The Electrical Requirements Nobody Puts in an AI Build Guide (20A Circuits, Watts at the Wall)

Bottom Line

• A continuous load may use 80% of the breaker. 15A at 120V = 1,440W. 20A = 1,920W. A 24/7 inference box is a continuous load.
• Single-GPU builds fit a 15A circuit. Even an RTX 5090 build, at ~800W at the wall.
• Dual RTX 5090 does not. ~1,400W DC is ~1,550W at the wall. That build needs a 20A circuit.
• Four RTX PRO 6000 Max-Q cards fill a 20A circuit. ~1,700–1,800W at the wall. Four Workstation Editions need 240V.
• A “1600W” PSU can be a 1300W PSU on US power. Read the low-line rating, not the box.
• A dedicated 20A circuit costs ~$250–900 in 2026. Less than the second GPU.

The Rule, Stated Precisely

The US National Electrical Code, article 210.20(A), says the branch-circuit overcurrent device must be rated at not less than 125% of the continuous load. The code defines a continuous load as one whose maximum current continues for three hours or more.

Standard breakers are listed for continuous operation at 80% of their rating. The breaker’s own terminations overheat at sustained full current. So 125% inverted is the “80% rule” that electricians use:

Two things follow that build guides get wrong.

First, the three-hour threshold is what makes this apply to you. A gaming PC at 800W for an hour is not a continuous load. A 24/7 local agent at 800W is. Same hardware, different code category.

Second, the limit is per circuit, not per outlet. The other outlets in the room usually share the breaker. A 1,300W workstation on a 15A circuit fits until someone plugs in a 400W space heater on the same wall.

This page is arithmetic on a published convention. Your local code, your panel and your wiring are the authority. Consult an electrician before you commit to a build above 1,400W.

From GPU Watts to Wall Watts

The number on the GPU box is DC power at the card. The breaker sees AC power at the wall. Three steps between them.

1. Add the platform. GPU total board power + CPU package power + ~100W for board, memory, NVMe and fans.

2. Divide by PSU efficiency. The 80 Plus program’s minimums at 50% load on 115V are 90% for Gold and 92% for Platinum. Efficiency drops at both ends of the load curve, so a supply running near its limit converts worse. Use 90% as a working figure unless your unit’s review says otherwise.

3. Compare to the continuous limit for your circuit.

Here is the ladder for the builds this site actually recommends. GPU figures are total board power from the vendor; CPU and platform are our estimates and are labelled as such.

The dual-5090 row is the one that surprises people. Two consumer cards, one consumer motherboard, and it is already past the 15A continuous limit. That build belongs on a 20A circuit, and it is the cheapest build on the site that does. Dual RTX 3090 vs one RTX 5090 compares the two-card builds; this table is why the 5090 pair also needs an electrician.

The four-card rows come from our 4x RTX PRO 6000 Max-Q workstation page. Four 300W Max-Q cards make a 384GB workstation possible. They do not make it fit an ordinary outlet.

The Spikes the Breaker Does Not See

The RTX 5090 spikes to about 901W for under a millisecond, roughly 57% above its 575W rating. That matters for the PSU, which is why our PSU guide puts a 5090 on a 1000W ATX 3.1 unit. It matters much less for the breaker. Thermal-magnetic breakers respond to sustained current, and a millisecond excursion is inside the PSU’s hold-up, not on the wall.

So size the PSU for transients and size the circuit for the average. They are different problems with different numbers.

The 1600W PSU That Is a 1300W PSU

This is the part that changes a four-GPU plan.

High-wattage ATX supplies are rated for their full output on 200–240V input. On low-line 100–120V input, which is every US wall outlet, several derate. The Seasonic PRIME TX-1600 is specified at 1,300W continuous on low-line power, as reported in the Tom’s Hardware and Hardware Busters reviews of the unit. The label says 1600W. On a US wall it is a 1300W supply.

Two consequences:

• A build that needs 1,600W DC needs a 240V circuit, or two supplies, or a unit whose manual states full output at 115V. Check the manual before you build. Do not trust the number on the box.
• This is the physical reason four 600W cards is not a home build. 2,400W of GPU alone is past what any single low-line supply delivers and past what a 20A 120V circuit permits.

For the builds under 1,300W DC, which is everything up to a dual 5090, the 1200W tier on our ladder covers it with headroom:

MSI MAG A1200PLS 1200W 80+ Platinum ATX 3.1 — the dual-GPU unit on our ladder, native 12V-2x6. Platinum’s 92% at half load is worth about 20W at the wall over Gold on a 1,000W build, which is small, but it runs cooler at the load fractions a 24/7 box sits at.

For a single RTX 5090:

Corsair RM1000e 1000W ATX 3.1 — NVIDIA’s stated minimum for the card, native 12V-2x6, and a build that lands at ~860W at the wall on a 15A circuit.

What a Dedicated 20A Circuit Involves

• Breaker: 20A in the panel.
• Wire: 12 AWG copper, minimum. 15A circuits are usually 14 AWG, which cannot carry 20A safely. NEC Table 210.21(B)(3) only permits a 20A receptacle on a circuit wired for 20A.
• Receptacle: NEMA 5-20R, the one with a T-shaped neutral slot. It accepts both 15A and 20A plugs.
• Dedicated means nothing else on the breaker. A 1,800W workstation and a 100W monitor on a 20A circuit is fine. A 1,800W workstation and a microwave is not.
• Cost: HomeGuide’s 2026 estimate is $250–900 installed, with typical jobs around $650. Distance from the panel and finished drywall push it up.

Put that against the hardware. A second RTX 5090 is $4,300–5,000 as of August 2026. The circuit that lets you run it is about 15% of the card. A four-card Max-Q workstation is a five-figure machine. An electrician’s visit is a rounding error on it and the one line item that decides whether it turns on.

We do not stock a link for a 20A power distribution unit or NEMA 5-20 power strip. Use one rated for the circuit, not a consumer 15A strip; the plug shape will tell you which one you have.

Above 20A: 240V

Past ~1,900W at the wall, a 120V circuit is the wrong tool. A 240V 20A circuit (NEMA 6-20) gives 3,840W continuous under the same 80% rule, and most high-wattage ATX supplies deliver their full rating on 240V. This is how four Workstation Edition cards get powered, and it is also how every rack does it. If your build is here, you have left the home-workstation category and the PSU, the circuit and the cooling are all an electrician’s conversation, not a parts list.

The Decision

See Also

• Can You Put Four RTX PRO 6000 Max-Q Cards in One Workstation? — the build that made this page necessary
• What PSU for a Local AI Rig — sizing the supply for transients, the other half of the power problem
• Local LLM Electricity Cost Break-Even — what the same watts cost per month at $0.18/kWh
• Undervolting Your GPU for a 24/7 Local Agent — the cheapest way to move a build down one row of the table
• Dual RTX 3090 vs RTX 5090 — the two-card builds, one of which needs a 20A circuit
• Motherboard and CPU for a Multi-GPU Rig — the platform watts in the table above

Sources

• NEC 210.20(A): overcurrent device rating not less than the noncontinuous load plus 125% of the continuous load; NEC Article 100 defines a continuous load as one expected to continue for three hours or more. Read via the IAEI Magazine and ECalPro explanations of the rule, which quote the article text; the 1,440W and 1,920W figures are arithmetic on 12A and 16A at 120V.
• NEC Table 210.21(B)(3) on 20A receptacles requiring a 20A branch circuit, and 12 AWG minimum conductor for 20A, from electrician guides to NEMA 5-20R installation
• 80 Plus efficiency minimums at 50% load, 115V: Gold 90%, Platinum 92% — CLEAResult 80 Plus program details, Tom’s Hardware explainer
• Seasonic PRIME TX-1600 low-line rating of 1,300W continuous — Tom’s Hardware and Hardware Busters reviews of the unit. Check the manual of any specific supply for its own low-line figure.
• RTX 5090 901W transient — Igor’s Lab and TechPowerUp measurements, as cited on our PSU page. GPU total board power figures from vendor specifications: RTX 5090 575W, RTX PRO 6000 Workstation Edition 600W, Max-Q 300W (NVIDIA)
• Dedicated circuit installation cost $250–900, typical ~$650 — HomeGuide 2026 estimate
• CPU and platform draws in the tables are our estimates and are marked as such; they vary by build
• Prices as of August 2026 from our hardware price reference: RTX 5090 $4,300–5,000 (Tom’s Hardware, TechPowerUp)