A Titanium PSU Label Does Not Set the GPU Server Input Power

Read the three quantities behind the offer

An offer that lists a 3,000 W supply, a Titanium label and 96% efficiency has not supplied one interchangeable power number. The wattage identifies rated output. The percentage needs both a load point and input conditions. Neither tells the buyer what the complete server continuously draws.

Let Prated be the PSU's rated DC output and Pout its delivered DC output at the condition being considered. Its output-load fraction is Pout / Prated; multiply by 100 to express that fraction as a percentage. Conversion efficiency has a different denominator: eta = Pout / Pin, where Pin is real input power at the stated converter boundary. Use eta as a decimal in the arithmetic. Rearranging gives Pin = Pout / eta and conversion loss = Pin - Pout.

A 50% output-load point therefore describes how much of a supply's output rating is being used. It does not mean 50% efficiency, 50% GPU utilization or half the rack's available power. Those quantities describe different objects. Here, useful DC output means power delivered by the converter; it does not mean useful AI work completed.

References: [PE3]

Keep the certification category and input conditions attached

CLEAResult's 80 PLUS program details distinguish desktop, industrial and redundant data-center PSU categories, including different AC and DC input categories. A badge identifies a PSU certification class within that framework. It does not certify the input consumption of every server containing a supply with that badge.

The linked generalized test protocol is revision 6.7.3, dated July 22, 2025. It identifies input voltage, frequency and output loading among the variables affecting efficiency. A comparison should retain the source's category, input conditions and output-load point rather than transferring a percentage between unlike entries.

For an offered server, ask the supplier to identify the actual supply option and the applicable supporting data. A generic Titanium summary cannot establish that an unspecified purchased configuration has the same efficiency under every operating condition.

References: [PE1][PE3]

Calculate the named Dell half-load point

Dell's PowerEdge XE9685L Installation and Service Manual, checked October 20, 2026, provides a specific example in Table 2, PSU Efficiency Level. Its 3,000 W output entry, with input labeled 240AC, lists Titanium efficiency of 96% at 50% output load and 94% at 100%. The two stated efficiencies already show why 96% cannot describe every load.

At the listed half-load point, Pout = 0.50 × 3,000 W = 1,500 W DC. Applying the listed efficiency gives Pin = 1,500 W / 0.96 = 1,562.5 W AC. The corresponding conversion loss is 1,562.5 W - 1,500 W = 62.5 W.

This is arithmetic from a tabulated PSU point, not measured consumption of an XE9685L server or a prediction of its workload demand. The calculation inherits that point's input and efficiency boundary. It supplies no efficiency for an intermediate output load.

The loss denominator also matters: 4% of 1,562.5 W input is 62.5 W. Adding 4% to 1,500 W output instead gives 1,560 W, which does not satisfy the stated 96% ratio.

References: [PE2]

The same DC demand can put two supplies at 25% load

Now consider an entirely stipulated arrangement in an unnamed system: a total useful DC demand of 1,500 W is divided equally between two supplies, each rated for 3,000 W output. This is an invented comparison, not the documented configuration or load-sharing behavior of the Dell server.

Each supply delivers 1,500 W / 2 = 750 W. Its output-load fraction is 750 W / 3,000 W = 0.25, or 25%. The total demand has stayed at 1,500 W, but neither supply is at the 50% output-load point used in the previous calculation.

If their efficiencies at the stipulated condition are eta1 and eta2, combined converter input would be 750 W / eta1 + 750 W / eta2. The cited 50% cell establishes neither value. It therefore cannot justify assigning 1,562.5 W input to this two-supply arrangement. Interpolating an efficiency at 25%, or assuming both efficiencies equal 96%, would introduce unsupported information.

For an actual offer, the operating supply population and supported distribution of output demand are necessary inputs. Installed supply count alone does not establish which supplies deliver output in a particular state. This example explains the missing evidence; it selects no redundancy arrangement, recommends no supply removal or disabling, and establishes no facility capacity.

Keep conversion efficiency inside the complete-server boundary

GPU demand alone does not identify the DC demand of the whole server. CPUs, host memory, storage, networking and fans also belong in the actual equipment accounting. The supplied configuration and operating state determine what is included; a guessed allowance cannot turn GPU-only power into supported server input.

Preserve the efficiency report's treatment of auxiliary loads. The generalized protocol distinguishes integrated-fan treatment for single-output and multiple-output supplies. Consequently, a converter calculation must retain any stated exclusions; the complete-server input comparison still needs all power consumed within its declared equipment boundary. No fan treatment for the Dell entry is inferred here.

When the supplier provides complete-system input evidence, keep the exact hardware configuration, input conditions, operating state and included equipment with it. This lets the buyer compare an input observation with another input observation, or an explicitly labeled estimate with another estimate. PSU output ratings and GPU activity percentages cannot replace either.

References: [PE3]

Energy needs time at each input state

An efficiency point describes conversion under a stated condition. Energy adds time. For intervals with mean complete-system input Pin,i in kW and duration ti in hours, total input energy is the sum of Pin,i × ti, in kWh. If input varies within an interval, its mean must represent that interval rather than its highest reading.

For a separate fictional example, stipulate complete-system AC input averaging 1.6 kW for six hours and 0.8 kW for eighteen hours. These invented input values are independent of the Dell efficiency table. Daily energy is 1.6 × 6 + 0.8 × 18 = 24 kWh.

Multiplying the higher 1.6 kW value by all 24 hours gives 38.4 kWh. That describes a different, constant-input assumption. It does not describe the stipulated day. Likewise, a proposed annual figure needs a supported distribution of states and durations across the year; a peak input point or a favorable efficiency cell cannot supply those hours.

If an estimate starts from DC demand, each state's conversion needs the efficiency applicable to that state's per-supply loads and conditions. A simple average of efficiency percentages does not establish total input energy. Pricing additionally needs the actual charging basis and terms; these examples calculate no bill, savings or return on investment.

References: [PE4]

Resolve the missing quantity before comparing offers

The practical question is whether an offer supports the quantity its summary claims. A PSU output rating can identify the supplied component. A qualified efficiency point can support a converter calculation at that point. A complete-server input claim needs the included configuration and operating basis. An energy claim additionally needs time. Evidence for one level does not fill the missing inputs at the next.

An offer with a lower PSU wattage or a higher badge efficiency therefore does not, by those facts alone, establish lower server input or lower annual consumption. Nor does the hypothetical two-supply example establish that fewer operating supplies would be preferable. Ask the supplier to resolve the actual configuration and state evidence instead of ranking incomplete offers by the badge alone.

Use the related GPU power-cap article for the separate question of which components and policy a limit controls. Use the PUE article for facility energy relative to IT energy, and the workload-benchmark guide for energy associated with comparable accepted work. Conversion efficiency answers none of those questions by itself.

Carry the selected equipment into the physical installation scope

If the equipment comparison leads to an installation in authorized third-party colocation space, bring the selected server configuration and the customer's confirmed physical requirements to the installation discussion.

Leviathan's Colocation Infrastructure Installation service covers agreed customer-side physical IT work, including equipment mounting, physical IT cabling, labeling and defined handoff records. The linked service and project inquiry support that physical scope; server energy evaluation remains with the customer's appointed platform and facility reviewers.

For an inquiry, share the nonconfidential site location, equipment population, requested physical tasks, available plans and intended work window so the team can discuss scope, requirements and availability.