How much can better data-center efficiency save?

Use this guide to: evaluate an energy or facility decision using explicit boundaries and assumptions.

Constructed worked example. The loads, PUE values and energy rates below are teaching assumptions, not operating records, supplier quotes or a project forecast.

The question: what does a lower PUE actually save?

Two facility options serve the same workload. Option A has an annual PUE of 1.30; option B has an annual PUE of 1.60. Before calling A the better project, isolate the energy effect. Keep the IT load, operating period and energy rate equal.

PUE is total facility energy divided by IT equipment energy. A lower value means less facility overhead for the same IT energy within a comparable measurement boundary. It does not establish how efficiently a model produces useful work.

Hold the workload constant

Average IT demand is 800 kW. Annual IT energy is 1,000 × 0.80 × 8,760 = 7,008,000 kWh. Multiplying that energy by annual PUE gives the assumed facility energy. This uses an annual energy ratio, not a claim that instantaneous demand is constant.

Follow the energy, then the charge

The difference is 2,102,400 kWh and $126,144 per year. That is the modeled benefit of changing PUE alone under these assumptions. It is not an estimate of net profit or total operating savings.

Change one assumption at a time

If both energy rates rise to $0.09/kWh, the same energy difference is worth $189,216. If average IT electrical load instead falls to 40%, with the original rates and the same assumed PUE values, the difference falls to $63,072. In practice PUE can change with loading and conditions; the fixed-PUE sensitivity is a controlled calculation, not an operating prediction.

Now consider a different comparison: A costs $0.08/kWh while B remains at $0.06/kWh. A's energy charge becomes $728,832, above B's $672,768. The lower PUE still saves energy, but the rate difference reverses the energy-charge ranking.

What the decision still needs

Ask whether the measurement periods and boundaries match, whether the facilities can support the same workload, and whether capacity is actually available. Then add demand charges, fixed fees, capital, maintenance, availability requirements and delivery dates. This example excludes those items.

For a real comparison, replace every assumed input with documented evidence. Keep a record of the source and test the downside. A useful next question is: how much extra annual cost could A justify before its modeled energy advantage disappears?

Recreate the scenarios in the energy calculator ↗

A real case: Google and DeepMind cooling optimization

In a historical company report, Google said machine learning reduced the energy used for cooling by up to 40%. The same report described a 15% reduction in overall PUE overhead after accounting for other inefficiencies. These are company-reported results from that deployment, not independently verified results for the scenarios above.

The measurement boundary matters. Cooling energy is one part of facility consumption. PUE overhead is the energy above the IT equipment requirement. Neither percentage can simply be applied to the entire electricity bill of another site.

AI Keeper's interpretation: the useful lesson is to measure the subsystem being improved and connect it to total facility energy. Before transferring a result, establish the baseline, operating conditions, control limits and verification method. Our A/B calculation isolates PUE assumptions; it does not reproduce Google's implementation or predict its savings.

Read Google's original case report ↗