HELSINKI / AGILITYPR.NEWS / October 28, 2026 /

The datacentre industry has made its decision on liquid cooling. Facilities running high-density AI workloads led the way. The major hardware vendors have designed their latest AI processor hardware around it. Direct-to-chip cooling, where liquid is routed through cold plates directly onto the processors, is now the reference architecture for any new AI-optimized facility. The question is no longer whether liquid cooling will become standard. It already has. The real question now is ‘What do accurate reliable measurements look like in this new scenario?’

Liquid cooling is not a new idea. IBM pioneered water-cooled mainframe computers in the 1960s, and the technology has been standard in supercomputers and high-performance computing clusters ever since. For the mainstream datacentre industry, however, it remained a niche concern. Interesting in principle, unnecessary in practice. Air worked well enough, and the economics of switching did not justify the complexity. That calculation changed, and fast. The trigger was artificial intelligence. As AI moved from research to large-scale commercial deployment, it introduced workloads that generate heat at a scale air cooling cannot handle.

A conventional server rack draws between 5 and 10 kilowatts. An AI rack today can draw ten to twenty, even thirty times that. At those densities, the physics of air cooling simply run out: heat capacity of air is relatively low, and moving enough air to extract the heat would require fan energy so large it becomes economically absurd, and even then, the hottest parts of the chip would not be reached in time. Liquid, with a heat capacity roughly 3,500 times greater than air by volume, is the only viable path forward.

Liquid cooling solves the thermal problem. But operating these systems reliably over time is a question the industry is still working out, and the measurement infrastructure that makes reliable operation possible is where much of that work remains.

The thermal trajectory has only one direction

To understand where the measurement challenge is headed, it helps to understand the rate at which the underlying hardware is changing.

According to AFCOM's 2025 State of the Datacentre Report, average rack density has already risen from 7 kW in 2021 to 16 kW today, and 79 percent of industry respondents expect the climb to continue. At the leading edge, NVIDIA's 2027 Kyber rack platform is projected to reach 600 kW with NVIDIA CEO Jensen Huang indicating future racks could eventually require a full megawatt. That would make a single enclosure as power-hungry as an entire small datacentre was a decade ago.

The gap between average and leading-edge deployments is widening with each new generation of AI hardware, and the leading edge sets the infrastructure requirements that everyone else eventually must meet.

Air cooling reaches its practical limit at roughly 20 to 30 kW per rack. At higher densities, fan energy alone makes air-only cooling economically unworkable. Liquid, with its vastly superior thermal conductivity, is not a preference at these densities, it is a physical requirement.

From one domain to two

Running a liquid-cooled facility is not a straightforward extension of running an air-cooled one. It is a categorically different operational environment, demanding new expertise across power, structural and mechanical engineering, fluid dynamics and chemistry, and instrumentation. And across all of these, measurement is the layer that makes the others controllable.

In a traditional air-cooled facility, the management of controlled environments is well understood. Temperature, relative humidity, differential pressure. Parameters, with decades of industry practice behind them, standardized in guidelines like those published by ASHRAE, the American Society of Heating, Refrigerating and Air-Conditioning Engineers, and served by mature sensor technology. Operators know what they are measuring, why, how that feeds to controls, and what the numbers mean.

Liquid cooling extends that domain significantly. Pipes carrying cold water have always been part of datacentre infrastructure, delivering chilled water to air handlers and fan wall units. But the scope and criticality of liquid measurement expands dramatically when cooling is brought all the way to the chips through CDUs and cold plates. The parameters that now require dedicated instrumentation include liquid temperature, pressure, flow rate, and level, as well as leak detection and coolant health indicators such as pH, turbidity, conductivity, and refractive index. Each carries its own failure modes, all requiring careful monitoring to ensure effective cooling loop performance and to maintain the needed quality of liquid, in the proximity of hardware that may be worth tens of millions of euros per square meter of floor space.

Liquid cooling takes over where air runs out of power, but air-side management remains essential for everything else. Direct-to-chip cooling takes the heat out from the hottest components, but the residual heat from the rest is cooled with air. As the hardware being protected becomes more expensive per square meter, the cost of any failure, whether from overheating, condensation, or coolant leak on the liquid side, or from inadequate airflow on the air side, grows with it. The stakes around every rack are higher, and so is the value of ensuring at all times, that the controlled environments are exactly as they should be. In liquid-cooled environments, the reliability and dependability requirements for measurement are even more critical than before.

Measurements in a liquid-cooled environment

Alongside the liquid-side parameters, ambient conditions in the data hall and in the critical parts of the cooling system must be actively managed. Dew point is where the risk concentrates. Liquid cooling’s efficiency depends on delivering cold liquid to the cold plates, but colder piping inside an air-conditioned space creates a condensation risk. When pipe surface temperature drops below the ambient dew point, water forms. Not a maintenance inconvenience, but a potential failure event with consequences that far exceed any maintenance budget. The standard practice is to maintain coolant supply temperatures at least 2°C above the measured ambient dew point. The coolant distribution unit is responsible for holding that margin, but only if it knows what the dew point actually is, continuously.

The measurement challenge across all these parameters is not exotic. It draws on well-established sensor physics. But applying it correctly, meaning knowing where to measure, how frequently, and how sensor data feeds into control decisions requires a thorough understanding of both the technology and the environment it operates in.

Dynamic facilities demand dynamic measurements

AI workloads are fundamentally different from traditional datacentre workloads. AI training jobs are intensive and intermittent. A cluster transitions from idle to full training load in minutes, generating a step change in thermal output that the cooling system must respond to in real time. The CDU managing that cluster needs to control the cooling loop, not with a delay of several minutes but instantly and continuously, ensuring supply and return temperatures are within limits, pressure and flow rates are as they should be, and the ambient dew point near the cold manifolds stays within tolerance as the system heats up.

Reliable measurements make dynamic control possible. Without accurate real-time data on critical parameters across the facility, the control systems are operating on delayed or inaccurate information. And the consequences cascade: inefficient cooling wastes energy, inadequate cooling risks hardware, and overcorrection wastes both. Every percentage point of efficiency improvement in a facility running hundreds of kilowatts per rack translates directly into operating cost and carbon impact. That connection between measurement quality and operational outcome is what makes getting the instrumentation right a business decision, not just a technical one.

What the next generation will require

The direction from here is more of everything that is already challenging. Rack densities will continue to increase. Each step in this progression adds measurement requirements that do not yet have the benefit of decades of industry practice behind them. Reliable measurements are the foundation on which everything else is built. They allow cooling systems to be controlled precisely rather than conservatively, energy efficiency targets to be achieved rather than merely aspired to. They give operators confidence that both the cooling infrastructure and the computer hardware it serves are operating within safe parameters. And they provide the early warning that separates a recoverable problem from an expensive one. Without trustworthy measurements, a liquid-cooled datacentre is operating on assumptions. And at these rack densities, assumptions are expensive. The industry is still building the understanding of what rigorous measurement looks like in this environment. That understanding is where the hard work of the next phase begins.

ENDS

Words: 1395

Sources

- AFCOM, 2025 State of the Datacentre Report

- ASHRAE TC 9.9, Thermal Guidelines for Data Processing Environments; Liquid Cooling Technical Committee guidance

Author: Anu Kätkä, Head of Datacentres and Critical Environments, Industrial Measurements, Vaisala