The call we've been getting more than any other lately starts with some version of the same sentence: "We ordered some GPU servers and I need to know if we can plug them in." Sometimes the servers are already on the loading dock. Sometimes the PO hasn't been cut yet. Either way, by the time we get there, IT has picked the workload, finance has approved the budget, and nobody's run the math with facilities. Here's what we usually find when we walk the closet — and why the answer is almost always the same.
What's actually in those boxes
A modern AI server is not a rack-and-stack 2U file server with a couple of extra Ethernet ports. It's a dense, power-hungry specialized appliance designed to run GPUs at maximum sustained draw for hours at a time. The scale has moved fast:
- An NVIDIA H100 GPU runs up to 700 W by itself. A DGX H100 server — 8 GPUs plus CPUs, NVSwitch fabric, InfiniBand NICs, and DPUs — draws roughly 10–11 kW per chassis at full load. [1]
- The newer DGX H200 sits in a similar power envelope at roughly 10.2 kW, and NVIDIA's own documentation specifies it requires a 200–240 V three-phase configuration — a standard wall outlet is explicitly called out as unsuitable. At 208 V three-phase, it draws about 28.3 A per phase at max load. [2]
- The DGX B200 (Blackwell) pushes this to about 14.3 kW per chassis, with each B200 GPU running 1,000 W air-cooled or 1,200 W liquid-cooled. [1][3]
- The GB200 NVL72 rack-scale system — 72 Blackwell GPUs in a single rack — demands up to 120–140 kW per rack, with mandatory direct-to-chip liquid cooling. [1]
For context: a traditional enterprise server rack — the kind your existing closet was probably designed around — typically lives in the 5–15 kW range per rack. [4] A single DGX B200 is already at the top of that range. A rack of them is somewhere none of your existing infrastructure was asked to go.
Why 125 A single-phase hits a ceiling fast
Many mid-enterprise buildings have exactly the kind of setup this article is named for: a data closet fed by a 125 A single-phase 208 V circuit — two legs of the building's 208Y/120V three-phase service, pulled into a subpanel in the IDF or MDF. It's been fine for years. For workloads up to maybe 15 kW per rack, it's still fine. But the AI server changes the ceiling.
Let's run the math, because this is where the reality shows up.
A single-phase 208 V, 125 A circuit delivers a theoretical 26 kVA of apparent power (V × I). But NEC 210.19(A)(1) requires that branch-circuit conductors and breakers be sized to handle 125% of continuous load — equivalently, continuous loads are limited to 80% of the branch-circuit rating. [5] A server running GPU workloads at sustained high utilization is the textbook definition of a continuous load (defined in the NEC as any load running at maximum current for three or more hours).
Apply the 80% rule and that 26 kVA circuit gives you about 20–21 kW of usable continuous capacity. That's enough for one DGX H200 running hard (10.2 kW), with a little headroom for networking gear and a small UPS — and not much else. Add a second DGX and you're over the line.
And that's before we talk about cooling load, PDU overhead, the cost of balancing two circuits that must not trip each other under startup surge, or the fact that large single-phase loads start to create noticeable voltage imbalance in your building's 208Y/120V panel as a whole.
Why three-phase is the answer (not just "more amps")
Here's the number that makes the conversation different. That same 125 A at 208 V, delivered as three-phase, gives you roughly 45 kVA of apparent power — because three-phase power is computed as √3 × Vline-to-line × I, not just V × I. [6] After the NEC 80% continuous-load derate, that's about 36 kW usable — nearly 75% more capacity out of the same ampere rating on the breaker.
You get that capacity by adding a third current-carrying conductor rather than by upsizing the others, which is a meaningfully cheaper copper bill than going from 125 A single-phase to 200+ A single-phase. And you get it in a form that's better for the equipment, too: three-phase power delivered to a rack is naturally load-balanced across three legs, reducing voltage drop, reducing the neutral-conductor heating that single-phase UPS installations have to plan for, and giving you cleaner power at the GPU's input stage.
Figures synthesized from industry references [7][6]; any actual installation should be load-calculated against the specific equipment nameplates and reviewed by a licensed electrical engineer.
Why it's almost always 208 V (and when it isn't)
In the US commercial building stock, the default three-phase service is 208Y/120V — three hot legs at 208 V line-to-line, 120 V from each leg to neutral. It's what most office buildings have at the service entrance, and it's what most mid-enterprise AI deployments will land on. Critically, every major AI server platform — DGX, HGX, and the big Supermicro/Dell/Lenovo reference designs — accepts 200–240 V input natively. [2]
The next step up is 480Y/277V, which is what you see in large data centers and purpose-built colocation space. It reduces conductor sizes further and supports the 100+ kW-per-rack densities that a GB200 NVL72 requires. But 480V requires dedicated step-down transformers for anything running at 208/120V (your lights, your office equipment, your existing PDUs), and it brings a different tier of electrical code attention. For most SMB, SLED, and mid-enterprise AI deployments — single DGX, small clusters, specialty rendering or inference racks — 208Y/120V three-phase is the right answer. 480V is a different project.
What else changes in the closet
Moving to three-phase isn't a single-line-item change. Here's the honest scope of what usually comes with it:
The PDU
Your existing single-phase rack PDU doesn't accept three-phase input. You're looking at a new three-phase rack PDU, typically a 30 A or 60 A unit with a receptacle like L21-30, L22-30, or an IEC 60309 plug. Good news: a three-phase rack PDU can still deliver 120 V convenience circuits for non-server gear, because each phase-to-neutral leg is 120 V. [6] You don't lose flexibility; you gain capacity.
The UPS
Single-phase UPS systems top out well below the capacities AI workloads demand. A three-phase UPS — typically a double-conversion design in the 20–60 kVA range for closet-scale AI, up to 200 kVA+ for dedicated AI rooms — is almost always the right path. If you're still on VRLA batteries and the system is end-of-life, this is a natural moment to have the Li-Ion vs VRLA conversation as part of the upgrade, because lithium UPS systems are built to handle the higher power densities and tighter footprints that AI deployments need.
The feeder
The conductors feeding your existing 125 A single-phase subpanel are probably not reusable for a new three-phase feeder — you need a third hot conductor, your conduit fill changes, and your existing breaker is single-phase. In most buildings where 208Y/120V three-phase is already available at the main panel, the path is extending a new three-phase feeder from the main panel into the closet, landing on a new three-phase subpanel. It's an electrical contractor job, needs a permit, and needs to be coordinated with a facilities outage. Plan on a half-day to two-day scope of work depending on conduit runs, panel capacity, and whether any AHJ complications surface.
The cooling question — and when it stops being optional
Power is half the conversation. The heat goes somewhere, and most existing closets weren't designed to move it. As a rough guide:
- Under ~10 kW/rack: Standard room cooling usually sufficient, as long as your airflow and hot/cold aisle setup are reasonable.
- 10–30 kW/rack: In-row cooling or rear-door heat exchangers become the right tool. Traditional perimeter CRAC units start to struggle with the concentrated heat load.
- 30–50 kW/rack: The upper end of what air cooling can realistically deliver in an enterprise environment. Above this, the physics starts to fight you. [4]
- Above ~50 kW/rack: Liquid cooling — direct-to-chip cold plates or rear-door heat exchangers with coolant distribution units (CDUs) — becomes necessary, not optional. NVIDIA's own GB200 NVL72 ships with mandatory direct liquid cooling. [1]
For most mid-enterprise AI deployments — one to four DGX-class systems in a single rack — you're in the 10–45 kW/rack range and air cooling is still workable, provided the room has the BTU capacity and the airflow is engineered properly. If you're going larger than that in a single room, the liquid cooling conversation is a project of its own, and it usually means a purpose-built room rather than a retrofit of an existing closet.
The planning order that actually works
We've walked enough of these projects now to have a clear sequence that keeps them on track. It looks like this:
- GPU platform and count first. How many DGX- or HGX-class systems, in what racks, running what workloads. This determines every downstream decision.
- Per-rack load calculation. Sum the nameplate, apply a reasonable diversity factor (GPUs running inference rarely pull 100% sustained, training often does), add cooling overhead, add PDU and monitoring. Get the honest kW per rack.
- Single-phase vs three-phase call. If you're above ~15 kW per rack, you're in three-phase territory. Below that, honestly evaluate whether you're planning to grow — if you'll cross the line in two years, do the three-phase work now.
- Feeder and panel review. Does the main panel have capacity? What's the route from main panel to IDF/MDF? Is there a permit question?
- PDU and UPS selection. Match to the load and the input voltage. Don't buy a single-phase UPS for a deployment that's obviously going three-phase in two years.
- Cooling capacity check. Does the room have the BTU capacity? If not, does in-row cooling make sense, or is the heat load pushing you toward a liquid cooling conversation?
- Schedule the work. Electrical, cooling, and equipment installation all in the right sequence, with the right downtime windows.
Doing these in order means every downstream decision is grounded in the ones above it. Doing them out of order — which is what happens when the GPU servers show up first and the facilities conversation starts second — means you pay for infrastructure twice, or you get a deployment that works for six months and then starts tripping breakers.
- IntuitionLabs (2026). NVIDIA HGX Platform: Data Center Physical Requirements Guide. intuitionlabs.ai/articles/nvidia-hgx-data-center-requirements
- Uvation (2025). NVIDIA DGX H200 Power Consumption: Key Facts & Requirements. uvation.com
- GPU.fm (2026). NVIDIA B200 GPU: Complete Pricing, Specs & Buyer's Guide. gpu.fm/blog/nvidia-b200-complete-buyers-guide-2026
- Introl. Liquid Cooling vs Air: The 50kW GPU Rack Guide. introl.com/blog/liquid-cooling-gpu-data-centers-50kw-thermal-limits-guide
- EE Power (NEC Commentary). National Electrical Code Basics: Sizing and Protecting Branch-circuit Conductors (NEC 210.19, 210.20, 310). eepower.com
- Enconnex. Examining Single-Phase vs. 3-Phase Power & PDUs. blog.enconnex.com
- Tripp Lite / Eaton. Power Distribution Units (PDUs) — single-phase vs three-phase capacity reference. tripplite.eaton.com