When a GPU rack breaks your UPS: the inrush current calculation your vendor doesn’t run

The calculation your GPU rack vendor didn’t run — and that can take you offline

Your UPS is sized for 1 MW. Your vendor sells you a rack with 4 NVIDIA H100 servers, “10 kW each, 40 kW per rack total, no problem for your UPS.” The rack arrives, you energize it during a non-critical window, and the UPS doesn’t flinch. The next morning, during a scheduled maintenance reboot, the UPS trips the power module overcurrent protection and transfers to static bypass, leaving the critical load exposed.

Your vendor wasn’t lying about the 10 kW nominal rating. What they omitted is the inrush current at energization.

What inrush current is — and why it’s different on GPU racks

Inrush current is the transient current peak that an electronic device demands at the exact moment of energization, while its internal capacitors charge from zero. For a standard server, that peak reaches ~3× nominal current over 20-50 milliseconds [Source: Enova, AI GPU Inrush Current Measurement, 2026].

On an AI GPU server with hardware like NVIDIA H100 or A100, the onboard capacitors are significantly larger (multiple high-power VRMs for the GPUs, memory banks, NVSwitches). The peak reaches 3-5× nominal current, with a longer decay window because of the higher capacitance.

Concrete numbers [Source: Enova, citing measurements with Raritan PX4 PDU, 2026]:

• H100 server nominal 6 A → inrush peak 30 A.
• A100 server nominal 5 A → inrush peak 20-25 A.
• Full DGX H100 (10.2 kW nominal) → measured inrush peak 40-50 A per PSU, multiplied by 6 PSUs.

And the event happens on every boot, every restart, every power cycle. It doesn’t appear in the logs because most standard PDUs don’t sample at the sub-millisecond resolution needed to capture the peak.

What your vendor omitted — the calculation that has to happen

Take a rack of 4 DGX H100 servers, with 6 PSUs per server, 3.3 kW per PSU [Source: NVIDIA DGX H100 User Guide, docs.nvidia.com]:

• Steady-state load: 4 × 10.2 kW = 40.8 kW.
• Nominal rack current (3Φ 415 V, assuming 0.95 power factor): ~60 A per phase.
• Simultaneous inrush (worst case): 24 PSUs starting in the same maintenance cycle. If each PSU inrushes at 2× nominal (~30 A peak), the rack pulls ~720 A instantaneous from the distribution circuit.
• A 100 A breaker on the rack PDU sees that peak and trips overcurrent protection.

Why did the first energization work but the reboot fail? Because on the first energization, the capacitors were discharged but the UPS had margin (it was conservatively sized and the capacitors charged on a ramp, not in a single event). On the reboot, the rack de-energizes cleanly and the simultaneous reconnection of 24 PSUs trips the breaker — or the UPS power module if bypass wasn’t active.

The math of correctly sizing the UPS

To size a UPS that supports a GPU rack without failing on reboot, the calculation isn’t steady-state power × 1.25 (the typical sizing factor). It’s:

Steady-state power × inrush diversity factor × N+1 redundancy factor

where:

• Inrush diversity factor recognizes that not all PSUs start at exactly the same millisecond. In practice, diversity reduces the aggregated peak to ~40-60% of the worst-case simultaneous peak. Even so, on a Tier III redundant N+1 UPS, the peak can exceed the capacity of a single module.
• N+1 redundancy factor adds 20-40% over nominal capacity to absorb the transient without having to transfer to bypass.

Example for a 4-server DGX H100 rack:

• Steady: 40.8 kW
• Worst-case simultaneous aggregated peak: ~150 kW (3.7× steady, assuming high-capacitance PSUs)
• Peak with 50% diversity: ~75 kW
• UPS must absorb 75 kW peak without dropping to bypass = size for ~95 kW nominal (a 100 kW module with margin).

Your original 1 MW UPS wasn’t mis-sized for 40 kW steady. But the typical individual power module is 100-250 kW, and a 100 kW module sees the 75 kW peak from the GPU rack and decides not to take it. That triggers the transfer to bypass — and leaves the critical load exposed for the duration of the transfer (typically < 4 ms under static bypass, but the event gets logged as a Tier II-III alarm in any serious monitoring system).

The questions you have to ask BEFORE installing the GPU rack

1. How many PSUs per server and what individual capacity? For a DGX H100, it’s 6 PSUs at 3.3 kW. For a generic HGX H100, typically 4-8 PSUs depending on the OEM.

2. What is the manufacturer’s declared inrush current? NVIDIA publishes this data in its electrical design datasheet. If your vendor doesn’t have it, that’s a red flag.

3. Does the UPS have high-speed static bypass with transfer time < 4 ms? If yes, the worst case is a transparent transfer. If not, the scheduled reboot leaves you with an outage of half a cycle (~10 ms in some electromechanical designs).

4. What energization sequence will you use for all racks in the batch? If you energize racks in parallel during a commissioning batch, the aggregated inrush peak can exceed the upstream PDU capacity. Sequencing rack by rack with 30-60 seconds between each one protects the distribution circuit and gives the UPS time to absorb each individual peak.

5. Are the rack PDU breakers C-curve (instantaneous) or D-curve (delayed)? A C-curve breaker trips at 5-10× In instantaneously. For server inrush, you want D-curve or adjustable breakers with configurable short-term threshold.

6. Does the rack have soft-start or energization sequencer? Some server chassis allow staggered PSU startup. If your rack has that option, the inrush diversity rises to 80-90% and the aggregated peak drops to ~20 kW, within the safe range of any Tier III UPS.

The mis-sized rack — operational symptoms

Symptom 1: scheduled reboot that triggers an alarm on the UPS module. The reboot executed, the module transferred to bypass, returned cleanly after 4-6 ms. The monitoring records the bypass as a Tier II-III event. If nobody reads the logs, this repeats every maintenance window.

Symptom 2: PDU breakers that trip when energizing new racks. If your installation procedure energizes racks one at a time, the first ones pass. When you energize the second or third rack simultaneously (common in batch commissioning), the aggregated current from the fourth rack onward exceeds the upstream PDU capacity.

Symptom 3: UPS power module breakers that trip under AI peak load. Modern AI workloads aren’t flat — they spike to 100% GPU usage during training, decay to 30% during inference. If your upstream breaker is calibrated to steady-state peak, not to the combination of inrush + AI peak, it will trip under sustained maximum load.

Symptom 4: PDU branch circuit breakers that trip after months of normal operation. This is the most deceptive symptom. The thermal calibration of the breaker drifts over time. A breaker that accepted 75 kW peak at commissioning may start tripping at 18 months because its thermal element has become more sensitive.

How to prevent it — the correct sizing procedure

5-step procedure before installing a GPU rack:

1. Request from your rack vendor: datasheet for each PSU with peak inrush current, inrush duration, and decay profile. If they don’t have it, don’t buy the rack without measuring first.

2. Calculate the aggregated peak of the full rack assuming a conservative simultaneity factor (70%). For 4× DGX H100 with 3.3 kW PSUs, peak ~75-90 kW.

3. Size the UPS module so the peak lands below 80% of its nominal capacity, leaving margin for sustained AI peak (40 kW + 100% spikes during training).

4. Configure PDU breakers with D-curve or adjustable, calibrated to 125% of the rack’s expected peak (not the steady-state).

5. Implement an energization sequencer on the rack. Staggered soft-start: PSU1 at T+0, PSU2 at T+200ms, PSU3 at T+400ms, PSU4 at T+600ms. Reduces the aggregated peak to 30-40% of the worst-case simultaneous.

If your rack vendor can’t deliver the PSU inrush data, there’s a transparency problem you need to escalate. NVIDIA publishes the data in its Design Guide. OEMs that don’t publish it are hiding information you need to do your commissioning safely.

What your UPS is NOT telling you about this GPU rack

Your UPS has telemetry: voltage, current, frequency, power factor, THD, internal temperature. It does NOT have sub-millisecond current sampling per branch. That means the inrush peaks your PDU breaker sees don’t appear in your UPS logs. The event lives in the breaker’s thermal curve, invisible to your monitoring system.

If you want real visibility, you need intelligent PDUs with high-frequency sampling (Raritan PX4, Vertiv Geist, Schneider AP8881) that record inrush events per outlet and let you verify post-mortem that the scheduled reboot didn’t leave a breaker on the edge of tripping.

Sources

1. NVIDIA — DGX SuperPOD Data Center Design Guide H100, Electrical Specifications, docs.nvidia.com
2. NVIDIA — DGX H100/H200 User Guide, Component Descriptions, docs.nvidia.com
3. Enova — AI GPU Inrush Current: Measurement & PDU Solutions, 2026
4. HighPoint — Why H100 & H200 GPUs Need 1300W PSU Support, 2026
5. IntuitionLabs — NVIDIA HGX Platform: Data Center Physical Requirements, 2026