Cooling: The Invisible Infrastructure
Clarity in a Changing World.
For informational purposes only. Not investment advice.
Introduction
Every unit of computation eventually produces heat. For most of the computing era, removing that heat was primarily an engineering requirement rather than a constraint on how much computing could be installed.
AI is changing that relationship. Advanced accelerators consume more power, while rack-scale architectures concentrate large numbers of processors into increasingly dense systems. NVIDIA’s current GB300 NVL72, for example, integrates 72 Blackwell Ultra GPUs and 36 Grace CPUs in a fully liquid-cooled rack-scale architecture. NVIDIA’s March 2026 reference architecture states that a full GB300 NVL72 rack can require up to 142 kW of power. (NVIDIA)
At these densities, thermal management becomes part of the effective capacity of the computing system. The question is no longer simply how many processors can fit into a data centre, but how much heat the facility can remove while keeping those processors operating efficiently.
Computing Density Changes the Cooling Problem
AI increases cooling requirements in two related ways. It raises total power consumption, but it also concentrates more of that power within the same rack and physical space.
Traditional data centres could distribute computing loads across relatively large areas, allowing air cooling to handle many workloads effectively. High-density AI systems create a different thermal environment. More accelerators, memory and networking equipment operate close together, producing substantially more heat within a smaller footprint.
The transition can already be seen in commercial infrastructure. In its January 2026 MegaMod HDX product announcement, Vertiv said the system was designed for AI and high-performance computing environments with rack densities from 50 kW to more than 100 kW per rack, using a combination of direct-to-chip liquid cooling and air cooling. These figures describe the design range of a specific Vertiv system rather than average rack density across the data-centre industry, but they illustrate the thermal requirements that infrastructure suppliers are designing around. (Vertiv)
As density rises, simply moving more air becomes progressively less efficient. Fans require power and space, while maintaining acceptable temperatures across tightly packed equipment becomes more difficult.
At sufficiently high densities, the cooling architecture itself has to change.
Liquid Cooling Has Entered Commercial AI Infrastructure
Liquid can transfer heat more effectively than air, allowing thermal energy to be removed closer to the components producing it. This makes liquid cooling particularly relevant as processors become more powerful and racks become denser.
The important change is that liquid cooling is no longer confined to specialised high-performance computing environments. It is now incorporated into commercially available AI systems.
NVIDIA describes the GB300 NVL72 as a fully liquid-cooled, rack-scale architecture, and its current product page lists the system as available. NVIDIA’s enterprise reference architecture also documents liquid cooling as part of the rack design. (NVIDIA)
That does not establish that all AI data centres are moving to liquid cooling. Lower-density workloads can continue to use air cooling effectively, while hybrid designs can combine both approaches. Vertiv’s MegaMod HDX, for example, integrates direct-to-chip liquid cooling with air-cooled architectures rather than treating the two as mutually exclusive. (Vertiv)
The evidence therefore supports a narrower conclusion:
Liquid cooling has entered commercial deployment for some high-density AI architectures, while air and hybrid cooling remain relevant across the broader data-centre market.
Cooling Determines Usable Computing Capacity
The relationship between cooling and computing is straightforward. A processor can sustain its intended operating performance only if the heat it generates can be removed.
This creates a distinction between nominal capacity and usable capacity.
A facility may have enough floor space for additional AI hardware but lack the thermal or electrical infrastructure required to operate that hardware at the intended density. Existing data centres designed for conventional cloud workloads can therefore face different constraints from facilities designed specifically around high-density AI systems.
Retrofitting can require changes to coolant distribution, pumps, heat exchangers, chillers, monitoring systems and sometimes electrical infrastructure. The economics depend on the existing facility, the intended hardware and the density of the workload.
New facilities can address these requirements from the beginning. Older facilities may require substantial modification.
This means aggregate data-centre floor space alone does not measure how much high-density AI capacity can actually be deployed.
The configuration of the infrastructure matters as much as its physical size.
Cooling Also Uses Electricity
Thermal management solves one physical problem while consuming another constrained resource: electricity.
Fans, pumps, chillers, cooling towers and other thermal systems require power. The precise amount depends on the facility, climate and cooling architecture, so there is no single relationship between liquid cooling and overall energy consumption.
The economic mechanism is nevertheless clear. If total electrical capacity at a facility is constrained, electricity consumed by thermal management is unavailable for computing.
Reducing the electricity required to remove a given amount of heat can therefore leave a greater share of available facility power for processors and related computing equipment.
This does not mean that cooling efficiency and capital efficiency are identical. The relationship is narrower:
Cooling efficiency can influence how productively a data centre uses its available electrical capacity.
That relationship becomes more relevant as both rack density and facility-level power demand increase.
Cooling Also Interacts With Water Availability
Water adds another dimension, but the relationship is more complicated than the simple assumption that liquid cooling necessarily means greater water consumption.
Some data-centre cooling architectures use evaporative processes that consume water. Others rely more heavily on closed-loop systems in which coolant circulates repeatedly through the facility. Chilled-water, direct-to-chip and hybrid architectures can also differ substantially in how they interact with external water resources.
Water requirements therefore depend on the cooling architecture, heat-rejection system, climate and operating conditions.
The distinction matters because data centres are physical infrastructure located within local resource systems. In regions where water is constrained, cooling design can affect both site selection and operating decisions. In other regions, the issue may be considerably less significant.
The appropriate conclusion is therefore not that liquid cooling necessarily increases water demand.
Water use is architecture- and location-dependent, and should be assessed separately from the broader transition towards higher-density thermal management.
Existing Data Centres Face a Different Problem
The transition towards higher-density computing creates an important distinction between new construction and the installed data-centre base.
Purpose-built AI facilities can be designed around higher electrical loads and liquid-cooling infrastructure from the beginning. Existing facilities were often constructed for much lower rack densities.
Their buildings may remain perfectly usable while their supporting systems become the constraint.
This creates a category that is increasingly relevant to infrastructure analysis: AI-ready capacity.
A conventional data centre and an AI-ready data centre may have similar floor space but very different abilities to support high-density accelerators. Electrical distribution, coolant infrastructure, heat rejection and the physical layout of the facility can all affect the amount of usable compute that can be installed.
The distinction also helps explain why AI infrastructure investment extends beyond processors. Existing buildings may require substantial supporting investment before they can accommodate new generations of computing equipment.
Suppliers Are Adding Thermal Capacity
Another indication of the change comes from the capital being committed by infrastructure suppliers.
In March 2026, Vertiv announced approximately $50 million of planned investment to expand manufacturing in Ohio. The company said the Ironton expansion, expected to become operational in the second quarter of 2027, was intended to increase production capacity for liquid-cooling and chilled-water systems used in advanced thermal-management applications. (Vertiv)
In April, Vertiv announced the acquisition of Strategic Thermal Labs, describing the transaction as adding expertise in cold-plate design, server-side liquid cooling and high-density thermal validation. (Vertiv)
In July, Vertiv announced further investment at its Tognana campus in Italy. The company said it expected the expansion to double its regional chiller production capacity by the end of 2026, with a new large-scale testing laboratory planned for early 2027. (Vertiv)
These announcements do not establish the future growth rate of the liquid-cooling market, nor do they establish the future profitability of Vertiv or other thermal-management suppliers.
They establish something narrower: one major infrastructure supplier is committing additional manufacturing capacity and engineering resources to thermal systems designed for high-density computing.
That is consistent with cooling becoming a more significant component of the physical infrastructure required for AI deployment.
Demand Evidence Extends Beyond Cooling Alone
Vertiv’s financial results provide additional context, although they need to be interpreted carefully.
In its second-quarter 2026 earnings release, Vertiv reported net sales of $3.274 billion, up 24% year on year, including 18% organic growth. The company raised its full-year 2026 guidance to approximately $14 billion of sales at the midpoint and 31% organic growth. Chief Executive Giordano Albertazzi said in the release that demand for AI and general compute continued to intensify. (Vertiv)
These results should not be interpreted as evidence that liquid-cooling revenue itself grew by a similar amount. Vertiv supplies a broader range of power, thermal and infrastructure products, and the reported sales figures do not isolate liquid cooling.
They do provide company-level evidence of stronger demand across the broader physical infrastructure serving data centres, alongside the specific manufacturing investments Vertiv is making in thermal management.
The distinction is important:
Growth in AI infrastructure demand does not tell us automatically how much economic value will accrue to cooling, or to any individual supplier.
That requires separate evidence.
Thermal Management Is Becoming a System
The cooling requirement extends beyond chillers or individual cooling units.
High-density liquid-cooled facilities can require cold plates, coolant distribution units, pumps, heat exchangers, piping, monitoring systems, controls and heat-rejection equipment. These components must interact with servers, electrical infrastructure and building systems.
As the systems become more integrated, the interface between the server and the facility also becomes more important.
Vertiv’s April 2026 acquisition of Strategic Thermal Labs illustrates this direction. In announcing the acquisition, the company specifically referred to engineering at the interface between server-side liquid cooling and supporting infrastructure, including flow, controls, serviceability and lifecycle reliability. (Vertiv)
This is a company description of its strategic rationale, not evidence that one architecture will dominate the market. But it illustrates how thermal management is moving from a collection of individual components towards a more integrated engineering problem.
As with the optical infrastructure discussed later in this series, greater demand does not mean that economic value will be distributed evenly across the supply chain.
The relevant question is:
Which parts of thermal management remain technically difficult, operationally critical or difficult to replace as cooling architectures evolve?
What Should We Watch?
The first indicator is rack power density. NVIDIA’s March 2026 enterprise reference architecture specifies up to 142 kW for a full GB300 NVL72 rack, while Vertiv is marketing infrastructure designed for rack densities above 100 kW. These figures refer to specific systems, not industry averages, but they show the direction of high-density AI architecture. (NVIDIA Docs)
The second is commercial deployment of liquid-cooled systems. The presence of commercially available liquid-cooled rack-scale platforms provides stronger evidence than product roadmaps alone.
Third is the balance between air, hybrid and liquid architectures. The relevant question is not whether one technology eliminates the others, but where increasing density changes the economically appropriate cooling method.
Fourth is energy and water efficiency. As power and local resources become constrained, thermal performance should be considered alongside the electricity and water required to achieve it.
Finally, supplier capacity expansion can help distinguish realised infrastructure demand from expectations alone. Manufacturing investments, acquisitions, customer deployments and utilisation of new capacity provide evidence about how the physical supply chain is responding.
Together, these indicators can show whether advanced cooling remains concentrated in the highest-density systems or spreads more broadly across AI infrastructure.
What Could Change This Outlook?
The most important uncertainty is efficiency.
Future accelerators may produce more useful computation per watt. Model architectures and software may also reduce the amount of computation required for a given task. If improvements in efficiency outpace increases in computing density, thermal requirements could grow more slowly than current infrastructure designs imply.
Density itself is not predetermined. AI computing could become more distributed rather than continuing indefinitely towards higher power per rack. Different combinations of training, inference and edge deployment could produce different thermal requirements.
Cooling technology is also unlikely to converge immediately on a single architecture. Direct-to-chip liquid cooling, air cooling, hybrid systems and other approaches can coexist because facilities and workloads differ.
The evidence therefore supports a transition towards more advanced thermal management as high-density AI systems expand. It does not establish that liquid cooling will replace air cooling across the data-centre industry.
The pace of adoption will depend on computing density, facility design, power availability, retrofit economics and the relative performance of competing cooling architectures.
Conclusion
Cooling is becoming more closely integrated with the productive capacity of high-density AI infrastructure.
Commercial systems such as NVIDIA’s GB300 NVL72 show that fully liquid-cooled rack-scale computing has moved into available AI infrastructure, while Vertiv’s 2026 manufacturing investments provide company-level evidence that suppliers are adding capacity for more demanding thermal systems. (NVIDIA)
These developments do not imply that every data centre will adopt liquid cooling, or that one cooling architecture will dominate. They indicate something more specific: as computing becomes denser, thermal design increasingly affects how much computing can be deployed within the power and physical limits of a facility.
The next constraint follows from the same process. Dense AI systems do not only generate more heat; they also require processors to exchange rapidly growing volumes of data.
In the next article, The Optical Network Behind AI, we examine why moving data is becoming part of the economics of compute—and why optical communications are moving deeper into AI infrastructure.