Data centre operators across Europe are currently grappling with the sobering reality that AI and software are rapidly developing while physics moves at its own pace, writes Ad Vink, enterprise business leader – EMEA at Apx Data Centre Solutions.
Over the past year, we have been seeing a sudden surge where standard power rack requirements jump from 5-8kW to a 20.5kW threshold and beyond.
That almost overnight acceleration is happening due to high-density AI technology such as the NVIDIA H200s and Blackwell architectures which have caused a massive leap in software performance. However, while development is booming, operators are paying the price on the ground. Facilities that are a few years old are designed for an era of heat dissipation that no longer exists and our industry is hitting a thermal wall that legacy infrastructure was never designed to climb. In the rush to adapt, many operators are making critical mistakes.
The dangers of density debt
When I see operators trying to make infrastructure and legacy air-cooled racks fit 20.5kW without correct retrofitting, I warn them about a dangerous consequence called density debt. Density debt occurs when data centre operators resort to quick fixes to cool extreme heat, incurring large costs and failing to future-proof their infrastructure.
Examples of this include leaving adjacent racks empty to spread thermal load which can lead to expensive waste and turning up CRAH fans to 100% which speeds up equipment wear and makes utility bills surge. Leaving adjacent racks empty means that operators are paying for floor space and power distribution that they can’t use. Full immersion requires massive upfront CapEx too, because of costs such as fluid handling, safety and building structural reinforcement.
The consensus at recent industry summits is clear: the transition to high-density AI is happening “much too quickly” for the physical capacity of most existing sites.
Summit experts added that density debt can lead to compromising Power Usage Effectiveness (PUE), often seeing it spiral from 1.2 to 1.6. This 33% increase in total energy consumption can impact a business’s ESG and sustainability goals, and means that organisations take a step backward. Stopping density debt relies on retrofitting that is engineered to be future-proof rather than create short-term solutions.
Air vs immersion
In response to the thermal wall, many operators assume they must choose between two extremes – traditional air cooling or a complete overhaul of facilities. Yet, it doesn’t need to be a binary choice.
Standard air systems hit a hard physical ceiling at around 20.5kW per rack because of air’s limited heat transfer capacity. Shifting directly to full immersion introduces massive spending, complex maintenance and logistical challenges. It’s simply financially unfeasible for many sites.
While it’s not a simple issue to solve, hybrid cooling is an option. I view it as a realistic bridge between legacy sites and modern AI density. It allows operators to upgrade high-density zones without abandoning their current investment projects. A hybrid cooling set-up offers a phased CapEx approach; a potential solution to high costs. Under a hybrid cooling model we retain legacy cooling for standard compute tasks while integrating direct-to-chip (DLC) or localised liquid loops into high-density AI racks.
Retrofitting localised DLC parts or closed-loop liquid systems into existing rows allows operators to respect the physical limits of their site. Using this approach is a way of avoiding risky structural modifications and keeping energy consumption controlled. It also ensures operators maintain strict safety standards as well as keeping PUE and sustainability targets on track.
Engineering out the fear
Introducing DLC cooling is likely to trigger some anxiety for engineers – even ones experienced enough to have worked in this sector for a few years. There’s no doubt that bringing fluid into a cabinet that has millions of pounds of AI chips goes against what they’re specifically trained to do in their careers—keeping liquids away from racks, avoiding leads, short circuits and downtime.
The biggest hurdle to actually adopting liquid cooling is mindset. Our modern DLC infrastructure is designed to neutralise risks and allay anxieties. It uses non-conductive dielectric fluids or isolated loops with integrated leak detection.
The designs also have negative pressure vacuum loops. The system pulls liquid through instead of forcing it through cold plates under high pressure. Air is drawn into the pipe if a hose or seal is ever compromised which stops liquid spraying out over costly silicon.
We have the technology to make liquid cooling safe, but we need to work together to shift our operational culture from hesitation to trust.
The future
A sustainable and future-proof data centre is built on careful, considered and balanced engineering. I see the 20.5kW rack as the new baseline for the era of AI innovation. Operators must audit their density limits before the thermal debt gets unpayable and embrace the possibility of a hybrid evolution.
Learn more Apx Data Centre Solutions.
