How to Calculate UPS Power Loss in a Data Center

Ask a vendor for the efficiency of a 1,000 kW UPS and you will get one number: 96.2 percent. Ask what it costs to run and you will get a shrug. The shrug is the honest answer, because a UPS does not have an efficiency — it has a loss curve, and the loss curve has three parts with completely different behavior. One part runs 8,760 hours a year whether or not a single rack is plugged in. One part follows every kWh you deliver. One part grows with the square of the load and only shows up when you push the unit hard.

Treat the 96.2 as a constant and you will misprice everything: the cost of 2N redundancy, the cost of oversizing for growth, the cost of running a hall at 20 percent load during the two years before the GPU pods arrive. On the 2 MW example this article carries end to end, the difference between a competent loss calculation and a nameplate-based guess is $96,000 a year — roughly the price of the UPS hardware every decade, paid to the utility for nothing.

This article shows how to turn any UPS datasheet into a loss model, how to read the model at your real load, how to convert it into annual kWh and dollars including the cooling bill for the heat, and how to verify it in the field without fooling yourself with meter error. It pairs with our UPS battery sizing guide, which handles the discharge side; this one handles the other 8,760 hours.

Full-load efficiency is the least useful number on the datasheet

Laboratory benchmarking by LBNL and EPRI found that UPS efficiency varies by 7 to 12 percentage points between part load and full load, stays relatively flat between 50 and 100 percent, drops steadily below 50 percent, and falls off a cliff below 20 percent [1]. Legacy installations measured in the field have been found operating at 60 to 80 percent efficiency [4].

Now the uncomfortable part: data center UPSs almost never operate near full load. A 2N arrangement splits the load across two independent systems, leaving each around 40 percent; a 2N+1 block can sit near 33 percent [4]. Nonlinear server power supplies knock another 1 to 2 points off [4]. So the datasheet's headline number describes a condition your UPS visits a few hours a year, while the loads you actually run — the ones the redundancy design forces — sit on the steep part of the curve.

A published design study makes the consequence concrete: in a 1,500 kW design built with 2N static UPS, efficiency fell from 90.9 percent at full IT load to 85.1 percent at half load and 76.2 percent at quarter load, dragging facility PUE from 1.27 to 1.48 and then 1.91 [11]. The UPS did not get worse. It just moved left on its own curve.

The three-term loss model

Schneider Electric's Data Center Science Center measured commercial UPS units across the load range, fitted the results, and published the model the industry has used since: total loss, as a percentage of rated capacity, follows a second-order polynomial in load fraction x [2]:

L(x) = a·x² + b·x + c   (all terms in percent of rated kW; x = load ÷ rating, 0 to 1)

• c — no-load loss. Fans, control electronics, power supplies for the logic boards, capacitor leakage, and in transformer-based units the iron of the output transformer. Constant from 0 to 100 percent load. The white paper notes this term alone can exceed 40 percent of a UPS's total losses [2].
• b·x — proportional loss. IGBT switching losses and semiconductor conduction losses that scale roughly with current [2].
• a·x² — square-law loss. I²R heating in busbars, cables, filters and transformer windings; invisible at part load, dominant at full load [2].

Once you have a, b, c you can compute the loss at any load, split any annual energy bill into "paid for being energized" and "paid per kWh", and compare two quotes on the load points that matter instead of the one the marketing department chose.

Fit the coefficients from a datasheet in ten minutes

You usually do not need the white paper's curve, because most large-UPS datasheets already publish the losses under a different name: heat rejection, in kW or BTU/hr, at 25, 50, 75 and 100 percent load. Heat rejection is the loss. Here is the 1,000 kW column of a transformerless double-conversion datasheet [3]:

Load	Stated efficiency	Heat rejection	Loss as % of rating
25% (250 kW)	95.5%	11.8 kW	1.18%
50% (500 kW)	96.4%	18.7 kW	1.87%
75% (750 kW)	96.5%	27.2 kW	2.72%
100% (1,000 kW)	96.2%	39.5 kW	3.95%

Table 1 — A 1,000 kW transformerless VFI unit, double-conversion mode, unity-PF load, energy storage disconnected [3]. Loss percentages are heat rejection divided by rating.

A least-squares quadratic fit through those four points gives a = 2.16, b = 0.96, c = 0.81 (percent of rating), with R² = 0.9995 and a worst residual of 0.03 percentage points. For comparison, the coefficients Schneider measured on a transformer-based double-conversion unit are a = 3.46, b = 0.96, c = 3.23 [2]. Read those side by side:

• The proportional term is nearly identical (0.96 vs 0.96). Power electronics is power electronics.
• The square-law term differs modestly (2.16 vs 3.46).
• The no-load term differs by a factor of four (0.81 vs 3.23 percent of rating — 8 kW vs 32 kW on a 1,000 kW machine).

That last line is the entire efficiency story of the last fifteen years. The industry did not make switching losses disappear; it took the output transformer — and the better part of 24 kW of iron and thermal loss — out of the power path [10].

Read the curve at your actual load

With the coefficients in hand, evaluate the model at the loads your architecture actually produces — not at 100 percent:

Load on a 1,000 kW unit	Modern transformerless [3]		2010-era transformer-based [2]
	Loss, kW	Efficiency	Loss, kW	Efficiency
5% (50 kW)	8.7	85.2%	32.9	60.3%
10% (100 kW)	9.3	91.5%	33.6	74.8%
25% (250 kW)	11.9	95.5%	36.9	87.1%
40% (400 kW)	15.4	96.3%	41.7	90.6%
50% (500 kW)	18.4	96.5%	45.8	91.6%
67% (667 kW)	24.2	96.5%	54.1	92.5%
75% (750 kW)	27.5	96.5%	59.0	92.7%
100% (1,000 kW)	39.4	96.2%	76.5	92.9%

Table 2 — Loss and efficiency versus load. The modern curve is flat from 40 to 75 percent; the legacy curve keeps sliding. Below 10 percent load, both units burn most of their loss doing nothing.

Three things to take from the table:

• The flat top is wide. From 40 to 75 percent the modern unit varies by 0.2 points. Redundancy that lands you anywhere in that band costs almost nothing extra — which is not true of the legacy unit, where the same redundancy spread costs 4 points.
• Below 10 percent, efficiency is a fiction. At 5 percent load the modern unit still burns 8.7 kW — 93 percent of it the constant term. A UPS idling through a phased build-out is a 77,000 kWh/year space heater before a single rack ships.
• The minimum is not at full load. The curve peaks around 60 to 70 percent. A 1,000 kW unit carrying 500 kW (18.35 kW of loss) actually beats a 500 kW unit carrying the same 500 kW at nameplate (19.70 kW), and a 1,250 kW unit at 40 percent (19.30 kW) sits between them. The sweet spot for a modern flat-curve unit is roughly 40 to 70 percent of rating — which is, conveniently, where redundancy puts you anyway.

From a loss curve to an annual energy bill

The model converts to energy almost trivially, because each term integrates differently over a year:

• No-load: c × rating × 8,760 h — independent of load.
• Proportional: b × rating × Σx·Δt — proportional to annual kWh delivered.
• Square-law: needs the load profile's second moment; small unless you run hot.

Take the reference case: 2,000 kW of IT load, 2N, four 1,000 kW modules, each carrying 500 kW — the 40-to-50 percent condition that 2N produces [4]. At $0.10/kWh, 8,760 hours:

Architecture	Loss per module	Total loss	Share of IT energy	kWh/yr	Cost/yr
Modern 2N (4 × 50%)	18.4 kW	73.4 kW	3.7%	643,100	$64,300
Legacy 2N (4 × 50%)	45.8 kW	183.1 kW	9.2%	1,603,900	$160,400

Table 3 — Same load, same redundancy, different era of power electronics. The difference is $96,000 a year, every year.

Now split each bill into its fixed and variable parts:

2N system	Total loss	Fixed (no-load)	Variable	Fixed share	Fixed cost/yr
Modern	73.4 kW	32.5 kW	40.9 kW	44%	$28,500
Legacy	183.1 kW	129.4 kW	53.7 kW	71%	$113,300

Table 4 — The no-load term is paid 8,760 hours a year. On the legacy system it is a $113,000 annual subscription for the privilege of being plugged in.

This is the number oversizing debates miss. Adding modules "for future growth" adds a constant term per module, forever. On a modern flat curve the penalty is small (Table 6 below); on a legacy curve it is brutal — six modules instead of four costs $46,600 more per year, because you just bought two more $18,700-per-year subscriptions.

The heat has to go somewhere: the cooling adder

Every lost kilowatt becomes heat in the electrical room, and the cooling plant pays to move it out. Schneider's design guide uses 0.4 kW of cooling energy per kW of UPS heat [2]; a modern plant with good waterside economization can approach 0.3. On the 2 MW reference case:

2N system	UPS loss	Cooling energy (0.29 kW/kW)	Energy cost/yr	Cooling cost/yr	Total/yr
Modern	73.4 kW	21.0 kW	$64,300	$18,400	$82,700
Legacy	183.1 kW	52.3 kW	$160,400	$45,800	$206,200

Table 5 — Cooling adds roughly 25 to 30 percent to the bill. This is also why UPS losses pollute PUE: they sit inside the denominator as facility load.

One percentage point of efficiency at 2 MW is 20 kW — $17,500 a year in energy alone, $22,500 once the heat is cooled. That is the exchange rate to use whenever someone offers you a point of efficiency for money.

Redundancy arithmetic: when N+1 versus 2N is an energy decision

The redundancy choice sets the load point of every module, so it interacts with the loss curve. Same 2,000 kW, three ways:

Arrangement	Load per module	Modern loss	Modern $/yr	Legacy loss	Legacy $/yr
2N — 4 × 1,000 kW	50%	73.4 kW	$64,300	183.1 kW	$160,400
N+1 — 3 × 1,000 kW	67%	72.5 kW	$63,500	162.3 kW	$142,200
2N+2 — 6 × 1,000 kW	33%	82.4 kW	$72,200	236.3 kW	$207,000

Table 6 — N+1 versus 2N, then over-provisioning. On the modern flat curve, 2N costs $800 a year over N+1 and 2N+2 costs $7,900. On the legacy curve the same availability ladder costs $18,200 and $46,600.

Read that as a design rule, not an argument: on a flat-curve UPS, buy availability without an energy tax; on a steep-curve UPS, every redundant module you energize is a standing cost. Modular systems blur the line with sleep modes that de-energize surplus modules at light load, which is one more reason the light-load rows of the datasheet — not the full-load row — decide the purchase [1].

What the datasheet does not say

Five adjustments sit between the fitted curve and a defensible energy budget:

• kVA is not kW. Loss follows real power. A "1,000 kVA" legacy unit at 0.9 PF is a 900 kW unit; a modern unit rated at PF 1.0 is 1,000 kW. Two quotes at "the same size" can be 11 percent apart in real capacity — and if you read efficiency at 90 percent of kVA thinking it is 90 percent load, you are on the wrong row.
• Nonlinear loads cost 1 to 2 points. The curve above was measured on linear load. Server PSUs push input current in narrow peaks, adding rectifier stress; budget the derating explicitly [4]. Our power quality guide covers what that current does upstream.
• Battery terms are real but small — except after an outage. Float charging a 2 MW block draws on the order of 0.5 to 1.5 kW continuously (call it $500 to $1,300 a year). The big term is recharge: a 10-minute full-load ride-through consumes 333 kWh, and recharge draws 1.10 to 1.20 times that from the input once charger efficiency is counted — 407 to 444 kWh, averaging about 71 kW over a six-hour recharge. That is a 3 percent bump on the block's daily energy, and it is why facilities with chronically dirty utility feeds see UPS losses their model does not predict. Sizing the storage itself is covered in our battery sizing article.
• Eco mode changes the curve, not just the number. Static-bypass modes quote 98 to 99 percent because the double conversion is switched out of the path; such schemes reduce UPS losses by around 70 percent in practice, and the model prices the 2 MW block at $17,700 a year versus $64,300. The price is a different failure envelope: the load now rides on the bypass path, and the transfer back into double conversion must complete inside the 10 to 50 ms hold-up window of the IT power supplies [12]. It is an availability decision first — see our ATS vs STS comparison for what sits on that path.
• Topology is the biggest lever of all. Transformer-based double conversion runs 90 to 92 percent at full load; transformerless 95 to 96 percent; both reach about 99 percent in eco mode [10]. Three terms, and only one of them — the constant — separates the topologies.

Weighted efficiency: why two "96 percent" claims are not comparable

Because part-load behavior dominates real energy use, the standards bodies stopped certifying full-load efficiency and moved to weighted averages. Three frameworks coexist:

• US DOE test procedure (10 CFR 430, Appendix Y1) — measures at 100/75/50/25 percent in descending order, 15-minute energy accumulation at 1 Hz or faster, steady-state confirmed within 1 percent, and weights the result 0 / 0.3 / 0.4 / 0.3 across 25/50/75/100 for units above 1,500 W [5]. The 25 percent point carries zero weight.
• ENERGY STAR UPS v2.0 — the same method, adding a no-load measurement and a 125 percent stabilization allowance [6].
• EU Code of Conduct for UPS, v2.0 — weighted efficiency per EN 62040-3, with two tiers, "standard" and "Elite"; the Elite tier is now the expectation for new data centers signing the Data Centre CoC [8][9][13]. For VFI units above 200 kW the thresholds are 92.0 percent standard and 93.5 percent Elite.

Apply the different weightings to our two fitted units and you get different numbers for the same hardware:

Weighting	Weights (25/50/75/100)	Modern unit	Legacy unit
DOE Appendix Y1 [5]	0 / 0.3 / 0.4 / 0.3	96.38%	92.44%
Equal quarters	0.25 each	96.15%	91.09%

Table 7 — Same units, two label numbers, 0.2 to 1.4 points apart. A vendor holding up a "96.4%" certificate and a buyer assuming a plain average are not discussing the same quantity.

The practical rule: never compare weighted numbers across frameworks, and never compare a weighted number to a single-load number. Ask which weights produced the figure, then ask for the four underlying load points so you can rebuild the curve yourself.

Measuring it in the field without lying to yourself

The DOE method translates directly to a commissioning test: measure input and output energy simultaneously over at least 15 minutes at 1 Hz or faster, confirm steady state by repeating after 10 minutes and checking the two efficiencies agree within 1 percent, then step through 100, 75, 50 and 25 percent with a load bank [5][6]. The subtlety is not the procedure — it is the meter.

Efficiency is a small difference between two large numbers. At 96 percent efficiency the loss is 4 percent of throughput, so a measurement error on either meter is magnified roughly 25-fold in the loss. Work the arithmetic:

Per-meter accuracy	Uncertainty on a 96.2% efficiency reading	Can you resolve a 1-point claim?
±0.5% of reading (typical revenue meter)	±0.97 points	No
±0.2% (class 0.2S)	±0.39 points	No
±0.1%	±0.19 points	Borderline
±0.05%	±0.10 points	Yes, barely

Table 8 — Measurement uncertainty at η = 96.2%, both meters erring in the same direction. A field crew with building-class meters cannot verify a "96 vs 97" marketing claim to better than ±1 point.

Three consequences:

• Demand the type-test report, not a re-measurement. Credible large-UPS loss curves are laboratory measurements, in Schneider's case by an independent third party [2]. Reproducing a 0.1-point claim in an electrical room requires 0.05-percent-class metering and clean waveforms — lab equipment.
• Do not bill from the front-panel readout. Panel meters are indicative, not revenue grade; if internal metering feeds tenant billing or an ISO 50001 baseline, validate it against a calibrated external analyzer at one load point before trusting the rest.
• Watch phase-angle error. At data center power factors, current-transformer phase displacement moves real-power readings more than ratio error does. This is the same metering-chain discipline our DCIM point-schedule guide applies to PUE boundaries.

Worked example, end to end

Put the whole model to work on a replacement decision. A 2 MW hall runs legacy transformer-based 2N. The loss model says the UPS layer costs $206,000 a year including cooling — $2.06 million over ten years, more than the switchgear feeding it. A modern transformerless 2N replacement costs $82,700 a year including cooling.

• Annual saving: $123,500.
• Ten years at $0.10/kWh — the same reference assumptions Schneider's white paper uses ($0.10, 8,760 hours, ten years, 0.4 kW of cooling per kW of heat [2]) — come to $1.24 million.
• Against installed hardware in the $500,000 to $700,000 range, energy alone pays the project back in roughly four to five years — before counting floor space, weight, and the smaller upstream switchgear that a 95-96 percent topology allows [10].

Then stack the options in order of risk appetite: N+1 instead of 2N saves another $1,200 a year (take it only if the availability model accepts it); 2N+2 for growth headroom costs $7,900 a year, which is the honest price of that headroom on a flat-curve unit; eco mode saves $46,600 a year and buys a different failure envelope that must be validated against PSU ride-through and utility quality first [12]. Every one of those numbers came from three coefficients fitted to four datasheet rows.

Ten mistakes that corrupt a UPS loss calculation

1. Comparing nameplate full-load efficiency. The units live at 33 to 50 percent load [4]. Price the curve, not the point.
2. Sizing in kVA and reading kW curves. Loss follows real power; PF 0.9 and PF 1.0 ratings are different machines.
3. Forgetting that 2N halves the load. Model the architecture's load point first, the hardware second.
4. Treating redundancy as energy-free. Each energized module adds a constant term for 8,760 hours a year.
5. Dropping the cooling adder. 0.3 to 0.4 kW of cooling per kW of loss, inside your PUE denominator [2].
6. Billing or benchmarking from the UPS front panel. Panel accuracy cannot resolve the differences you are deciding on.
7. Reporting two decimals from ±0.5 percent meters. The loss is a difference of large numbers; ±1 point of uncertainty is the floor for typical field metering.
8. Ignoring the nonlinear-load penalty. Subtract 1 to 2 points for real IT loads [4].
9. Booking eco-mode savings without the risk review. The transfer into double conversion must fit inside the PSU ride-through window [12].
10. Leaving battery energy out of the model. Float is small; recharge after every outage averages about 71 kW for six hours on a 2 MW block.

Eight questions for the UPS supplier

1. Efficiency and heat rejection at 25/50/75/100 percent, double-conversion mode, at PF 1.0 and PF 0.9, with energy storage connected and disconnected — four numbers minimum.
2. No-load loss in kW, measured with the inverter running and the output open.
3. Which standard produced the headline efficiency — which IEC 62040-3 edition, which weighting, which load points [7]?
4. A third-party type-test report with the raw input/output measurements, not a summary slide [2].
5. Efficiency and transfer behavior in eco/ESS mode: transfer time both directions, and behavior on input disturbance.
6. Performance derating under nonlinear load and at 35 to 40 °C ambient, including auxiliary consumption at full fan speed.
7. Battery recharge profile: charger efficiency, recharge time from full discharge, and float power.
8. Expected auxiliary and filter losses after ten years — and what the efficiency curve looks like then, not just on day one.

The takeaway

UPS energy loss is not a percentage; it is a three-term curve you can fit from any competent datasheet in ten minutes. The constant term is a subscription billed 8,760 hours a year, and it is where old and new topologies actually differ. Read the curve at the load your redundancy architecture produces, add 25 to 30 percent for cooling, respect what your meters can and cannot prove, and the efficiency line item stops being a brochure argument and becomes an engineering number you can put in a financial model. If you are weighing UPS options for a build or a refresh, send us the load profile and the one-line — we will run the loss model against the shortlist and show you the ten-year number for each.

References

1. LBNL / EPRI. High Performance Buildings: Data Centers — Uninterruptible Power Supplies, 2005. Efficiency versus load (7–12 points of variation, steep decline below 50%, cliff below 20%), efficiency definition per IEC, and manufacturer/field test methods. datacenters.lbl.gov (PDF)
2. Schneider Electric, Data Center Science Center. Making Large UPS Systems More Efficient, White Paper 108, rev. 3. Three-term loss model (no-load, proportional, square-law), no-load share above 40%, TÜV-measured coefficients, cooling at 0.4 kW per kW of heat, and the $0.10/kWh / 8,760 h / 10-year cost basis. on365.co.uk (PDF)
3. Power Solutions. TLE Series UPS 750/1000 kW Technical Datasheet — system efficiency and heat rejection at 25/50/75/100% in double-conversion and eBoost modes. power-solutions.com (PDF)
4. NIH. Sustainable Data Center Design Guide — legacy UPS operational efficiency of 60–80%, typical load points of 2N (~40%) and 2N+1 (~33%), the 1–2 point nonlinear-load penalty, and efficiency ranges by topology and load. orf.od.nih.gov (PDF)
5. US DOE. 10 CFR 430, Appendix Y1 to Subpart B of Part 430 — UPS test procedure: 100/75/50/25% reference loads, 15-minute energy accumulation at ≥1 Hz, steady-state within 1%, load weightings 0/0.3/0.4/0.3 above 1,500 W. ecfr.gov
6. US EPA. ENERGY STAR Program Requirements for UPS, Final Version 2.0 — test method including the 0%-load measurement and stabilization requirements. energystar.gov (PDF)
7. IEC 62040-3, Uninterruptible power systems (UPS) — Part 3: Method of specifying performance and test requirements. webstore.iec.ch
8. European Commission JRC. Code of Conduct on Energy Efficiency of AC Uninterruptible Power Systems (UPS), Version 2.0, 2021 — weighted efficiency per EN 62040-3, standard and Elite tiers; for VFI above 200 kW, 92.0% standard / 93.5% Elite. e3p.jrc.ec.europa.eu (PDF)
9. Energy XPRT / CEMEP. Launch of the Code of Conduct for the energy efficiency of UPS 2021–23 — the two-tier structure and the Data Centre CoC requirement to install Elite-class UPS in new facilities. energy-xprt.com
10. ABB (C. Kufahl). UPS design: Transformer-free vs transformer-based — double-conversion efficiency of 90–92% (transformer-based) vs 95–96% (transformer-free), eco mode near 99% for both, and the downstream design consequences. power-backup.ro (PDF)
11. Brunel University. UPS System: How Can Future Technology and Topology Improve the Energy Efficiency in Data Centers? — 2N static UPS design: efficiency 90.9% → 76.2% and PUE 1.27 → 1.91 as IT load falls to 25%. bura.brunel.ac.uk (PDF)
12. Schneider Electric. UPS Eco Mode Can Deliver Big Savings — The Green Grid eco-mode evaluation: PSU ride-through typically 10–50 ms, transfer-time matching, and utility quality as the gating conditions. blog.se.com
13. Riello UPS. General Catalogue — signatory statement that the entire online range complies with the Elite classification of the UPS Code of Conduct 2021–23. riello-ups.co.uk (PDF)