Titanium Efficiency as a TCO Argument

The Three Cost Lines Efficiency Moves

An efficiency tier affects three separate budgets, and only one of them appears on the electricity invoice. The first is the direct energy cost of the loss inside the supply. The second is the cooling energy required to remove that loss from the building, which is the first cost multiplied by your facility's power usage effectiveness. The third is thermal headroom — the capacity you free inside the chassis, which shows up as density, acoustics and reliability rather than as a line item, and which is why high-density designs converge on Titanium even before anyone runs the numbers.

Build the model in that order. Each step is a multiplication, each assumption is stated, and the whole thing fits on one page.

Step 1 — Turn Efficiency Points into Watts

Start from the gate points rather than the headline. The 80 PLUS internal redundant rating tests 10%, 50% and 100% load, and our platforms sit on the Titanium floors of 90%, 96% and 91%.

Rating	10% load	50% load	100% load	Input to deliver 5 kW at 50%	Loss
Gold	82%	92%	89%	5,435 W	435 W
Platinum	90%	94%	91%	5,319 W	319 W
Titanium	90%	96%	91%	5,208 W	208 W

A note on how to read this table, because two different figures circulate for the same comparison. Elsewhere on this site we quote the Gold half-load case as roughly 430 W per 10 kW node per hour. Look at the table and you can see exactly what that number is: it is the whole loss the Gold unit carries at 5 kW delivered, not the difference between tiers. The difference between Gold and Titanium at the same operating point is the roughly 230 W you actually stop paying for; against Platinum it is about 110 W. Both framings come from the same 92% and 96% figures, and the payback conclusion holds under either, so this article runs the conservative delta-based case and states the headline case alongside it rather than letting the two look like different claims.

The 10% column deserves attention for a different reason. Titanium is the only tier with a hard requirement at 10% load, and 10% is exactly where a lightly loaded redundant array spends its idle hours. On a system sitting at a tenth of rated capacity, the gap between 82% and 90% is proportionally larger than the gap at any other point — a fact that matters enormously for storage arrays, networking gear and enterprise servers that are sized for peak and run well below it.

Step 2 — Heat Is Paid for Twice: the PUE Layer

Every watt lost inside a power supply leaves the chassis as heat, and the cooling plant has to remove it. That is why a watt of PSU loss costs more than a watt of delivered compute: facility energy scales with PUE, while IT energy does not.

Using the conservative 230 W delta at the half-load point, 8,760 hours of always-on operation and a flat tariff of $0.10 per kWh — state your own tariff, this is an illustration and not a quote — the numbers land like this:

Facility PUE	IT-side energy for the 230 W delta	Facility energy per node per year	Annual cost per node
1.10 (liquid-cooled, efficient)	2,015 kWh	2,216 kWh	~$222
1.20	2,015 kWh	2,418 kWh	~$242
1.40	2,015 kWh	2,821 kWh	~$282
1.60 (air-cooled, mixed estate)	2,015 kWh	3,224 kWh	~$322

The same model on the headline 430 W framing, with the cooling overhead expressed as a 10–30% adder on top of the direct energy cost, returns roughly $377 per node per year in electricity and about $415–$490 all-in. Whichever version you present, the structure is what matters: a per-node annual saving that can be divided into an array-level premium to produce a payback period.

Two more PUE consequences are worth putting in the deck. First, cooling capacity is capital, not just energy. A 230 W per-node reduction across 1,000 nodes is 230 kW of heat that the cooling plant never has to be sized for — and on a per-rack basis, roughly 1.8 kW freed in an eight-node rack. Second, the higher your PUE, the stronger the case, which is why the efficiency argument is most persuasive in exactly the air-cooled mixed estates that are least likely to have run the calculation.

Step 3 — Thermal Headroom Becomes Density

The freed thermal budget is the least quantified and often the most valuable line. In a 1U node where the power bay sits last in the airflow path, every watt of PSU loss raises the inlet temperature of everything downstream. Removing loss widens the margin at the hottest point in the chassis, which converts into three things: higher sustained boost clocks, slower fans and lower acoustic output, or more compute in the same rack unit — the choice is yours, but only if the margin exists.

That is why the efficiency decision and the density decision are the same decision. A 2400 W CRPS unit in the 73.5 × 40 × 185 mm envelope runs in the 75 W/in³ class, and losses at 96% are small enough to hold inside a single 40 mm fan's acoustic envelope. At 92%, the same power pushes materially more heat into the same 1U of exhaust, and the fan has to work harder to move it. The efficiency premium buys back airflow you can spend on silicon.

Step 4 — Reliability Is the Line Finance Never Models

Component temperature is the dominant stressor behind power supply field failures, and the rule of thumb for electrolytic capacitors — the components that most often set a supply's service life — is that life roughly halves for every 10°C increase in operating temperature. A cooler-running Titanium unit therefore does not merely cost less to power; it is on a longer failure curve than a Gold unit at the same output.

Our platforms are engineered to a 250,000-hour MTBF at 40°C per Telcordia SR-332 target, with continuous output from −5°C to +55°C ambient. Run the prediction over a fleet and the absolute numbers look alarming in the way all MTBF arithmetic does: 1,000 nodes with six modules each is 6,000 modules, and the predicted rate works out to roughly 210 module replacements a year. Real fleets usually do better, which is precisely why the useful move is not to quote the absolute figure but to model the marginal difference — put your loaded cost per replacement event against the failure-rate difference between a cooler unit and a warmer one. That delta is the smallest number in this article and frequently the largest in the payback.

Building the Payback Calculation

Payback is the array-level Titanium premium divided by the annual saving per node. Titanium units carry a price premium over Gold of typically low double-digit percentages at the 800 W–2400 W level, narrowing at volume, and the saving accrues per node while the premium is quoted per unit — so run the division at array level, not module level, or you will overstate the payback by the number of modules in the array.

Array-level premium	Payback at $420 saved per node per year	Payback at $490 saved per node per year
$400	~1.0 year	~0.8 year
$800	~1.9 years	~1.6 years
$1,200	~2.9 years	~2.4 years
$1,600	~3.8 years	~3.3 years

Two things to say out loud when you present this. First, the premium is quoted per unit and the saving per node, so a six-module array divides its premium by one node's saving. Second, the comparison horizon is your refresh interval, not the supply's service life — if you replace the platform in three years, a payback beyond three years does not happen, and you should discount the tail rather than present it.

Modelling Mistakes That Break the Answer

• Using the 100% load point. At full load the Gold-to-Titanium gap is 89% against 91% — two points. At 50% it is four points, and at 10% it is eight. Servers do not live at 100%.
• Ignoring the 10% floor. Redundant arrays split the load across every installed module, so a lightly loaded node sits lower on the curve than a single-unit calculation suggests. This is where the tier difference is largest.
• Assuming a single-unit load factor. In an N+1 array the load is divided by the number of modules, which changes both the operating point and the efficiency you should read from the curve.
• Leaving cooling out. Direct electricity is only the first of the two energy lines; the PUE multiplier is not optional and it rises with the inefficiency you are trying to cost.
• Using an average tariff where a marginal one applies. The kilowatt-hours you avoid are marginal consumption, and marginal pricing is often higher than the blended rate on the invoice.
• Comparing at a horizon longer than the refresh. An eight-year saving on a platform you replace in three is a modelling artefact, not a business case.
• Valuing the freed thermal budget at zero. If the margin is real, it has a use — density, acoustics or reliability. Assign it a value or state explicitly that you are not.

The One-Page Fleet Model

If the analysis has to fit on one page, these are the eight inputs it needs:

• Delivered power per node at the average operating point, in kW — not the peak, and not the nameplate.
• Average load factor of the power array as a percentage of installed capacity.
• Operating hours per year, which for always-on equipment is 8,760.
• Your marginal electricity price per kWh, stated as an assumption.
• Facility PUE, taken from measured data rather than a design target.
• The tier difference in watts at that operating point, read from the 10/50/100% table.
• The array-level Titanium premium from your actual quotation, not a list price.
• The refresh interval, which caps the horizon over which savings may be counted.

Feed those eight numbers in and the payback falls out in one division, with a sensitivity table around the two inputs you are least sure of — usually the load factor and the PUE. That is the version of the argument that survives a finance review, because every number in it is either measured or labelled as an assumption.

If you would rather not build it from scratch, send us the load profile and your tariff and PUE, and we will return the comparison against the Titanium gate points with the arithmetic shown. Our production 550 W–2400 W CRPS series carries those efficiencies today, with samples in 2–4 weeks and volume in 6–10 weeks including burn-in; the underlying design detail is on the 80 PLUS Titanium capability page.

Related Reading