One of the most repeated pieces of efficiency advice in the data centre sector is to run hotter. Inlet temperatures have crept up across most operators over the past decade, on the rationale that warmer air saves cooling energy. The advice is widely given and rarely qualified. Our most recent research, published in IEEE Transactions on Sustainable Computing, asks whether the advice holds. The short answer is: sometimes. The longer answer matters.

What the paper actually tests

The paper is titled Is Hot IT a False Economy? An Analysis of Server and Data Center Energy Efficiency as Temperatures Rise. The methodology pairs cross-generational server measurement (how different CPUs and server models respond to temperature change) with an idealised data centre simulation, then extends the simulation across multiple geographic regions to capture how location affects the answer.

What the experiment asks: is there always a net energy and carbon benefit to raising the temperature in the data centre. The answer, on measured data, is no.

What changes with temperature, beyond cooling

The classic argument for higher inlet temperatures is that fans run less and chillers do less work. Both are true. What the same argument omits is that the IT itself responds to temperature in ways that work against the saving. Three findings stand out:

  • Leakage current increases with temperature. As silicon heats, transistors leak more current at idle. This is a physical property of CMOS, not a cooling problem, and it cannot be fixed by pressurising the airflow.
  • Inlet temperatures below 30°C don't affect performance, only efficiency. Above that point, performance starts to degrade. The relationship is not linear and not symmetric across generations.
  • Idle and max utilisation are hit hardest. The sweet spot for tolerance is in the middle of the utilisation curve. Servers running at very low or very high utilisation respond worst to rising inlet temperatures.

The cross-generational point also matters commercially. Earlier server generations are more affected by temperature rises than later ones; advice that suits a fleet of recent hardware can damage the case for keeping older, refurbished kit in service.

The order that should drive the decision

The most impactful single takeaway from the paper is a re-ordering of how the design and operational decisions should be made.

The conventional order is building → cooling → hardware → temperature. Build the facility, install the cooling, populate with hardware, set the temperature.

The order the paper argues for is location → hardware → cooling → temperature. Start with where the facility sits, because the local climate and grid mix set the boundary conditions. Choose hardware that matches both the workload and the location's thermal profile. Specify cooling against that combination. Set the temperature last, against the specific tolerance of the installed system.

That re-ordering treats the data centre as a system of systems rather than a stack of independently-optimised components. Most existing facilities have not been designed that way; most blanket advice on temperature inherits the assumption that they have.

What this means in practice

For an operator currently raising inlet temperatures as an efficiency lever, the implication is that the decision needs to be tested against the specific estate. Sometimes the energy saved on cooling is offset (and occasionally outweighed) by the efficiency loss at the IT layer. Sometimes it isn't. The measurement to do that test is the hardware-side performance and energy data the paper is built on.

For an operator designing a new facility, the implication is more structural. Site selection matters more than the cooling specification, because the location drives both the local climate envelope and the energy and water profile of the grid the facility will pull from. Most current site-selection criteria (connectivity, power availability, planning permission) don't fully account for the operational efficiency consequences a few years downstream.

Where the paper sits

Is Hot IT a False Economy? is the second peer-reviewed paper from the Interact team in IEEE Transactions on Sustainable Computing, a journal in the top 5% of academic publications globally. The first, in 2020, established the slowdown in Moore's Law and the carbon arithmetic for refresh decisions involving refurbished hardware. The two papers, taken together, are the methodological backbone of the dataset our customer reports draw on.

The full paper is available on ResearchGate.