Single-Phase vs Three-Phase Power for AI Compute

The first conversation that derails an AI build is not about GPUs. It is about the wall. Someone orders an 8-GPU server, the installer arrives, looks at the available circuit, and the project stops for three weeks while an electrician retrofits the room. This article is the version of that conversation you want to have before you sign the purchase order.

The audience is buyers, lab managers, and integrators in Europe who need to decide whether their existing electrical install is enough, or whether a three-phase upgrade is on the bill of materials. We will assume nothing about your background in power systems — but we will be honest about where the math actually matters.

AC power basics, in the smallest dose that works

Alternating current at the wall has three numbers worth knowing.

Voltage. The pressure the grid pushes electrons with. Europe runs on a nominal 230 V phase-to-neutral. North America runs on 120 V for general outlets, 208 V or 240 V for higher-power equipment. Voltage is measured between two conductors — either between a "live" wire and neutral, or between two live wires of different phases.

Frequency. How many times per second the AC sine wave reverses direction. 50 Hz in Europe and most of Asia, 60 Hz in North America and a few outliers. For server PSUs this barely matters — virtually all modern switching power supplies accept either. But it does matter for motors, transformers, and any equipment with a synchronous frequency dependency.

Current. How many electrons per second flow. Measured in amperes. The breaker rating you see on the circuit (16 A, 32 A, 63 A) is the maximum continuous current the wire and breaker can carry safely.

Power, the thing that actually drives your GPUs, is voltage × current. At 230 V and 16 A you get 3,680 W of theoretical capacity. In practice continuous draw is derated to 80% of the breaker rating in most jurisdictions, so a 16 A circuit is good for about 2,950 W of sustained load. Keep that number in your head — it is the single most-quoted figure in this article.

Single-phase, split-phase, three-phase

The grid delivers power as three sine waves, each offset 120° from the others. What you get at the wall depends on how the building was wired from the transformer.

Single-phase. One live wire and one neutral. This is what every European home outlet looks like: 230 V phase-to-neutral, one breaker, one circuit. Simple, ubiquitous, limited in capacity. A "single-phase 16 A" circuit is the default living-room socket in the EU.

Split-phase. A North American oddity: 240 V transformer secondary, center-tapped to neutral, giving 120 V on each half and 240 V across both. This is what powers a US clothes dryer or EV charger. It does not exist in Europe. When an American forum says "240 V" they almost always mean split-phase 120/240 V, not the European single-phase 230 V. The wiring, breakers, and outlets are different. Mentioned here only so you do not confuse internet advice from US sources with EU reality.

Three-phase. Three live wires (L1, L2, L3) plus a shared neutral. In Europe the standard is 400 V between any two phases and 230 V from each phase to neutral. This means a three-phase outlet gives you both — high-voltage three-phase for big motors and PDUs, and three independent 230 V single-phase circuits if you split it correctly. This is the format every industrial install, every datacenter, and every serious workshop runs on.

Three-phase math, the version you need:

The headline: a 400 V 16 A three-phase circuit carries the same 11 kW that would otherwise need a 48 A single-phase 230 V feed. Same copper, three times the power, balanced load on the neutral.

EU electrical reality

These are typical, not guaranteed. The first thing you do before sizing a server install is look at the actual breaker panel and count circuits, not at what the building "should" have.

A few things that catch buyers out:

• A flat with a 400 V cooker socket is not a three-phase install. Some EU apartments have one three-phase outlet wired for an induction range or instantaneous water heater. That circuit cannot reasonably feed an inference server without rewiring and a separate breaker.
• "32 A" alone is meaningless. A 32 A single-phase circuit at 230 V is ~7.4 kW. A 32 A three-phase circuit at 400 V is ~22 kW. The voltage matters as much as the current.
• The breaker is not the only limit. The wire gauge, the panel feed from the street, and the building's main fuse all cap your real capacity. A 32 A circuit on a 25 A main fuse trips upstream before downstream.

When single-phase is enough

For a meaningful slice of buyers, single-phase 230 V is fine. The honest rules:

One 4-GPU AI server, consumer cards. A 4× RTX 5090 build draws 1.8–2.4 kW sustained under load. On a 230 V 16 A circuit (2.95 kW continuous limit) you have enough headroom for the server itself plus a workstation, a monitor, and the room's own lights. Most home offices and small labs run exactly this configuration.

One 4-GPU server, workstation cards. A 4× RTX Pro 6000 Blackwell at 600 W TDP each is the borderline case. The card-level draw is 2.4 kW; once you add the host platform, fans, and pumps the total lands at 2.8–3.0 kW. That is right at the continuous limit of a 16 A circuit. We deploy these on 32 A circuits whenever possible, or on a 3-phase rack PDU with one server per phase.

8× L4 or 8× L40 inference build. Low-power datacenter cards (L4 at 72 W, L40 at 300 W) keep the whole server inside the budget of a single 16 A circuit. An 8× L4 system runs around 1.0 kW total; an 8× L40 lands near 3.0 kW and benefits from a 32 A feed.

Two servers, 32 A circuit. If you have a 230 V 32 A single-phase circuit (7.4 kW continuous, ~5.9 kW derated) you can put two 4-GPU consumer-card servers on one feed. This is the realistic ceiling for single-phase deployments. Beyond two servers, the math stops working.

If your deployment fits in this table and is not going to grow, you do not need three-phase. Skip the rest of this article and call the electrician for a dedicated 16 A or 32 A run with a B-curve breaker.

When you NEED three-phase

Three-phase becomes mandatory the moment any of the following is true:

A single rack draws more than 7 kW continuous. A 32 A single-phase circuit at 230 V is hard-capped at 7.4 kW theoretical, ~5.9 kW derated. Anything above that needs either multiple single-phase feeds (clumsy, ugly, and you still hit the panel feed) or three-phase. For an 8-GPU 5090 server (4.5 kW) plus the host, networking, and a UPS, you are already at 5–6 kW for one server. Two servers in a rack is 10 kW. Three-phase is the only sane answer.

You need 32 A+ continuous draw on a single feed. Most EU residential and small-commercial installs cap single-phase branch circuits at 32 A. Beyond that the wire gauge, breaker selection, and panel design assume three-phase. If you ask an electrician for a "63 A single-phase" outlet for a server, they will look at you funny and then quote you for three-phase.

Multiple servers in one room. Once you cross two 8-GPU servers, you are in three-phase territory whether you like it or not. The wattage simply does not fit on single-phase EU circuits without ugly compromises.

You want one PDU per rack, not three. Running three separate single-phase circuits to a rack to power three servers is theoretically possible. In practice the cable management is awful, you cannot balance the load, and any half-decent rack PDU above 5 kW is three-phase only. You will end up with three-phase anyway, by way of the PDU.

Future-proofing for one more server. This is the case we see most often in practice. A lab orders a single 4-GPU build, the integrator is in already, the panel has room, and the marginal cost to pull a three-phase feed now is a fraction of the cost to redo the install in eighteen months when the second server arrives. If you have any plan to grow, three-phase from day one is the right call.

Three-phase advantages, concretely

Beyond raw capacity, three-phase has three real engineering advantages.

Less copper for the same power. A three-phase 400 V 32 A feed carries 22 kW on four conductors (L1, L2, L3, N) sized for 32 A each. The equivalent single-phase delivery — 22 kW at 230 V — would need 96 A on a single live-neutral pair, which is a much heavier cable and a much bigger breaker. For long cable runs in industrial buildings the copper savings are real money.

Balanced load, low neutral current. When the three phases are loaded equally, the neutral return current is approximately zero (the three sine waves cancel). This means the neutral conductor can be sized smaller, transformers run cooler, and your power factor stays clean. On a single-phase install every amp drawn flows back through the neutral.

Natural distribution across multiple devices. Three-phase PDUs hand you three independent 230 V single-phase legs (L1+N, L2+N, L3+N). Plug one server into each leg and you have automatic load balancing — assuming the servers are roughly equal in draw. We size racks with this in mind: three identical 4-GPU servers per three-phase 32 A PDU is the cleanest topology you can build.

Cleaner waveforms under heavy switching load. GPU PSUs are switch-mode and pull current in bursts at twice line frequency. Three-phase systems handle this load profile better because each phase only sees a third of the total switching activity. This shows up as lower harmonic distortion on the building feed and, in extreme cases, fewer nuisance trips of upstream RCDs.

The trade-off: three-phase is more expensive to install, requires a certified electrician who knows what they are doing, and adds complexity to your PDU and UPS sizing. None of these are show-stoppers above 7 kW per rack. Below that, single-phase is just simpler.

How a three-phase PDU actually distributes power

This is the part most buyers do not see until installation day, so it is worth spelling out.

A three-phase rack PDU takes one 400 V 5-pin input (L1, L2, L3, N, PE) and exposes a strip of standard C13 or C19 outlets. Internally it splits the input into three banks:

Three-phase PDU splits one 400 V input into three independent 230 V banks — one server per phase for balanced load.

A few practical implications:

• The servers themselves remain single-phase devices. Standard ATX or redundant server PSUs accept 100–240 V single-phase. There is no such thing as a "three-phase server PSU" in the AI build class. The three-phase is purely a building-distribution choice; the PDU does the splitting.
• Load balancing is on you. If you put two servers on L1 and nothing on L2 or L3, you defeat the point. Metered or switched PDUs show per-phase current; check them at install and rebalance if needed.
• Per-phase breaker matters. A 32 A three-phase PDU is usually internally fused per leg at 16 A or 20 A. You cannot pull 30 A through a single outlet bank just because the input feed is 32 A per phase. Read the PDU spec sheet.
• Dual-PSU servers on three-phase get interesting. If you have a server with two redundant PSUs and want true A/B power, you need two separate three-phase feeds (ideally from two breakers, sometimes from two utility sources). The PSU 1 goes to PDU A on (e.g.) L1, the PSU 2 goes to PDU B on L1. This is colo-grade design; small labs rarely need it.

400 V vs 230 V at the server side — the equipment story

A reasonable question: if three-phase delivers 400 V between phases, does my server somehow run on 400 V directly?

For AI servers in the Kentino class — Supermicro and Bone64c platforms with ATX or CRPS redundant PSUs — the answer is no. Server PSUs accept 100–240 V single-phase AC. They auto-range across that band. You feed them one of the 230 V legs out of a three-phase PDU and the PSU does not know or care that the building runs on three-phase upstream.

Some hyperscaler and large datacenter equipment uses 200–415 V or "high-voltage DC" (HVDC) inputs, where the PSU is fed phase-to-phase or directly from a rectified 380 V DC bus. These exist — they are more efficient at scale because each conversion step loses 1–3% — but they are not what an EU lab or small-commercial install will deploy. For the lineup Kentino actually builds (4-GPU and 8-GPU servers on PCIe-attached GPUs, standard CRPS PSUs), assume 230 V single-phase feed to the PSU regardless of upstream topology.

Honest EU install reality

Stripped of sales gloss, the picture for European buyers in 2026 is:

• One 4-GPU server: single-phase 16 A is fine. This is the bulk of small-lab and workstation deployments.
• One 8-GPU server: single-phase 32 A works, three-phase is cleaner. A 32 A circuit at the panel is not always available; if you have to pull new wire anyway, pull three-phase.
• Two 8-GPU servers or anything above ~7 kW continuous: three-phase, no exceptions. The single-phase math falls apart, the panel does not support it, and the PDUs you would want do not exist in single-phase variants.
• Any rack you intend to grow into: three-phase from day one. The marginal install cost is small, the rework cost later is large, and your future self will thank you.

There is a temptation among first-time buyers to push single-phase as far as it can go, in order to defer the three-phase install. This works for one server. It fails predictably at the second.

What to ask the electrician for

When the electrician comes to scope the install, hand them this list. (We are happy to do it for you on a Kentino-built deployment, but the words are useful even if not.)

1. Available capacity at the main panel. How many amps does the building feed allow? In an EU residential the main fuse is often 25 A or 32 A total; you cannot install a 32 A server circuit if the whole flat is on a 25 A main.
2. Existing three-phase availability. Some buildings have three-phase to the panel but only single-phase distributed. The cost to enable three-phase to a specific socket is much lower than running three-phase to the building from the street.
3. Dedicated circuit for the server(s). Not shared with lighting, HVAC, or office outlets. A dedicated B-curve breaker matched to PSU inrush behavior is what you want — C-curve trips less but allows more current before opening, A-curve trips too aggressively for switch-mode PSUs.
4. Outlet type and location. For three-phase, specify CEE 16 A or 32 A red (IEC 60309) outlets unless you have a specific PDU model in hand. For single-phase, Schuko CEE 7/4 (Germany, Austria, Netherlands) or the local standard (CZ/SK uses CEE 7/5, the French / Belgian variant) is fine.
5. RCD class. Switch-mode PSUs with PFC have small earth leakage currents. Type AC RCDs may nuisance-trip with multiple servers; Type A or Type B is the right choice. Spell this out — many electricians default to Type AC.
6. Cable run length. If the panel is far from the server room, voltage drop becomes a real concern at high current. The electrician needs to know the run distance to size the cable correctly.
7. Future capacity headroom. Ask for the breaker and wire to be sized one step above your current draw. The cost difference between 25 A and 32 A copper is small; the cost to redo it later is large.

If the electrician cannot answer questions 1–3 from looking at the panel, find a different electrician. This is fundamental work for any commercial install.

Safety call-out

We will say this once, plainly, because it matters.

Never DIY three-phase, split-phase, or anything beyond a domestic single-phase circuit. The voltages involved — 400 V phase-to-phase, with fault currents that can hit thousands of amps before the breaker opens — are lethal in milliseconds, not seconds. The clearance distances, conductor sizes, RCD selection, and earthing scheme are not intuitive and the failure modes do not give second chances.

In the EU, three-phase work is universally restricted to certified electricians (a "revize" in CZ/SK, a "qualification BR/B2V" in FR, a "Elektrofachkraft" in DE). The certification is not bureaucratic gatekeeping — it exists because the consequences of a wrong connection on a 32 A three-phase circuit include arc flash, electrocution, and building fires. Insurance will not pay out on unauthorized electrical work.

This applies even if you are technically competent with low-voltage electronics. A 12 V hobbyist who has soldered hundreds of boards still does not have the testing equipment, the regulatory knowledge, or the practiced safety behavior to do a building-grade three-phase install. Hire it out. The day-rate of a certified electrician is two orders of magnitude cheaper than a fire.

---

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.