A new power paradigm
Data centre design is evolving as occupants demand more power and more progress on sustainability.
Even as capacity is ramping up across Europe, data centre design is undergoing an evolution whose outcome is not yet certain.
AI workloads rely on high-power processors, such as GPUs, tensor processing units and custom AI accelerators. This is pushing rack densities beyond the standard 50kW per cabinet.
As a result, data centre occupants are demanding ever greater density: 73% of survey respondents expect customer demand for power density to increase in the next three years, including 24% who expect it to increase significantly (Figure 4).
Figure 4. Occupants’ growing demand for power density
How do you expect demand from data centre occupants for the following attributes to change in the next three years? (% of respondents)
This concentration of power means more heat. This, in turn, is stoking demand for liquid cooling, which is more efficient for high-heat applications than conventional, adiabatic cooling.
Data centre designs are therefore pivoting to liquid cooling. “I can't tell you how many jobs we've seen change to liquid cooling over the past six to nine months,” says Byrne.
In line with last year, operators we surveyed expect liquid cooling to be the dominant model by 2030, accounting for 61% of their capacity (Figure 5). But they have made little to no progress since last year’s survey: back then, respondents expected to achieve 42% liquid cooling in 2024. For 2025, that figure has dropped to 40%.
This slow progress reflects the industry’s inertia. Although the hyperscalers are all moving towards high-power density with liquid cooling, an industry standard design has yet to emerge.
“There’s a bottleneck,” explains STACK Infrastructure’s Baylis. “We’re waiting for customer design standards and approach to be finalised.”
When asked which liquid cooling solution their organisation is most likely to adopt, respondents’ most common answers are rack-based immersion cooling (33%) and direct-to-chip cooling (24%). There is evidently no consensus on a single approach yet.
Another area of uncertainty is quite how much power density will really be needed to support AI. The emergence of DeepSeek, a large-language model whose performance rivals that of industry leaders without the need for high-end GPUs (its creators claim), has challenged assumptions about the continued need for more power.
“Some operators are building gigawatts, but others are holding back,” says arch.law’s Lamb. “Efficiencies in software engineering could mean that we don’t need the power we’re all expecting – or, at least, we might not need it all in the same place.”
Staying on track in sustainability
The backdrop to this evolution in power and cooling is a continued effort to improve the sector’s energy sustainability. Our survey respondents’ highest priority for data centre design is reducing emissions and increasing energy efficiency, with 58% ranking this in their top three.
This is driven by customers’ sustainability demands, constraints on the energy supply and the need to keep bills in check – with energy comprising a significant chunk of a data centre’s lifetime cost, any improvement in efficiency pays a substantial return.
But the shift to high power density will increase the total energy consumption of a data centre, piling pressure on an already limited supply of power from the grid. This means operators must continue to drive efficiency and eliminate emissions from their energy footprint.
Some organisations have adopted onsite renewable generation, such as solar – 38% of survey respondents have done so, more than any other sustainability measure in the survey (Figure 6). A further 33% plan to adopt this in the next one or two years.
Figure 6. Data centre operators adopt on-site renewable generation
Which of the following has your organisation adopted in response to environmental regulation or to meet sustainability commitments? (% of respondents)
However, on-site generation will only ever provide a fraction of a data centre’s energy needs, especially in Northern European climates. And its adoption may be driven as much by the need to achieve certifications, such as the Uptime Institute’s Tier Three classification system, as by the sustainability benefits.
Instead, energy decarbonisation will be driven by uptake of offsite renewable sources, through power purchase agreements (PPAs) and direct grid connections. In some cases, this will require investment in local energy infrastructure, such as private grid connections or microgrids.
Another knock-on effect of the shift to liquid cooling is an increasing use of water and, with it, a growing sustainability risk. The water consumption of data centres is under growing scrutiny, and regions where water is scarce may soon impose regulatory restrictions on access to the municipal supply. Operators may therefore need to invest in onsite water treatment and closed-loop liquid cooling to reduce their offsite water usage.
This is yet another example of how the evolution of power and cooling design, combined with today’s sustainability constraints, increases the sector’s reliance on specialist equipment and the engineering expertise to install and operate it. Unfortunately, both are in short supply.
The nuclear option
An ever-expanding thirst for energy has prompted some hyperscale data centre operators to consider small modular reactors (SMRs) to power their facilities. Nearly three quarters (73%) of our respondents expect these to be widespread in Europe within 15 years.
SMRs face many hurdles before they are likely to become commonplace, however. First, they are likely to face opposition from their local communities, many of whom are already sceptical of data centre projects. Some projects have offered local communities the incentive of cheap energy, but this would require local grid upgrades.
Second, the EU‘s strict regulatory environment means progress is likely to be far slower than in other jurisdictions. Last year, hyperscalers Google and Amazon signed the first corporate deals for data-centre SMRs in the US.
Meanwhile, the European Industrial Alliance on Small Modular Reactors aims to facilitate the first SMRs projects in Europe in the early 2030s. (The UK government has got as far as shortlisting suppliers for its SMR development scheme but has not set a target launch date.)
Third, widespread SMR adoption would require a rapid increase in the supply of suitably trained engineers. Romania’s nuclear agency recently estimated that the country needs 5,000 nuclear engineers in the next decade—more than the current capacity of all its university training schemes.
SMRs may one day power Europe’s data centre sector but they will not eliminate its energy constraints any time soon.
Share this report:
