
On November 2, 2026, OpenAI announced its Grid Protectors Program, describing plans to work with utilities so its future compute campuses can reduce electricity demand when local grids are under strain. The program describes future operating arrangements, with its planned Georgia campus intended as a pilot.
For infrastructure buyers, this raises a practical question: what must the cooling and power systems keep doing while the IT load falls, stays lower and returns? A useful equipment brief needs to describe those transitions as carefully as the peak load.
What is AI data center load flexibility?
AI data center load flexibility is the ability to adjust facility electricity demand within agreed operating limits while meeting the requirements of the workloads that remain active. The opportunity depends on the workload, equipment, controls and connection agreement.
EPRI's October 27 design insight considers generation, electrical storage, flexible cooling and thermal storage against different flexibility needs. Its framework distinguishes requirements such as notice, activation duration and annual use. The engineering implication is to start with an operating brief before choosing equipment.
1. Define the event before sizing the equipment
Ask the owner and IT team to describe the intended reduction and recovery. A peak-load schedule alone leaves important decisions unresolved. The procurement brief should identify:
- Which workloads can change and which services must remain available.
- The intended reduction, notice period and event duration.
- The expected rate of load reduction and subsequent recovery.
- The cooling and electrical limits that must remain protected.
- Who authorizes the event, monitors it and ends it if conditions change.
Keep planned demand reduction separate from an electrical outage. Both deserve review, but they can require different control sequences and acceptance criteria.
2. Specify CDU behavior across the load range
A coolant distribution unit rating describes capacity under stated conditions. It does not, by itself, demonstrate how a complete liquid-cooling loop behaves during a load transition. OCP's liquid-distribution guidance treats temperature control, flow, pressure and loop boundaries as distinct design considerations.
For the proposed CDU and connected servers, request a documented operating range. Confirm the permitted temperatures, minimum and maximum flow, available pressure and control responsibility with the relevant suppliers. Then ask what happens when the load changes or a pump becomes unavailable.
Source 3: Open Compute Project
- Identify the signals used to control pumps and valves.
- Confirm which flow and temperature limits apply at reduced load.
- Describe how the system returns to the required cooling condition before or during IT recovery.
- Include the agreed alarms, fallback behavior and operator intervention points.
These questions complement the piping and fluid-quality review in Coolnet's existing CDU integration article. Their purpose here is to turn a changing workload profile into a control sequence that can be reviewed and tested.
3. Review the full path from the server to outdoor heat rejection
Follow the heat beyond the CDU. Show the facility circuit, pumps, outdoor heat-rejection equipment and any supporting air cooling on the same system review. Identify the power and controls needed by each part of that path.
In Coolnet's published liquid-cooled modular configuration, CDUs connect liquid-cooled racks to the facility circuit and outdoor dry coolers, while in-row air conditioning handles the remaining air-side heat. That is a reference arrangement; the liquid/air split must follow the selected servers and project design.
Use that distinction in the operating brief. Which cooling equipment remains active during the reduction? What ambient conditions constrain recovery? Which signals confirm that the required cooling is available? Review the actual installation and equipment schedules before assigning any flexible cooling capacity.
4. Protect the UPS reserve required by the reliability plan
UPS output capacity and available battery energy answer different questions. ENTSO-E's May 2026 report notes that using UPS capacity for grid support is constrained by battery duration and the facility's required protection level. An installed UPS should therefore not be assumed to provide an unrestricted flexibility resource.
If battery participation is proposed, request a separate review of compatible controls, battery condition, discharge limits, recharge requirements and the reserve needed for the agreed backup duty. Define what happens if a supply disturbance occurs during the event or before the battery has recovered.
Make the ownership clear: the electrical design and operating agreement should state who approves battery use and which reliability requirements take priority. Check the selected UPS and battery system rather than inferring these capabilities from a product-family name.
5. Include the return to load in the acceptance plan
Reduction is only part of the sequence. Recovery brings its own coordination questions: when can IT demand rise, which cooling conditions must already be available, and what happens if a required signal is missing?
For the equipment in the project scope, agree a test sequence and measurable limits before acceptance. The following is a planning structure to adapt to the design, not a universal operating procedure.
| Operating state | Review question | Evidence to request |
|---|---|---|
| Normal operation | Are IT, cooling and power conditions within the agreed envelope? | Baseline trends, settings and available backup reserve. |
| Reduced load | Do the remaining workloads retain their required cooling and power? | Load, coolant temperature, flow, pressure and alarm records. |
| Recovery | Is the return to load coordinated with cooling availability and battery recovery? | A time-aligned sequence log with the agreed acceptance limits. |
Record the assumptions, pass/fail observations and unresolved interfaces. A single stable operating point does not answer every transition question; the acceptance record should cover the conditions promised in the project brief.
Bring a load profile into the infrastructure discussion
For a new AI data center phase or retrofit, bring the expected workload profile, server cooling limits, proposed event duration and recovery requirements into the same review as the equipment list. This gives facilities, IT, controls and procurement teams a shared set of conditions to evaluate.
Coolnet's published portfolio covers cooling, CDUs, critical power and integrated data center systems. Share your project inputs with the team to discuss equipment selection and interfaces for your deployment. Any demand-response function or site performance target needs its own project-specific assessment and verification.
Frequently asked questions
Does a larger CDU automatically make an AI data center load-flexible?
No. A larger rating does not establish the behavior of the servers, liquid loops, outdoor cooling or controls during reduction and recovery. Specify the required operating range and test the relevant transitions for the selected equipment.
Can an existing data center UPS automatically be used for demand response?
Do not assume so. Review the exact UPS, battery system, control capabilities, operating agreement and protected backup reserve. The project must define permitted use and recovery conditions before assigning battery capacity to a flexibility event.
Does liquid cooling eliminate the need to review air-side cooling?
Not necessarily. Establish the heat captured by liquid and the heat still handled by air for the selected equipment. Coolnet's published modular liquid-cooling reference includes supporting in-row air conditioning; other server and facility configurations require their own assessment.
Sources
- OpenAI: Introducing the Grid Protectors Program (2026-11-02).
- EPRI DCFlex: Designing for Flexibility: How Onsite Gen, Energy Storage, and Flexible Cooling Technologies Align With Different DC Flexibility Needs (2026-10-27).
- Open Compute Project: Data Center Liquid Distribution Guidance & Reference Designs, Revision 1.0 (2024-01-08).
- ENTSO-E: Data centres and the power system: expected trends, challenges and opportunities (2026-05).