform. The 80% derate applies in full — there is no "intermittent burst load" carve-out for AI workloads, because the load is exactly the kind of sustained draw the derate exists to protect against.

A 4× 5090 server at 2.4 kW sustained sits at 81% of a B/C 16 A breaker's continuous rating, which is precisely on the line. With a B-curve breaker the line is fine on average and trips on inrush. With a C-curve breaker the line is fine on average and trips occasionally on coincident transient peaks across the four GPUs. With a C 20 A or C 25 A breaker the line has real headroom and trips never. This is why our default recommendation for a 4-GPU 5090 build is a dedicated C 20 A or C 25 A circuit, not the 16 A circuit that the math "fits" on.

Trip-curve selection for the K-AI lineup

Two notes on this table:

• D-curve is rarely the right answer. D 10–20× In is for transformer-energising and motor-starting inrush that can ring for hundreds of milliseconds. PSU inrush is sub-millisecond to a few-millisecond event. A D-curve breaker tolerates the inrush trivially but loses sensitivity for short-circuit protection at the load end of a long cable run. Stick with C unless you have a specific reason.
• Going up one breaker size is almost always cheaper than going up one curve. A C 20 A on a properly-sized cable is a €15 part swap from C 16 A. Going from B to C on the same rating is also a part swap, but argues with the original electrician's design choice. Either approach works; we usually do both — bump to C-curve and one rating size up — to leave room for the second PSU you will eventually add.

Why vendor inrush data sheets understate the real number

Three reasons the data-sheet inrush figure is the floor, not the ceiling:

1. Vendor measurements are at 230 VAC nominal. EU mains can sit at 245 V or higher in some regions and during low-load hours. Inrush peak scales linearly with line voltage above the partially-charged capacitor's voltage. A 100 A inrush at 230 V is 107 A at 245 V — small effect, but it stacks with the others.
2. Source impedance assumed in the data sheet is a "typical" lab source, often a few hundred milliohms. A real installation with the panel in the same room as the rack and a short feeder run can deliver source impedance below 100 mΩ. Lower source impedance means less voltage sag during inrush, which means a higher peak current. We have seen the same PSU draw 80 A peak on a long cable run and 140 A peak on a short one.
3. DC-bus precharge state matters on a quick power cycle. A PSU switched off for a few seconds and back on may see warm bulk caps with residual charge — usually that helps, less inrush. But a PSU switched off for a few seconds with a cold NTC that hasn't reset and partially-discharged caps can produce a peculiar second-strike inrush that exceeds the cold-start figure. This is rare, but it shows up in racks where someone is power-cycling via a switched PDU during debugging.

The practical answer is to assume the worst-case peak is roughly 2× the data-sheet figure when sizing the MCB curve. This is the reason the table above lands on C 20–25 A for 4-GPU builds and C 32 A or three-phase for 8-GPU builds, even though the steady-state math would fit on smaller breakers.

Differential protection: RCD type for AI server racks

The RCD (residual current device) protects against earth leakage that exceeds a safe threshold — typically 30 mA for personal-protection circuits. The mechanism is a current transformer that sums the live and neutral currents; in normal operation the sum is zero, and any imbalance (current going to earth somewhere) shows up as a residual that trips the device above its threshold.

Two ways an AI server rack confounds an RCD:

Steady-state Y-capacitor leakage. Every switch-mode PSU has Y capacitors in its EMC input filter, sitting between live-to-earth and neutral-to-earth. These shunt high-frequency switching noise to earth. They also pass a small 50 Hz current to earth as a side effect — typically 0.5–1 mA per PSU on consumer ATX, 1–3 mA per PSU on server CRPS units with heavier filtering. IEC/EN 60950-1 and its successor IEC 62368-1 cap this at 3.5 mA per IT device for safety reasons. A 4× PSU rack can therefore present 4–12 mA of steady-state earth leakage before any fault. On a 30 mA RCD this is "fine on paper" but eats nearly half the trip threshold.

DC-component leakage from active PFC. Modern server PSUs use active power-factor correction (PFC) — a boost converter on the input that pulls current proportional to the line voltage waveform, achieving PF > 0.95. The PFC stage produces a small but non-zero DC component in any earth leakage that does occur (under fault conditions, mostly). Type AC RCDs respond only to pure sinusoidal AC residual current — they will not detect DC-component faults. Type A RCDs detect AC and pulsating DC, which covers most single-phase electronic loads. Type B RCDs detect smooth DC residuals as well, which is what active PFC and any three-phase rectifier at the load can produce.

Type B RCDs are 3–5× the cost of Type A. For a single 4-GPU AI server in an existing office circuit, Type A is usually fine — the leakage is dominated by Y-capacitor AC current, the DC component is small, and the cost of a Type B retrofit on the whole panel is hard to justify. For a dedicated AI lab with multiple servers, Type B on the upstream feed is the right answer, with downstream 30 mA Type A RCDs at the socket level for personal protection on cleaning sockets and so on.

Selectivity: the upstream/downstream RCD trap

One nasty failure mode in labs that bolt extra RCDs onto an existing panel: a 30 mA RCD upstream of another 30 mA RCD does not provide any selective discrimination. Either device may trip first on a fault, and once one trips the other may follow as the leakage redistributes through the trip mechanism's own paths. The result is the entire panel goes dark on what should be a localised single-circuit fault.

The correct topology for an AI lab:

Correct RCD selectivity topology: 100 mA Type B selective upstream of AI circuits, 30 mA Type A only on auxiliary and office sockets.

The "S" or "selective" designation on an upstream RCD adds an intentional 40–200 ms time delay to the trip mechanism, ensuring downstream 30 mA devices have time to clear a fault first. Without the time delay, two RCDs in series at the same threshold race each other. A 100 mA Type B selective upstream, with 30 mA Type A downstream on auxiliary sockets only and no RCD on the dedicated server circuit (relying on the upstream 100 mA for fault clearing), is the topology that actually works for a multi-server AI rack.

Worked example 1 — 4-GPU 5090 K-AI server

Setup: a single K-AI 96 Turin with 4× RTX 5090, 2× 1500 W ATX PSUs (split delivery per W04), in a small lab on a freshly-installed dedicated circuit.

Sustained load: ~2.4 kW at 230 V = 10.4 A continuous. At 80% derate, the smallest breaker that legally fits this is 13 A; the next standard size up is 16 A.

Inrush peak (worst case): 2× 150 A = 300 A combined, lasting < 5 ms. A B 16 A breaker (magnetic 48–80 A) trips on this. A C 16 A breaker (magnetic 80–160 A) trips on the high end of the worst case. A C 20 A breaker (magnetic 100–200 A) holds.

RCD: single PSU pair, Y-cap leakage estimated at ~3–5 mA combined steady-state. A 30 mA Type A RCD has ~25 mA of headroom — fine. If the building has only Type AC RCD protection, replace the upstream device with Type A or Type B; do not stack a second 30 mA device.

Specification handed to the electrician:

• Dedicated 230 V single-phase circuit, C 20 A MCB (or B 20 A if the supplier does not stock C-curve in that rating; C is preferred).
• 30 mA Type A RCD at minimum, Type B if the building is being upgraded anyway.
• Schuko outlet (or local-country single-phase 16 A connector) at the rack. C19/C20 cable from the outlet to the PDU.
• No other loads on the same circuit.

This circuit will hold the 4-GPU 5090 server through cold starts, sustained training, and the occasional power-cycle with no nuisance trips.

Worked example 2 — 8-GPU 5090 K-AI server

Setup: a single K-AI 256 Turin Dual with 8× RTX 5090, 2+2 CRPS modules at 2 kW each, intended for sustained training in a small server room.

Sustained load: ~4.5 kW at 230 V = 19.6 A continuous. At 80% derate this needs at least a 25 A breaker; the next standard size is 32 A.

Inrush peak (worst case): 4× 80 A = 320 A combined for CRPS modules, < 5 ms. A C 32 A breaker (magnetic 160–320 A) is on the edge. A C 40 A breaker (200–400 A) holds with margin.

RCD: 4× CRPS with PFC, Y-cap leakage estimated at ~8–15 mA combined steady-state, plus a non-zero DC component. A 30 mA Type A RCD will eventually nuisance-trip on this, and may not detect DC fault current correctly. Type B 100 mA selective is the right specification.

Specification options:

• Option A — single-phase: C 32 A or C 40 A MCB on a dedicated 230 V circuit, 100 mA Type B selective RCD upstream. CEE 32 A blue (single-phase IEC 60309) outlet at the rack. This works but is at the upper edge of single-phase distribution; the cable run sees significant current.
• Option B — three-phase (preferred): 400 V three-phase 16 A CEE red outlet at the rack, with the four CRPS modules distributed across phases at the PDU (per P02/P03). Each phase carries ~6.5 A sustained, which is well inside any C 16 A MCB. RCD per leg or shared 100 mA Type B selective on the three-phase feed. This is the cleaner answer for any 8-GPU build that has three-phase available.

Worked example 3 — four-rack lab planning

Setup: planning a robotics + AI lab with four racks, each holding 2× 8-GPU K-AI servers, 2× 4-GPU servers, networking, and humanoid charging. Rough sustained per rack: 2× 4.5 + 2× 2.4 = 13.8 kW; with networking, switches, and chargers round to 15 kW per rack, 60 kW total.

Single-phase distribution is dead on arrival here — the sustained per-rack figure already requires three-phase. The plan:

• Building feed: 400 V three-phase, minimum 100 A per phase service. Existing buildings often need a utility upgrade for this; check the panel before signing the lease.
• Lab sub-panel: 400 V three-phase, fed from the main on 4× 63 A or 4× 100 A conductors per the cable run.
• Per rack: 400 V three-phase 32 A CEE red feed from the sub-panel (~22 kW theoretical, 17.6 kW continuous — fits the 15 kW per rack with 17% headroom). C 32 A three-pole MCB upstream of each rack feed.
• RCD strategy: 100 mA Type B selective per rack feed, 300 mA Type B selective on the sub-panel main as fire protection. Personal-protection 30 mA Type A RCDs only on auxiliary sockets in the lab (workstation desks, cleaning sockets), not on the server feeds.
• Phase balancing: per P03; phase-striped outlet-metered PDU per rack with documented outlet-to-phase mapping.
• Staggered startup: see below.

The staggered-startup trap

A specific scenario that bites labs the first time: a power outage trips the upstream breaker, mains returns, and every server in the rack tries to cold-start at the same instant. The total inrush is N × per-PSU peak across all servers in the rack. For a four-server rack this can briefly hit 1000–2000 A on the single feed.

The upstream MCB sees this and trips. The room goes dark again. The sysadmin power-cycles the panel. Same thing happens.

Two solutions, both of which we deploy by default on Kentino multi-server builds:

1. Switched PDU with staggered startup sequencing (per P02). The PDU's outlets are configured to power up in sequence with a 2–5 second delay between each outlet. Total time to bring up four servers is 8–20 seconds; the inrush events do not overlap.
2. PSU "delayed start" or "soft start" in BIOS / IPMI. Many server-grade platforms have a configurable AC-power-recovery delay that randomises or staggers the start time within a small window. Combined with the PDU stagger, this provides a second layer of decorrelation.

Without one of these, an 8-server rack on a single MCB will trip on every single power-recovery event. We have seen it; it is reliably reproducible; it has nothing to do with the average load.

---

This is part of the Kentino Wiki, a reference series on AI compute, robotics, and the systems that connect them. Comments and corrections welcome at info@kentino.com.