Upgrading my house electrical system for AI workloads

I upgraded my house electrical system to support a local AI rack, meaning multiple multi-GPU machines running large models on my own hardware instead. The house went from 100A to 200A service, and I ran two dedicated circuits to the rack: a standard 120V for networking gear, and a 240V / 30A on a NEMA L6-30R twist-lock for the GPU workstations. That 30A outlet gives me 5,760W of continuous power to work with in the future as local models keep getting bigger and more power hungry.

A full panel on 100A service

My house had 100A service and the main panel was full. There were no free slots for new breakers at all, so the rack was blocked before I even got to the question of how many watts I needed.

I also needed an EV charger, which is a 40-50A continuous load on its own and needed a separate 240V / 60A breaker. So even if I had found panel space, 100A service was not going to carry an EV charger plus the rack.

Going to 200A

This is a bigger job than it sounds like and it meant a new external meter box, new 200A main panel with more breaker slots, and an electrician doing the panel swap, grounding and inspection. Most of the first day was trenching along the side of the house:

The new meter main is the gray box in the photo at the top of this post, shot with the shield still covering the socket. The utility came out and did a disconnect and reconnect in the span of 20 minutes, and during that window the electrician moved the drop coming in from the street off the old conduit and onto the new one. After that the meter went in and the trench got backfilled:

Inside, the new Siemens panel. The 200A main breaker is at the top, and the important part for me was the row of empty slots below the populated ones:

The two circuits I ran to the rack

120V / 20A

A NEMA 5-20R receptacle on its own breaker, for networking gear, smaller machines and accessories. The point is just that the rack does not share a circuit with whatever else is running in the house.

240V / 30A

This is the one that matters. I had a Leviton 2620 installed, a 30A 250V flush-mount locking receptacle. It is a NEMA L6-30R, which is what nearly every 240V rack power distribution unit (PDU) expects.

At 240V and 30A you get 7,200W on paper, or 5,760W continuous after the 80% derate. That covers two dual-GPU workstations with plenty of headroom left for spikes.

Picking a PDU

Almost all 240V rack PDUs take an L6-30P twist-lock plug, so once the L6-30R is on the wall the choice gets easy. Two things need to line up: the input plug has to be L6-30P to match the receptacle, and the outlets have to include C19 if you are feeding a high-wattage PSU, because that is what the other end of the cable needs.

As an example of something that fits, the Tripp Lite 5800K (PDUH30HV19) is a 1U basic PDU with a 12 ft L6-30P input cable and four C19 outlets, rated 208/240V at 24A, with two 20A breakers covering two outlets each.

That 24A works out to 5,760W at 240V, which is the same as the continuous rating of the circuit feeding it, so the PDU is not the thing that limits you. Note that all four outlets are C19, so anything low-power in the rack needs a C20 to C14 cable rather than an ordinary C13/C14 cord.

The twist-lock connection on the input is also nice because it will not slowly work its way loose the way a straight-blade plug can.

PSUs in this class are 240V only

The one I used is the Seasonic PRIME PX-2200, a 2,200W 80 PLUS Platinum fully modular unit. It is just one option at this wattage and nothing below depends on that particular choice. Most consumer PSUs accept 100-240V. This one does not. It only operates on 200-240V AC input and will not turn on at 120V at all, which is normal once you get into this power class.

Because of the wattage it also uses a C20 inlet (rated for 16A) instead of the C14 you see on lower-wattage PSUs. So the chain from the wall looks like this:

Wall (L6-30R) -> PDU (L6-30P input, C19 outlet) -> C20-to-C19 jumper -> PSU (C20 inlet)

The part that trips people up is the cable. You need a C20 to C19 jumper: the C20 end goes into the PDU’s C19 outlet, and the C19 end goes into the PSU’s C20 inlet. It is not the same as the common C13/C14 cable you use for servers and monitors.

I use a StarTech C20 to C19 cable, 14 AWG, 15A / 250V. Two feet is deliberately short because the PSU sits right next to the PDU and I did not want loose loops of cable in the rack. If you need slack, the Tripp Lite 6 ft version works fine. Look for 12 or 14 AWG rated at least 15A at 250V.

Why 240V and not just another 120V circuit

Start with what one machine actually draws. My dual RTX Pro 6000 workstation pulls over 1,200W from the wall under sustained inference load, and that is before transient spikes. A single RTX Pro 6000 is rated at 600W TDP, and GPUs happily spike above their TDP for short periods. Two of them plus a 7950X3D, DDR5, storage and fans puts one workstation at roughly 1,300-1,400W under full load.

A 120V / 15A circuit gives you 1,800W on paper, but code derates continuous loads to 80%, so the real number is about 1,440W. Two GPUs already push right up against that, and going to four is not a matter of shaving watts somewhere. Four RTX Pro 6000 plus the rest of the machine lands north of 2,600W, which is over 21A continuous at 120V. A four-GPU node needs 240V, full stop. And all of that is before anything else in the rack, which was sharing the same circuit.

A PSU in this class will not start on 120V anyway. Even if you found a 120V unit rated for 1,600W+, you would be pulling over 13A continuous on a circuit rated for 12A. You could pull a 120V / 20A circuit instead, but at that point you are already running new wire, so you may as well go 240V.

Most of this is a North American problem. Europe and most of the rest of the world run 230V to ordinary wall sockets, so a PSU that refuses to start on 120V is a non-issue there and a normal outlet already delivers something like 3,700W. If you are reading this from that side of the Atlantic, you get for free the part that cost me a panel upgrade. You would still want a dedicated circuit once you are running several machines, but you would not be doing it just to reach 240V.

The better reason is headroom. One workstation at 1,400W draws 5.8A at 240V, on a circuit good for 24A continuous. That is not headroom for one machine, it is room for several. A single L6-30 run carries multiple nodes plus whatever else ends up in the rack, which is the whole point: I can keep adding machines without calling an electrician again. On 120V every new node is another dedicated circuit and eventually another panel slot, and panel slots are what I ran out of in the first place.

Where it ended up

The rack has its own 120V and 240V feeds, each on a dedicated breaker. The dual RTX Pro 6000 workstation runs on the 240V side through the PX-2200. The second machine and the networking gear are on the 120V circuit. Even at full inference load the 240V circuit sits nowhere near its continuous limit, which was the point of running 30A rather than the minimum that would have worked. There is room for several more machines on that one circuit.