AI Capacity Is Becoming a Thermal Planning Problem

AI infrastructure planning is increasingly becoming a question of how much compute a facility can physically cool, not simply how many processors it can install. Higher-performance AI systems concentrate substantial power into fewer racks, increasing the amount of heat that facilities must remove from increasingly dense spaces. Research from Lawrence Berkeley National Laboratory shows that liquid cooling can transfer heat from components such as CPUs and GPUs more efficiently than air cooling, while high-density computing is driving greater interest in these systems. That changes the economics of capacity expansion because additional compute can require changes to the thermal infrastructure supporting it. A facility may have electrical capacity available while still facing limitations in rack density, heat rejection, coolant distribution or mechanical infrastructure. The result is a more complicated definition of capacity. Installed compute may exist physically, but the facility still needs sufficient thermal capability to operate that compute reliably.

For AI infrastructure buyers, this distinction matters because capacity has traditionally been discussed in terms of power, floor space and hardware. Cooling was often treated as an engineering requirement that followed those decisions. Higher-density AI deployments are making that separation harder to maintain. Uptime Institute’s 2025 cooling survey found that high rack density was the leading factor driving direct liquid cooling adoption, while traditional air cooling remained widely used across the industry.

More GPUs Do Not Automatically Mean More Usable Capacity

Adding processors does not automatically create equivalent usable AI capacity. The practical limit can depend on whether the facility can deliver the required power, remove the resulting heat and maintain operating conditions across the intended workload. That makes thermal capacity increasingly important to the economics of adding AI compute. The distinction is particularly important when infrastructure is designed around high-density racks. A cooling system that works effectively for one rack configuration may require additional distribution equipment, control systems or facility modifications when rack density increases. Direct liquid cooling can address some of those thermal challenges by moving heat closer to the source rather than relying entirely on room-level air movement. Berkeley Lab research has demonstrated that direct liquid cooling can reduce the cooling burden placed on air-based systems. The economic question, therefore, is not simply whether liquid cooling costs more than air cooling. The more useful question is what amount of usable compute each cooling architecture enables within the available infrastructure. A more expensive cooling system can make economic sense when it allows a facility to deploy more compute within its existing power and space constraints. Conversely, liquid cooling can become difficult to justify when the workload does not require high rack density.

Liquid Cooling Changes the Expansion Equation

Liquid cooling changes the expansion equation because it connects compute density more directly with the infrastructure responsible for removing heat. The technology does not eliminate cooling costs. Instead, it changes where those costs occur and how closely they are tied to compute deployment. A liquid-cooled environment requires more than cold plates or other component-level equipment. Operators must consider coolant distribution, pumps, heat exchangers, cooling distribution units, monitoring, maintenance and appropriate fluid management. These systems create additional operational requirements alongside their thermal benefits. Current industry research also shows that liquid cooling adoption remains gradual, with integration into existing infrastructure continuing to influence deployment decisions. That makes the economics highly dependent on the design of the facility. A purpose-built high-density environment may be able to integrate liquid cooling into its electrical, mechanical and rack architecture from the beginning. A legacy facility may face a different calculation if substantial modifications are required before the same technology can support new AI workloads.

Cooling Capacity Can Become a Hidden Expansion Limit

The hidden constraint emerges when installed infrastructure and usable infrastructure stop meaning the same thing. A facility can have available floor space and electrical capacity but still lack sufficient thermal capability for the rack densities required by a particular AI system. That gap can reduce the amount of compute that can actually operate at the intended performance level. The same issue can appear at the customer level when contracted infrastructure supports a particular configuration but future hardware requires different thermal characteristics. Long-term capacity agreements can therefore make thermal headroom a relevant commercial consideration. Buyers may need to understand not only how much compute is available today, but also what range of hardware and rack configurations the cooling infrastructure can support. This is where liquid cooling becomes an economic planning decision rather than simply a mechanical one. The value comes from the usable capacity that the cooling architecture enables. If a facility can support higher-density systems without repeatedly rebuilding its thermal infrastructure, expansion economics can change.

The Economics Depend on What Capacity Actually Means

The phrase “AI capacity” can hide several different measurements. It can refer to the number of GPUs installed, the amount of electrical power allocated to those GPUs, the number of racks available or the amount of compute that can operate under defined thermal conditions. These measures do not necessarily produce the same answer. Liquid cooling makes the distinction more visible because the cooling architecture becomes closely connected to rack configuration and hardware selection. Different liquid-cooling systems can also use different approaches to coolant distribution, operating temperatures, flow requirements and system boundaries. Berkeley Lab has worked on specifications for liquid-cooled racks and transfer fluids precisely because common technical specifications can help address interoperability and deployment challenges. For buyers, the implication is straightforward: cooling capacity should be evaluated alongside compute capacity. A procurement decision that focuses only on processors can overlook the infrastructure required to operate those processors at the expected density. The more useful commercial metric is therefore usable AI capacity under defined power, thermal and operational conditions.

Expansion Planning Needs Thermal Headroom

Thermal headroom also matters when infrastructure is expected to evolve. AI hardware does not remain static, and future systems can impose different power and cooling requirements than the equipment installed today. That does not mean every future upgrade will require liquid cooling, but it does mean that cooling architecture can influence how easily a facility accommodates higher-density equipment. The issue becomes particularly important for infrastructure designed around rapid expansion. A facility that has limited thermal headroom may need additional cooling equipment before it can add the next generation of compute. That can introduce capital expenditure, engineering work and deployment time that were not visible in the original hardware plan. The economic value of liquid cooling therefore depends on the expansion pathway. It can be strongest where increasing compute density would otherwise require major changes to conventional cooling infrastructure. It can be less compelling where workloads remain within the practical limits of existing air-cooled systems.

Liquid Cooling Could Shift Where Expansion Happens

Liquid cooling can become one factor in facility selection as AI infrastructure moves towards higher-density deployments. Locations with suitable electrical infrastructure, cooling architecture and expansion potential can offer different economics from facilities that require extensive retrofits. The technology does not determine location by itself, but it can influence the practical amount of compute a site can support. That distinction could become more important as operators build infrastructure specifically around dense AI workloads. Uptime Institute’s research shows that direct liquid cooling adoption is still gradual, while high rack density remains a major reason operators consider it. That suggests the technology is likely to remain closely associated with workloads where thermal density creates a clear operational requirement. For AI buyers, this creates a more useful way to assess capacity. Instead of asking only how many GPUs a facility can host, they can ask how much compute the site’s electrical and thermal infrastructure can support, under which rack configurations and with what expansion margin.

Buyers Need Evidence Behind Cooling Claims

The commercial value of liquid cooling ultimately depends on evidence rather than technology labels. Buyers may need to compare power delivery, cooling performance, rack density, maintenance requirements and expansion capability as part of the same infrastructure decision. That does not mean every AI deployment should move to liquid cooling. Air cooling remains widely deployed, and liquid cooling introduces its own design and operational requirements. Uptime Institute’s 2026 analysis also highlights continuing challenges around resiliency design and operational complexity, reinforcing that higher cooling performance does not automatically produce a better business case.

The stronger argument is narrower. Where AI workloads require substantially higher rack density, liquid cooling can change the amount of usable compute that a facility can support within its physical infrastructure. That can change the economics of expansion because the relevant investment is no longer just the cost of adding processors. It includes the thermal infrastructure required to keep those processors productive.

The Real Value Is Usable AI Capacity

The next phase of AI infrastructure economics will depend less on the number of processors that can be purchased and more on how much compute can be operated consistently. Cooling is becoming part of that equation because every high-density AI system converts electrical power into heat that the facility must remove. Liquid cooling does not make that heat disappear, and it does not guarantee lower infrastructure costs. What it can do is provide a different method for managing concentrated thermal loads. Berkeley Lab’s work and broader industry research show why that distinction matters as high-performance computing becomes increasingly dense. For end users, the practical lesson is to treat cooling as part of capacity planning rather than as a technical detail added after compute is purchased. The most valuable AI capacity will be the capacity that can operate reliably within defined power, thermal and operational limits. As rack densities rise, that definition could make liquid cooling an increasingly important part of the economic calculation behind AI capacity expansion.