Cooling procurement / Evidence review
AI Cooling Savings Claims: What Vendors Must Prove
A cooling vendor says its AI delivers 20% savings. Is that 20% of chiller energy, cooling energy, the utility bill—or an adjusted result?
Sections in this article
You have two proposals for AI-assisted cooling. Vendor A promises “up to 20% savings.” Vendor B shows an audited reduction in measured chiller kWh over a specified operating period. Those statements do not tell you the same thing, and neither alone tells you whether the project will pay for itself.
This guide is for the commercial and engineering review meeting: the questions to ask, the evidence to request, and the acceptance language to agree before an optimization platform touches live setpoints. For the detailed mathematics of weather-normalized baselines, see our separate measurement and verification guide.
First identify what the percentage actually measures
Ask the vendor to complete this sentence: “The reported percentage is a reduction in ____ measured at ____ over ____ compared with ____.” If any blank is missing, the claimed outcome is not yet defined well enough for procurement.
| Marketing claim | What must be clarified | Why the distinction matters |
|---|---|---|
| “20% less chiller power” | Average kW, energy (kWh) or only the compressor? | Pumps, fans and operating hours may change |
| “20% less cooling energy” | Which meters and auxiliary devices are included? | Scope may omit tower fans or CRAH units |
| “PUE improved by 0.10” | Compared at what IT load and season? | The ratio moves when the workload changes |
| “20% off the energy bill” | Does it include IT energy, demand charges and tariffs? | It is a much larger financial assertion |
Terms like “up to” or “best observed day” need an explanation of how the result was selected and whether less favorable periods were excluded. The same applies to a pilot that happened to coincide with a cooler month.
The six-item evidence pack to ask for
- Measurement boundary. A marked diagram of every meter, what it includes and how IT energy is separated from facility overhead.
- Baseline method. A documented comparison period, variables such as IT load and outdoor conditions, and why those variables were selected.
- Actual readings. Time-stamped measured energy, not only screenshots of the optimization dashboard.
- Exceptions. Maintenance outages, changed rack allocations, failed sensors and periods excluded from calculations.
- Safety outcomes. Server inlet or coolant temperatures, rejected commands, manual overrides, comfort or process limits and hardware-related alarms.
- Commercial assumptions. Licenses, integration, sensors, operating support and claimed tariff or demand-charge savings.
IPMVP offers structured approaches to measurement and verification. The right option depends on whether the isolated cooling system or the whole facility is the measurement boundary. Do not impose a universal 12-month baseline, fixed R² threshold or compulsory alternating on/off schedule without first assessing whether the method is suitable and safe.
Worked example: 20% cooling savings is not 20% off the bill
Assume a hypothetical data center runs for one hour at a stable workload with 1,000 kW IT power, 300 kW cooling power and 100 kW of other facility overhead. Total facility power is 1,400 kW. The vendor lowers measured cooling power by 20%—a reduction of 60 kW—while everything else is unchanged.
Here the cooling saving is genuinely 20% of the cooling measurement, but only about 4.3% of total site power under the stated assumptions. Real billing may also include demand charges, time-of-use prices and changes in IT fan power. The Green Grid defines PUE at the facility and IT-energy boundary; it does not make PUE improvement synonymous with a corresponding bill reduction.
Design a pilot that could disprove the claim
Choose acceptance conditions before installation. An honest pilot should be able to report “insufficient evidence” or “not worth the cost,” not only success. Agree which operating regimes must be represented, how non-routine changes are handled and when the experiment must stop for operational reasons.
Report the central estimate and uncertainty, not just the best interval. Compare like-for-like periods or an appropriately modeled baseline.
Respect manufacturer operating envelopes, redundancy policy and approved human intervention rules. Site engineering defines the limits.
Include implementation, model support, maintenance, telemetry and effects on other energy consumers.
An on/off experiment may be useful where engineers determine switching modes is safe and periods are sufficiently comparable. In other facilities, a controlled baseline model or staged deployment is preferable. Rapid alternation can introduce carryover effects because building thermal mass and cooling plants do not reset instantly.
What should appear in the purchase agreement?
- A precisely defined metering boundary and agreed calculation method.
- Access to raw or sufficiently granular data needed for independent verification.
- Acceptance criteria for both savings and operational safety.
- Named responsibilities when equipment, software or tariffs change.
- A controlled rollback plan and a record of who can change setpoints.
- Clarity on fees if measured savings are materially below the proposal.
Ask for a sample report before signing. If the provider cannot show how sensor quality, missing data and baseline changes are explained, that gap will only be harder to resolve after deployment.
Practical takeaway: a credible AI cooling proposal is not the one with the largest percentage. It is the one with a measurable claim, a safe test design and enough data for the customer to check the result.
Sources and further reading
Primary frameworks and reference materials. Engineering decisions require site-specific validation; legal classifications require assessment of the actual use case.