Datacenter UPS Sizing Guide
kVA vs kW math, load profiling, redundancy topology, and brand selection for AI-era density.
Key takeaways
- Size UPS capacity in kW, not kVA - power factor matters for modern IT loads (close to 1.0 PF).
- Plan for 70-80% load utilization at steady state; never run a UPS at 100% rated capacity.
- N+1 is the working minimum for production datacenters; 2N (full A/B feed) is standard for Tier III+.
- AI training clusters push rack densities from 8-15 kW historical to 40-100+ kW - rethink branch circuit and UPS sizing.
- Runtime is a battery question, not a UPS question: most UPS configurations run 5-15 min internal, then handoff to generator.
In this guide
Load profiling: what to measure
Before sizing a UPS you need three numbers per rack or per circuit: steady-state draw, peak transient draw, and future growth headroom. Don't trust nameplate ratings - those are worst-case maximums, and most real loads sit at 40-70% of nameplate.
Tools
- In-line power meter on the PDU branch for 7-14 days. Captures real diurnal variation.
- Out-of-band BMC/iLO/iDRAC telemetry for server CPU + PSU draw over a week. Gives you idle vs load delta per workload type.
- Switch SNMP for PoE budget consumption if your rack hosts edge networking gear.
| Load type | Idle draw | Steady-state | Peak transient |
|---|---|---|---|
| 1U general-purpose server (Xeon Silver, 128GB) | 80-120W | 200-350W | 500-650W |
| 2U dual-socket (Xeon Gold, 512GB, NVMe) | 150-220W | 450-700W | 1000-1300W |
| 4U GPU compute (4x H100 or A100) | 400-600W | 2200-3000W | 4500-5500W |
| 8U GPU compute (8x H100, DGX-class) | 700-1000W | 5000-7500W | 10000-12000W |
| 48-port PoE++ switch with 960W PoE pool | 50-80W | 200-400W | 1000-1100W |
| SAN array, 24-drive 2U | 250-350W | 400-550W | 700-850W |
kVA vs kW: the math that matters
UPS capacity is rated in two ways. kVA (apparent power) is the volt-amp product the inverter must deliver. kW (real power) is the work the load actually draws. They're related by power factor:
kW = kVA × Power Factor (PF)
Older equipment with capacitor-input power supplies sat around 0.7-0.8 PF. Modern server PSUs are nearly always Active PFC, running 0.95-0.99 PF. That changes the sizing math.
| UPS rating | PF 0.8 (legacy) | PF 0.9 (mixed) | PF 1.0 (modern) |
|---|---|---|---|
| 3 kVA | 2.4 kW | 2.7 kW | 3.0 kW |
| 10 kVA | 8.0 kW | 9.0 kW | 10.0 kW |
| 20 kVA | 16.0 kW | 18.0 kW | 20.0 kW |
| 40 kVA | 32.0 kW | 36.0 kW | 40.0 kW |
| 100 kVA | 80.0 kW | 90.0 kW | 100.0 kW |
Unity-PF UPS systems
Vertiv Liebert EXL, APC Symmetra PX, Eaton 9395P and most modern modular UPS systems are rated at unity power factor (1.0 PF). The kVA and kW ratings are the same number. Older models with 0.8 or 0.9 PF rating force you to oversize by ~20% to get the kW you actually need.
Redundancy topologies
| Topology | Components | Survives | Cost vs N | Typical use |
|---|---|---|---|---|
| N | 1 UPS sized for full load | Utility outage | 1.0x | Branch office, edge closet, SMB |
| N+1 | 2 UPS, either covers load alone | Utility + single UPS fault | ~1.6x | Mid-size IT, Tier II datacenter |
| 2N | 2 independent A/B paths to load (dual-PSU servers) | Full single-path failure (UPS, PDU, branch breaker) | ~2.0x | Tier III datacenter, financial, healthcare |
| 2N+1 | 2N plus a spare in each path | Two simultaneous faults | ~2.5x | Tier IV datacenter, mission-critical |
| Distributed redundant | 3+ UPS feeding shared bus through STS | Single UPS fault, with hot-swap maintenance | ~1.5-1.8x | Modular datacenter, hyperscale |
AI-era density: what changed
Historical enterprise rack density ran 5-15 kW per rack. AI training clusters using 8x H100 or H200 GPUs per chassis pull 7-12 kW per node, and four nodes fit in a 42U rack. That puts a single rack at 30-50 kW - and inference racks with NVL72 or similar can push 80-120 kW.
This breaks several assumptions in legacy datacenter design:
- Branch circuit capacity. A standard 208V/30A circuit delivers ~5 kW. AI racks need three-phase 60A or 100A drops, or 415V power straight to in-rack PDUs.
- UPS topology. Per-rack UPS becomes economically painful at 50+ kW per rack. Centralized 250-1000 kVA UPS with bus-tie redundancy is the working pattern.
- Cooling tightly couples to power. 40+ kW per rack requires rear-door heat exchangers or direct-liquid cooling; air alone struggles past 25 kW.
- Battery runtime trade-off. At 100 kW per rack, even 5 minutes of battery means a substantial battery cabinet. Most AI deployments shorten UPS runtime and prioritize fast generator transfer.
Brand comparison
| Brand / line | Topology | Capacity range | Sweet spot | Strengths |
|---|---|---|---|---|
| Vertiv Liebert GXT5 | Online double-conversion | 1-20 kVA | Mid-size IT closets, branch datacenters | Wide voltage window, LCD with PF readout, hot-swap battery |
| Vertiv Liebert EXM2 / EXL | Modular online | 30-1100 kVA | Datacenter and AI-cluster build-outs | Unity PF, modular scaling, 97%+ ECO mode |
| APC Smart-UPS SRT | Online double-conversion | 2.2-20 kVA | Small to mid datacenter | Strong management software, broad dealer base |
| APC Symmetra PX | Modular online | 10-500 kVA | Mid-tier datacenter | Mature modular platform, common installed base |
| Eaton 9PX | Online double-conversion | 1-22 kVA | IT closet through mid-rack | Unity PF, virtualization integration, ABM battery management |
| Eaton 9395P | Modular online | 250-1200 kVA | Tier III/IV datacenter, AI clusters | Variable Module Management, ESS mode 99% efficient |
| Tripp Lite SmartOnline SU-series | Online double-conversion | 1-30 kVA | Mid-size IT, edge compute | Strong value, included PowerAlert software |
| Tripp Lite SmartPro SMART-series | Line-interactive | 0.5-3 kVA | Workstation, network closet, light IT | Strong value at low capacity, simple operation |
Build recommendations by facility size
Edge / branch office (under 10 kW total)
- 2-3 kVA online or line-interactive at top of each rack
- 5-15 min runtime (single battery pack)
- No generator - graceful shutdown is the goal
- Pattern: Eaton 9PX 3000, Vertiv Liebert GXT5-3000, Tripp Lite SU3000RTXL
Mid-size server room (10-50 kW)
- One 20-40 kVA online UPS with external battery cabinet
- Optional N+1 with a second matching unit and STS
- 10-20 min runtime, generator handoff if uptime SLA demands
- Pattern: Vertiv Liebert GXT5-20kVA, APC Smart-UPS SRT 20kVA, Eaton 9PX22K
Datacenter / colo cage (50-300 kW)
- Modular online UPS at 80-160 kVA, scalable to N+1
- External battery cabinets sized for 5-10 min runtime
- Standalone generator with 30-second transfer
- Pattern: Vertiv Liebert EXM2-100, Eaton 9395P-200, APC Symmetra PX
AI cluster / hyperscale (300 kW+)
- Centralized 500-1000+ kVA modular UPS at unity PF with bus-tie redundancy
- 415V three-phase distribution to in-row or in-rack PDUs
- 5-7 min battery (generator transfer), or 15+ min for missions where rolling restart is unacceptable
- Pattern: Vertiv Liebert EXM2-625 in N+1, Eaton 9395P-1000, dual-feed 2N for Tier III+
Runtime planning
Runtime is not a UPS function - it's a battery function. The UPS converts and conditions; the battery holds energy. Battery sizing is a separate exercise.
| Load | 5 min | 15 min | 30 min | 60 min |
|---|---|---|---|---|
| 3 kW | Internal pack | 1 external EBM | 2 EBMs | 4 EBMs or 1 battery cabinet |
| 10 kW | Internal + 1 EBM | 3 EBMs | 1 battery cabinet | 2 cabinets |
| 40 kW | 1 cabinet | 2-3 cabinets | 5-6 cabinets | Lithium recommended |
| 100 kW | 2-3 cabinets | 6-8 cabinets | Lithium recommended | Lithium only practical |
FAQ
How much UPS headroom should I build in?
Plan for 70-80% utilization at steady state. That gives you 25-40% growth headroom plus margin to absorb transient peaks. Running a UPS above 90% capacity for sustained periods accelerates battery wear and leaves no margin for the inrush event.
Is online double-conversion always required?
For datacenter and AI work, yes. Online UPS provides clean sinewave output continuously and is the only topology that fully isolates the load from utility events. Line-interactive is acceptable for branch offices, IT closets, and workstations where occasional brief transfers to battery are not a problem.
Can I run two UPS in parallel for N+1 without a special model?
No. Parallel UPS operation requires units designed for paralleling - they share a control bus and synchronize output. Most enterprise online UPS from Vertiv, APC, Eaton support parallel; smaller line-interactive units typically do not. Check the spec sheet for "parallel capable" or "modular".
What happens when the generator transfers?
The UPS holds the load on battery during the 5-30 second transfer window. Once generator output stabilizes (frequency and voltage in spec), the UPS resumes charging from generator power. Generator must be sized to support UPS recharge plus the IT load - typically 1.5-2x IT load capacity.
How long do UPS batteries last?
VRLA batteries: 3-5 years at typical room temperature, 18-30 months in hot rooms. Lithium-ion (LFP): 8-12 years. Always specify temperature monitoring and replace before end-of-life - a failed battery on transfer is a complete outage. Most enterprise UPS have battery monitoring that flags impending failure.
Do I need a maintenance bypass?
For anything in production, yes. A maintenance bypass switch lets you fully isolate and service the UPS while the load stays powered from utility. Without it, every UPS service window is a downtime window. External wraparound bypass cabinets from Vertiv, APC, Eaton are designed for this.
Power infrastructure built for the AI era
Channel-direct sourcing on Vertiv, APC, Eaton, Tripp Lite, Ubiquiti. Working integrators on the phone.
Browse Power & UPS UPS Topology Guide Rack PDU Guide