Power applications on hub → Scenario
GPU rack power calculator (AI kW & kVA)
Screen kW per GPU, kW per rack, and row-level demand for AI training or inference clusters before transformer, UPS, PDU, and liquid-cooling design.
Who this scenario is for
Best for: AI infrastructure planners, colocation engineers, and electrical designers estimating rack and row kW from GPU count, TDP, and host/network overhead before facility quotes.
Not ideal for: Network fabric architecture, chip-level thermal CFD, or final NEC ampacity studies—those need specialized design beyond screening math.
Quick answer
Rack kW ≈ (GPU count × GPU kW) + host kW + network kW + 15% margin. Per-GPU planning often uses 0.3–0.7 kW for datacenter accelerators—verify OEM TDP for your SKU. kVA ≈ rack kW ÷ PF (often 0.95).
That formula screens one server. It is not a rack-scale system. NVIDIA’s DGX GB200 user guide states rack power consumption of approximately 120 kW for an NVL72 rack. The GB300 NVL72 reference architecture lists a full rack requiring up to 142 kW. Use those as product examples, then roll your own average in the AI data center power calculator.
GPU power reference (planning TDP)
| Accelerator class | Planning kW / GPU | Notes |
|---|---|---|
| Previous-gen datacenter GPU | 0.25–0.35 kW | Lower TDP training/inference mix |
| H100-class / high TDP | 0.35–0.70 kW | Workload and power cap dependent |
| Multi-GPU node (per GPU share) | Same ranges | Sum all GPUs in chassis |
| Host CPU + memory | 0.5–1.5 kW per node | Add once per server, not per GPU |
| ToR switch (allocated per rack) | 0.3–0.8 kW | Share across racks if metered separately |
Last reviewed: June 2025. TDP varies by SKU, firmware cap, and workload.
Rack kW examples by GPU count
| Configuration | Subtotal kW | +15% margin | kVA @ PF 0.95 |
|---|---|---|---|
| 4× GPU rack | 4 kW | ~4.6 kW | ~4.8 kVA |
| 8× GPU rack | 6 kW | ~6.9 kW | ~7.3 kVA |
| 8× GPU @ 0.7 kW each | 7.6 kW | ~8.7 kW | ~9.2 kVA |
| 16× GPU dense tray | 10 kW | ~11.5 kW | ~12.1 kVA |
Row and facility screening
| Task | Given | Screening result |
|---|---|---|
| Row kW (10 racks) | 10 × 6.9 kW | ~69 kW IT row |
| PDU / branch | 7 kW rack @ 208V 3φ PF 0.95 | ~20 A/phase + margin → see PDU sizing |
| Transformer (AI row) | 69 kW @ PF 0.95 + 25% margin | ~91 kVA — see step-down transformer |
| UPS kVA (single rack) | 7 kW @ PF 0.95 + 20% margin | ~8.8 kVA frame screening |
| PUE (liquid-assisted hall) | 500 kW IT, PUE 1.2 | 600 kW facility — see PUE planning |
Three-phase ampacity context: 3-phase power calculation examples. Data center transformer methodology: Transformer sizing for data centers.
UPS kVA screening for GPU racks
| Rack kW | Planning kVA | Typical use |
|---|---|---|
| ~5 kW | ~6–7 kVA | Small inference rack |
| ~7 kW | ~8–10 kVA | One 8× GPU server |
| ~12 kW | ~15 kVA | One dense GPU node |
| ~20 kW | ~25 kVA | Higher single-cabinet screening |
Verify backup minutes separately in UPS for server rack—do not size UPS on kVA alone. A rack-scale liquid-cooled system is a different problem: NVIDIA documents about 120 kW for a DGX GB200 NVL72 rack and up to 142 kW for a GB300 NVL72 rack. Those figures are not a universal AI rack rating.
Liquid cooling note
Direct-to-chip or rear-door heat exchangers may improve facility PUE but do not reduce electrical feed kW at the rack PDU—size power for full IT draw. Cooling pump power may appear in facility numerator for PUE, not in rack IT kW.
Key variables
- GPU TDP vs cap: Firmware power caps lower average draw but peaks matter for breaker coordination.
- Host and storage: CPU, NVMe, and NIC watts add to GPU subtotal per node.
- Network: ToR and spine allocation depends on meter boundaries.
- Margin: 15–25% common; AI refreshes may need upper band until metered.
- Redundancy: Dual-cord servers need A+B feed sizing policy.
- Simultaneity: Row sums may use <1.0 only with measured job alignment.
Worked example: one 8× GPU server
8 GPUs × 0.5 kW = 4 kW; host + NIC ≈ 2 kW; subtotal 6 kW; 15% margin → ~6.9 kW for that server, not for a full AI rack. At PF 0.95 → ~7.3 kVA for branch screening. At 208V 3-phase → ~19 A/phase before margin—coordinate with PDU sizing. To roll many racks into facility MW, use the AI data center power calculator with your own kW per rack.
Verify rack kW — load calculator
Roll up GPUs, hosts, and network gear in the UPS Load Calculator, then convert to kVA and check upstream redundancy.
See also: Data center transformer guide · kW vs kVA · Rack power density
Next steps
- AI data center power calculator — racks × kW/rack × PUE.
- UPS Load Calculator — device rollup to kW.
- UPS Capacity Calculator — kVA with PF and margin.
- UPS Redundancy (N+1) — module count at kVA.
- Rack power density — general IT rack rollup.
- PDU sizing — branch ampacity.
- Step-down transformer sizing — hall kVA.
- PUE planning — facility overhead.
- Server rack UPS runtime — backup minutes.
AI infrastructure planning path
- GPU kW per unit — reference table on this page.
- One server — 8× GPU worked example on this page, or a documented rack-scale kW such as NVIDIA’s ~120 kW NVL72.
- AI data center power — IT MW and facility MW.
- Rack kW / density
- PDU sizing
- UPS Load Calculator
- Server rack UPS runtime
- UPS redundancy (N+1)
- UPS Battery Calculator
- Step-down transformer
- Transformer Size Calculator
- Liquid cooling — note on this page; IT kW unchanged at PDU.
- HVAC Capacity
- PUE planning
- Generator Size Calculator
Assumptions and disclaimer
GPU TDP and power caps vary by workload, SKU, and vendor. Values on this page are planning estimates for early electrical and facility screening—not procurement or stamped design. Confirm with meter data and OEM specs before ordering PDUs, whips, and transformers.
Frequently asked questions
Back to Power applications on hub.
How much power does an H100 GPU use?
Planning range ~0.3–0.7 kW per GPU plus host/network; confirm SKU datasheets and power cap policy.
How do I size UPS for a GPU rack?
kVA = kW ÷ PF with 15–25% margin. Check N+1 and runtime separately—see UPS capacity and server rack runtime guides.
Does liquid cooling change rack kW?
IT kW at the PDU stays similar; facility PUE may improve. Size electrical feed for full rack demand.
How many kW per AI rack is typical?
There is no single figure. One multi-GPU server is often under 15 kW. NVIDIA’s DGX GB200 user guide states about 120 kW for an NVL72 rack, and the GB300 NVL72 reference architecture lists up to 142 kW for a full rack. Those are product examples. Meter the SKU you buy, then use the AI data center power calculator.
What power factor should I use for GPU loads?
Screen PF 0.95–0.99; use PDU meters when available for kVA conversion.
How do I scale from one rack to a row?
Sum rack kW; apply simultaneity only with measured evidence—training clusters often use 0.85–1.0.
Do GPU racks need dual power feeds?
Most use dual cords. Size each feed for full rack load if the other cord is offline during maintenance.
How does GPU rack power affect transformer sizing?
Roll up row kW to hall kVA with overhead, margin, and harmonic derating on the step-down transformer scenario page.
What is the difference between TDP and metered GPU power?
TDP is a design rating; peaks may exceed averages. Size branches for peak or capped power policy.
Should I include network switches in rack kW?
Yes—include ToR and in-rack network unless metered on a separate branch.
Often planned together
- PDU sizing — branch ampacity for high kW racks
- Step-down transformer sizing — hall kVA from AI row kW
- PUE planning — facility overhead for AI halls