Thermal Flexibility Infrastructure: Using Underground Cold Storage for AI Data Centers
AI data centers are usually discussed as electricity projects. Servers need power, utilities need to connect them, and developers look for generation that can run continuously. Cooling is often treated as a supporting system, even though a large facility cannot operate without it.
That framing leaves out a useful design question: can cooling be produced at one time, stored, and dispatched when the grid or the facility needs it most?
The U.S. Department of Energy is studying Cold Underground Thermal Energy Storage, or Cold UTES, for this purpose. The approach injects cold water into the subsurface, stores it underground, and draws it back when cooling demand peaks. This article uses thermal flexibility infrastructure as a proposed industry frame for the equipment, geology, controls, and contracts that make that kind of time-shifted cooling possible. The phrase is an editorial concept, not a DOE standard or a proven commercial category.
What thermal flexibility infrastructure means
Thermal flexibility infrastructure is the layer that lets a facility change when it produces and consumes cooling without losing operational control. It can include underground storage, heat exchangers, pumps, sensors, control software, site models, and agreements that define when the reserve may be charged or discharged.
The basic idea is simple. Use electricity during an off-peak period to create a cold reserve. Hold that reserve in a suitable subsurface formation. Use it during a peak period, reducing the electricity required by conventional cooling equipment at that moment.
The engineering is not simple. The system must match local geology, water conditions, cooling-loop design, server heat loads, operating temperatures, and reliability requirements. A useful reserve is not just a tank underground. It is a coordinated thermal system with a measurable charge and discharge behavior.
Why this concept is emerging
Data-center planning has focused on generation and transmission because those constraints are visible in interconnection queues. Cooling creates a different constraint. It is a continuous service with a time profile, and part of that profile may be flexible.
The DOE's Geothermal and Data Centers page says data centers' share of U.S. annual electricity consumption rose from 1.9 percent in 2018 to 4.4 percent in 2023. It cites a 2024 report that projects a range of 6.7 to 12 percent by 2028. Those figures are reported observations and projections, not a prediction that every site will reach the upper end of the range.
As computing loads grow, cooling becomes a place where operators can look for flexibility without asking servers to stop useful work. The opportunity is not to make cooling disappear. It is to move some cooling production to a better time and use the stored service later.
What the DOE research shows
The DOE's Geothermal Energy Storage page describes underground thermal energy storage as a way to store thermal energy in the subsurface and extract it later. It lists aquifer thermal energy storage, borehole thermal energy storage, and reservoir thermal energy storage as distinct configurations.
Cold UTES is a more specific configuration. The DOE says it injects cold water into the subsurface, stores it, and draws it back to offset peak cooling demand. Its project page describes a National Laboratory of the Rockies project, with Lawrence Berkeley National Laboratory as a sub-laboratory, that is exploring this use for energy-intensive operations such as data centers.
The project uses an off-peak charge and a peak-hour discharge. The DOE compares this cycle to a conventional battery in terms of when the system is charged and discharged, while noting that Cold UTES can also operate at seasonal time scales. The comparison describes the scheduling function, not an equivalence in materials, efficiency, or risk.
The DOE also describes other geothermal storage projects and ongoing technical and economic work. These pages document funded research and emerging technology. They do not establish that Cold UTES has demonstrated commercial savings across data centers or that a suitable site exists everywhere.
The system has four connected layers
1. Site thermodynamics
The subsurface must be able to store and release cooling in a predictable way. Aquifers, bedrock, and deeper reservoirs behave differently. Water movement, heat transfer, drilling conditions, and environmental constraints affect what a site can support.
2. Cooling-loop integration
The stored cold must connect to the facility's existing cooling architecture. That can involve heat exchangers, pumps, controls, and a fallback path when the reserve is unavailable. A design that works in a laboratory model may still require substantial retrofit work at a live facility.
3. Dispatch control
Operators need to decide when to charge and discharge. Time-of-use prices, grid conditions, weather, maintenance, and computing demand can all affect that decision. The control system must also preserve temperature and uptime requirements for the equipment it protects.
4. Measurement and contracts
A flexible thermal asset needs a way to prove what it delivered. Operators, utilities, and project financiers may need measurements of energy shifted, cooling delivered, reliability maintained, and costs incurred. Without that evidence, it is difficult to distinguish a useful grid service from a promising diagram.
New vocabulary for an emerging industry
Thermal flexibility infrastructure describes the complete layer that makes cooling shiftable. It includes physical storage, controls, data, and operating agreements.
Cold reserve is a proposed term for stored cooling capacity that can be dispatched later. The term focuses attention on availability and duration rather than on the storage medium alone.
Cooling dispatch describes the decision to use stored or mechanical cooling at a particular time. It places cooling beside other managed loads without claiming that the systems have identical behavior.
Site thermodynamics is a useful way to describe the combined geological and facility model. It asks whether a physical site can support the required storage cycle, not merely whether a technology works in the abstract.
These terms help separate four questions that are often collapsed into one: Can the site store cold? Can the facility use it? Can software schedule it safely? Can the result be measured and paid for?
Where startups can enter
The asset layer is capital intensive, but not every business opportunity requires drilling wells. A startup might build site-screening software that combines geological information with cooling-loop requirements. Another might provide controls that coordinate thermal storage with time-of-use prices and facility load. Other possibilities include instrumentation, performance verification, retrofit planning, and data systems that help utilities evaluate a flexible cooling resource.
These are opportunity hypotheses, not evidence of customer demand. Founders should begin with a narrow facility type and a specific operating decision. A tool that predicts when to charge a cold reserve may be useful only if the operator can trust the prediction, override it, and verify the result afterward.
Hypothetical example: a regional data-center operator has a cooling system with predictable overnight slack and recurring afternoon peaks. It studies whether an underground reserve could be charged at night and dispatched during those peaks. A software startup helps model the site, estimates the usable reserve under different temperatures, and records every charge and discharge event. The operator uses those records to decide whether a larger pilot is justified. This example is hypothetical and does not describe a customer or a verified project.
What the concept does not prove
Cold UTES is not a universal replacement for electrical batteries, backup generation, or conventional chillers. It stores a thermal service, so its value depends on the cooling demand and the equipment around it. A site may also face geology, water, permitting, drilling, contamination, or retrofit constraints.
Nor does the DOE project prove that data centers can freely curtail cooling. Cooling requirements are tied to equipment safety and workload conditions. Any flexibility claim needs site-specific evidence, clear operating limits, and a fallback plan.
The most reliable near-term use of this vocabulary is as a design checklist. Ask which part of the cooling load is flexible, what physical reserve can serve it, which measurements verify delivery, and who carries the risk when the reserve underperforms.
Conclusion
Thermal flexibility infrastructure treats cooling as a service that can be scheduled, measured, and coordinated with the grid. Cold UTES is one emerging approach to that problem, and DOE research shows why it deserves attention without settling its commercial future.
For founders, the opportunity is not to repeat that underground storage is promising. It is to make the system legible: model the site, connect storage to the cooling loop, dispatch it within safe limits, and prove what happened. That work could become the software and verification layer around a new class of energy-intensive computing infrastructure.