Your factory added AI compute, but how is cooling cost measured?

When cooling demand disappears into factory overhead, part of the running cost is lost. That makes expansion tricky to budget and claims of cooling savings harder to check.

What you’ll learn:

  • When cooling demand disappears into factory overhead, part of the running cost is lost.
  • A lower electricity bill after an AI rollout does not establish a cooling improvement.
  • Have the plant engineer and IT owner agree on the map, then identify what is directly metered, what is shared and what must be estimated.

Consider a factory adding AI servers beside a production line to run a vision-inspection system. The project team can see what the servers cost and how much electricity their power feed uses. Cooling is harder to price: It comes from plant equipment that was already there.

See also: Smart Industry's 2027 State of Initiative Survey is now open

When the project's cooling demand disappears into factory overhead, the team loses part of the running cost. That makes the next expansion harder to budget and any claim of cooling savings harder to check.

Before approving more servers, ask IT and facilities to explain how the first installation is cooled and which electricity use they can attribute to it.

Draw the boundary before buying a meter

In this hypothetical installation, liquid-cooled servers connect through a coolant distribution unit to a shared chilled-water plant. Some server heat may still leave through room air. The cooling boundary therefore needs to include the relevant liquid path and any additional room-cooling demand, rather than assuming the water connection captures everything.

Map the electrical feeds as well. Does the server meter already include a cooling unit's pumps or fans? Are those loads on a separate panel? Counting the same equipment twice can make a careful-looking cost estimate wrong.

Have the plant engineer and IT owner agree on the map, then identify what is directly metered, what is shared and what must be estimated. Finance should be able to follow those distinctions without needing to interpret a piping drawing.

Separate heat removed from electricity bought

At a liquid connection, flow and the difference between supply and return temperatures can be used to estimate heat transfer, with the appropriate properties of the actual fluid. Integrating that rate over time gives thermal energy removed. Meter selection, installation and measurement uncertainty matter, particularly when temperature differences are small.

That thermal energy is not the electricity consumed by the cooling plant. Removing a given amount of heat can require different electrical input as outdoor conditions, plant loading and control sequences change. A pump serving the AI loop may have a directly measurable electrical cost, while the servers' share of the central chiller remains an allocation.

See also: Give operators the authority to stop flawed AI workflows

Suppose the AI branch accounts for 20% of measured heat removal during an interval. Allocating 20% of the shared plant's electricity to it might be a useful accounting convention.

It does not prove that shutting down those servers would reduce plant electricity by 20%. Shared equipment may continue running or change its operating point.

Label the result accordingly. A useful estimate is more credible when its limits are visible.

Compare equivalent work

A lower electricity bill after an AI rollout does not establish a cooling improvement. Perhaps the weather changed, a production line ran fewer hours, or the inspection workload fell. The U.S. Department of Energy's guidance on energy information systems stresses the role of relevant operating conditions when evaluating performance.

For a factory inspection application, compare intervals with similar production and compute demand. Track model or software changes that could alter processing requirements. Keep inspection coverage, quality and response-time requirements in view: doing less useful work is not an efficiency gain.

A lower electricity bill after an AI rollout does not establish a cooling improvement. Perhaps the weather changed, a production line ran fewer hours, or the inspection workload fell.

A site might track cooling electricity per thousand completed inspections, provided it defines what counts as a completed inspection and keeps the allocation method consistent. Another application will need a different denominator. The metric should help the operations team explain a change, not simply produce a smaller number.

Podcast: Why system integration is critical to AI-ready manufacturing

Use synchronized records and retain enough detail to examine the events behind a weekly total. A commissioning baseline, captured before routine changes accumulate, gives the team a much stronger starting point than a reconstructed estimate months later.

Put IT and facilities in the same review

IT can report server workload without seeing the chiller. Facilities can report chiller operation without knowing that a model update changed the computing load. Finance receives the cost of both.

Assign one person to bring those views together. A short recurring review should ask what changed in useful output, what changed in cooling electricity and which assumptions still need checking. New meters should answer a specific unresolved question, rather than expand a dashboard that nobody owns.

See also: Relay research finds critical gaps in factory-floor information-sharing, leading to failures

The next investment decision can then use three clearly separated figures: directly measured cooling electricity, allocated shared electricity and any remaining estimate. Energy cost calculations also need the applicable tariff; peak demand charges may require a separate assessment.

At the next budget review, the team should be able to show where its cooling-cost figure came from, what assumptions it contains and what expansion might change. A documented estimate can support that discussion. An unexplained share of the factory's overhead cannot.

About the Author

Brenda Pak

Brenda Pak

Brenda Pak is co-founder and CEO of Reliability Engine. She writes about liquid-cooled AI infrastructure and the operating decisions that determine whether a deployment is ready for use. Her writing appears in Reliability Engine Insights.

Sign up for our eNewsletters
Get the latest news and updates