"We have 200 kW of IT load, let's buy a 400 kVA UPS with N+1 redundancy." From that sentence comes a 320,000 EUR quote, and four years later the UPS runs at 35% of nameplate capacity, the batteries have aged sooner than they should have, and operational losses have raised the electricity bill by 12% for no reason. Sub-1MW datacentres are the segment where money is lost quietly — nobody hands out a certificate for "well-sized UPS," but the gap between precise and grossly oversized sizing is 40–60% CAPEX and 8–15% OPEX year after year.
Mistake 1: confusing kVA and kW
The most common mistake in vendor quotes and, frankly, in client documents too. UPS units are sold in kVA (apparent power). IT load is calculated in kW (active power). The relation is `kW = kVA × PF` (power factor).
For **modern IT equipment** (servers from 2018+, Cisco Nexus 9k switches, HPE Synergy blade chassis) PF is typically 0.95–0.99 — power supplies are active PFC-corrected. For older blade chassis HP c7000, IBM BladeCenter or 2010-era telco gear, PF drops to 0.82–0.88.
The practical impact: 200 kW IT load at PF 0.95 = 211 kVA. At PF 0.85 (a mixed population of old and new) = 235 kVA. If a vendor offers 400 kVA "to be safe," you have 1.7× oversizing. **The price difference between 250 kVA and 400 kVA Eaton 93PM:** 38,000 EUR + ~30% higher standing losses in green mode + needlessly more space and cabling. A realistic reserve is 20%, not 100%.
Mistake 2: trusting nameplate values
The server spec sheet says a 1,600 W power supply. The client takes 40 servers × 1,600 W = 64 kW and adds a 20% reserve = 77 kW. Reality: the server at real utilisation (CPU 35–55%, idle GPU, light disk activity) draws 380–550 W, which is 24–35% of nameplate. A fully populated rack with 20 such servers has a **real peak of 8–12 kW, not 32 kW per nameplate**.
For AI workloads with NVIDIA H100/B200 nameplate and real load are closer — full FP16 training holds the GPU at 90–98% TDP. But for ordinary enterprise IT (databases, virtualisation hosts, file servers, K8s control plane) the gap between nameplate and measured load is 30–50%.
The practical step: ask for **historical metrics from the PDU** (Schneider APC, Raritan PX3, Vertiv Geist) for the last 6–12 months if you're migrating from an existing environment. For greenfield, use vendor sizing tools (Dell ESSA, HPE Power Advisor) with realistic workload profiles — not maximum.
Mistake 3: batteries aren't sized by runtime — they're sized by cycles and temperature
"We want 15 minutes runtime at 200 kW load." From that the calculation produces 50 kWh of batteries. The client buys 6 strings of VRLA 12V 100Ah because "the calculator said so."
The real equation is `runtime = (cell count × Ah × V × DoD × η_inverter) / load_W`. For VRLA batteries: - **DoD (Depth of Discharge):** 0.80 if you want a 4–5 year life, 0.90 if 2–3 years is fine and you'll swap them sooner - **η_inverter:** 0.92–0.95 on Eaton 93PM in double-conversion mode - **Temperature derating:** at 25 °C capacity is 100%, at 30 °C 90%, at 35 °C 75%, at 40 °C 55%. A battery room in Slovak datacentres reaches 28–32 °C in summer without dedicated HVAC
The practical impact: 50 kWh "paper" capacity at real DoD 0.80 and 30 °C → 36 kWh of usable energy → 10.8 minutes of runtime, not 15. The client thinks they have a 15-minute reserve, but reality is 10–11 minutes.
**Lithium-ion (LFP) UPS batteries** from Eaton, Vertiv Edge Li-ion or CSB DCS remove most of these deratings — DoD 0.90 without affecting life, operating temperature 0–45 °C without significant derating, 10+ years of life. Price: 2.3–2.8× more per kWh than VRLA. **TCO over 8 years still wins** in 70% of cases, because VRLA needs replacing once or twice in the same period.
Mistake 4: N+1 is confused with 2N or 2(N+1) without knowing the difference
This is cost mistake #1 in the sub-1MW segment. Three topologies, three completely different price tags:
- **N (single, non-redundant):** one UPS for the whole load. No critical datacentre allows this — failover means UPS bypass + manual cable switch. **CAPEX reference 200 kW:** 95–110k EUR (Eaton 93PM 250 kVA, batteries, switchgear).
- **N+1:** for example 3 × 100 kW modular UPS, one of which is redundant. On failure of one unit the others take over its load. **CAPEX 200 kW:** 140–165k EUR.
- **2N:** two independent UPS systems, each fully sized for the whole load. Servers with dual-PSU fed from two independent paths. **CAPEX 200 kW:** 240–280k EUR.
- **2(N+1):** two independent systems, each with internal N+1 redundancy. Tier IV uptime. **CAPEX 200 kW:** 420–470k EUR.
The question the client must ask: **what tolerance to a single UPS failure do we actually need?** For most enterprise IT (internal file server, ERP, CRM, web applications) N+1 is enough — the probability of simultaneous double failure is 10⁻⁶ per year. For payment processing, online trading, real-time bidding 2N is justified. 2(N+1) makes sense **only** with Tier IV certification with a concrete business case — only a handful per year in the EU.
In practice clients buy 2N "to be safe" for a use case that tolerates 10 minutes of outage per quarter. The gap vs N+1 is 100k EUR upfront and 25% more electricity forever (two systems running in parallel, each at ~50% load, a sub-optimal operating point).
Mistake 5: ignoring the difference between static bypass and maintenance bypass
A UPS has **two bypass modes** that often get confused:
- **Static bypass** (automatic bypass) — the UPS's internal electronics switch the load to utility power in milliseconds if it detects an internal fault. Always part of the UPS unit. The client doesn't trigger it manually — the firmware does.
- **Maintenance bypass** (external bypass switch) — a manual switch external to the UPS that allows you to **completely disconnect the UPS** from the supply path for service or replacement, while the load stays on utility. For maintenance this is mandatory, otherwise you have to power down the whole load.
Sub-1MW projects often don't install a maintenance bypass. Reason: it costs 8–15k EUR extra (rotary switch, breaker, cabling). The client doesn't realise that **without a maintenance bypass the UPS can't be replaced without an outage**. Over a 5–7 year UPS life that saving comes back as 4–8 hours of outage during the swap.
**SOTA practice for sub-1MW datacentres:** Eaton 93PM with integrated static bypass + external wrap-around maintenance bypass cabinet (Eaton MBP, Schneider QMB-160) + Kirk Key interlock so the technician can't switch the maintenance bypass by mistake on a live UPS. Price: 12–18k EUR. Payback: the first service cycle.
Concrete models in the 100–500 kVA segment
In this segment there are three genuinely competitive platforms:
- **Eaton 93PM** (100–500 kVA) — modular, hot-swappable power modules in 25 or 50 kW, double-conversion efficiency 96.5%, ESS (Energy Saver System) up to 99%. Good partner for LFP batteries via Eaton 9PXM EBM Li-ion. **Price 250 kVA:** 42–48k EUR for the unit alone.
- **Schneider Galaxy VM** (160–400 kVA) — the best service ecosystem in the EU, integrated EcoStruxure IT for monitoring. Efficiency 96%, eConversion mode 99%. **Price 250 kVA:** 38–44k EUR.
- **ABB DPA UPScale ST** (100–200 kVA modular platform) — the most modular approach, decentralised parallel architecture (DPA), no single point of failure within the unit. Higher entry cost, but unique for expansion without outage. **Price 200 kVA:** 46–52k EUR.
For sub-200 kVA Vertiv Liebert APM (modular, good price at the low end) or Riello Master MPS for cost-sensitive projects also makes sense.
A 10-minute decision framework
1. **What is real IT load (kW), not nameplate?** Ask for historical PDU data or a realistic vendor estimate. Add a 20% reserve, not 100%. 2. **What is the power factor of the mixed population?** Default 0.95 for modern servers, 0.88 for a mixed environment. 3. **What is acceptable runtime on outage?** For a datacentre with a diesel generator 5–8 minutes is enough (generator start time + buffer). For one without a generator plan 15–30 min. 4. **What is the battery room temperature?** If > 28 °C, increase battery count substantially or switch to LFP. 5. **What topology do you really need?** N+1 for 80% of use cases, 2N for business-critical, 2(N+1) only with a real Tier IV business case. 6. **Plan a maintenance bypass.** Non-negotiable.
A practical tip in the tender process
Demand explicitly in the quote: **sizing design with explicit PF assumption, battery runtime calculation at real room temperature, topology choice with business justification** (not "we usually do 2N"), and **option list for LFP vs VRLA batteries with an 8-year TCO comparison**. If the vendor can't bring these numbers together on an hour-long call, they're oversizing the system — that's not what you want to pay for.
---
*We do power design + UPS sizing for datacentres from 50 kW to 5 MW including retrofit projects. If you have a specific use case in tendering, we'll walk through the numbers and topologies in a 60-minute workshop with your real load profile.*