AI is changing cooling - and enabling more energy-efficient heat recovery

Air cooling vs. liquid cooling
While traditional server environments
have largely relied on air cooling, AI
servers and HPC (High Performance
Computing) create much higher power
densities in a smaller footprint. According to the technical material UniCool
works with, the most demanding environments are moving from previous
rack loads of 2-15 kW to 300 kW and
above in the future.
From a design perspective, a load
above 40 kW per rack rules out air
cooling as an option; the shift to liquid
cooling becomes mandatory.
From a thermodynamic perspective,
the most important parameter for heat-transfer efficiency is volumetric heat
capacity [kJ/m³K]. Liquid has a value
more than 3,725 times higher than air;
therefore, the volume of liquid required
to transport the same amount of heat is far lower than the volume of air required,
as shown in the table below.
Direct-to-Chip Liquid Cooling
In direct-to-chip cooling, liquid is routed directly to the GPUs that generate the most
heat. A Coolant Distribution Unit (CDU)
connects the server cooling loop to the
rest of the data center’s cooling system.
From there, the heat can either be rejected
through dry coolers/chillers or transferred
for heat recovery. CDU requirements are
defined by the chip manufacturers.
Higher supply temperatures in the 35-
45°C range, depending on the system
and server requirements, can provide two
benefits.
Benefit 1: Higher temperatures enable
more free cooling
First, they allow more hours of free cooling, where heat can be rejected through
dry coolers without the same need for
mechanical cooling. In 2025, the highest
temperature was 34°C, so mechanical
cooling would not have been necessary.
However, during the 2026 heat wave,
temperatures exceeded 37°C, so extreme
weather conditions must be considered in
the cooling-system design.
Benefit 2: From waste heat to a heat
resource
The higher temperatures also make data
center waste heat easier to use. When the
required temperature lift is reduced, energy consumption is reduced as well.
According to UniCool, the simplest application is often to use the heat locally. If it
can be used to heat offices, greenhouses,
or other buildings
on the same site, a
heat pump may potentially
be unnecessary.
If the heat is instead supplied
to a district heating network, factors
such as distance, temperature requirements, heat-pump requirements, and heat
demand become decisive. The data center
produces heat year-round, while demand
from an external off-taker varies.
It is therefore not enough simply to identify available waste heat. There must also
be a technically and economically viable
route from the server to the off-taker.
“The data centers we help build today
are prepared so that the heat can be
harvested. If there is no off-taker from the
start, district heating or other users can be
connected later,” says Jimmy Hoffmann
Hansen.
That is why UniCool recommends assessing heat-recovery opportunities already
during the design phase. Even if there is
no heat off-taker from day one, the facility
can be prepared for a later connection.
Design considerations
According to UniCool, it is important not
to select a chiller based solely on maximum capacity. Load profile, temperatures,
operating hours, free cooling, and any
potential heat off-taker should be assessed together.
This also applies to existing data centers. One solution can be to establish a
separate high-temperature loop for new
AI racks while traditional server environments continue to run on their existing
cooling system. This allows capacity to be
expanded in stages rather than requiring a
complete rebuild.
It should also be kept in mind that some
equipment still has lower temperature
requirements, such as UPS systems, batteries, switchboards, and other technical
equipment. Cooling is still required to keep
these rooms at around 22-25°C.
The design must also perform on hot
days
Developments inside the data center are
only half the equation. The surrounding
climate is changing as well.
This summer, UniCool has seen older
cooling systems for critical infrastructure
come under pressure at ambient temperatures close to 40°C. Some of these
systems were originally designed for maximum temperatures of around 26-28°C.
New EU measures
At the same time, EU rules are placing
greater emphasis on data center energy
performance and the potential to use waste heat, making early planning even more
relevant. EED 2023/1791 of September 13,
2023, on energy efficiency and amending
the regulation introduced these measures.
EU 024/1364 of March 14, 2024, on the
first phase of establishing a common EU
scheme for rating the sustainability of data
centers addresses heat reuse from larger
data centers.
Among the key KPIs intended for focus
are Power Usage Effectiveness (PUE),
WUE, and Energy Reuse Factor (ERF), defined as Reused Energy / Total Data Center
Energy.
Whereas PUE simply measures how
much energy is wasted, ERF quantifies how much energy is returned to the
system. This represents a clear paradigm
shift - from a model in which heat is regarded as a waste product to an approach in
which heat is viewed as a resource. The
greater the share of energy that is recovered and reused, the lower the ERF value
will be.
In this way, the integration of direct-to-chip systems can directly influence
the company’s decarbonization strategy,
as defined in the GHG Protocol (Greenhouse Gas Protocol).