For a modern invention, the data center is draped in traditional operations. AI demands have disrupted the old ways of doing things, meaning a rapid shift to systems-level design is needed. What can the data center industry learn from aerospace and automotive engineering? We discuss the topic with Peter de Bock, VP Data Center Energy & Cooling at Eaton.
That background shapes how he frames the challenge facing the industry. The data center as it exists today, De Bock argues, is a relic of its telecom roots. There’s a real estate model built around leased space, centralized air conditioning, and an infrastructure used to racks drawing 10 to 40 kilowatts. That model still works adequately for standard enterprise IT. For AI, it does not.
The AI combustion
Nowadays, the jump to hundreds of kilowatts is taking place rapidly. Admittedly, we’re only talking about the most extreme setups which we can collectively refer to as AI infrastructure. Nevertheless, these set the pace for future deployments that may very well scale up to previously unheard of power demands. Nvidia’s Vera Rubin systems will demand up to 600 kW per rack. De Bock notes that this type of wattage brings data center cooling demands per unit volume to that of internal combustion engines. Except combusion engines only have that kind of output intermittently. In an ideal deployment, AI infrastructure is constantly running, saturating any inadequate cooling solution in short order.
The solution is not just to upgrade the cooling. De Bock envisions an end-to-end infrastructure renewal that requires changes from the electrical grid all the way down to the silicon chips themselves. Namely, cooling needs to be done with liquid water rather than air, with a direct copper heat exchanger next to the AI accelerators to maximize efficiency, fully integrated into the plumbing of modern data centers and acting in lockstep with power delivery, again closely looped in with both the local grid as well as power distribution units.
The power density problem
We often hear about Moore’s law being dead. Chips simply aren’t shrinking as fast as before, their footprints are expanding and their power demands are rising beyond historic thresholds for traditional IT designs. De Bock adds context to this development with another, perhaps less well-known observation that highlights this paradigm shift. Dennard scaling, a law of sorts stating that ever-shrinking chips will require a roughly equivalent amount of power, has hit its physical limits, so chipmakers are compensating with larger silicon footprints and by clustering processors tightly together. Millions of cores must communicate constantly, and the thermal consequence of such an approach is unavoidable.
Data centers can no longer be built as generic buildings with standalone air conditioners bolted on, De Bock says. They must be engineered “holistically” as industrial AI factories, he states. As a result, Eaton is collaborating closely with Nvidia and other IT vendors to optimize for future growth as well as the elevated demands even today’s top-of-the-line chips possess. The broader shift, in his framing, is less about the technology itself and more about the discipline required to design for it, moving from telecom engineering to system-level thinking closer to aerospace or automotive manufacturing. In such arenas, close collaboration is far more common to get aircraft and rockets off the ground and cars driving safely and reliably for decades.
Urgency, as it happens, can’t be denied here. De Bock points to a global AI data center power requirement exceeding 150 gigawatts, roughly 15 times the entire grid of New York City. That same urban area was recently compared in prominent research to highlight current CO2 output from AI infrastructure, although companies are also working to reduce their footprint there. At any rate, meeting the demand itself within any reasonable timeframe makes the current, component-by-component approach to data center construction untenable. Eaton is among the companies exploring prefabricated, pre-tested modular compute pods. These are factory-built rather than assembled on site, and they may serve as one response to this constraint. The push toward high-voltage, liquid-cooled designs capable of handling up to megawatts per rack reflects a broader industry reckoning with the same pressure.
What’s wrong with air cooling?
The thermodynamics of air cooling, De Bock says, have never been a good match for this class of hardware. PUE (Power Usage Effectiveness), the standard metric used to measure efficiency masks just how bad the fit is. This places server fan energy on the IT side of the ledger rather than the cooling side, which makes PUE artificially lower, making the facility seem more efficient than it actually is. A facility claiming a PUE of 1.1, De Bock argues, may actually sit closer to 1.4 once that accounting quirk is taken into account. Traditional air-cooled facilities meanwhile spend roughly 40 percent of their total energy just on cooling.
The temperature mismatch compounds the problem. Silicon transistors throttle at around 85°C to 95°C, yet standard facility loops pump refrigerated coolant through at 20°C to 30°C. Chilling fluid to near-room temperature to cool chips that can tolerate 85°C is thermodynamically illogical, our conversation with De Bock underscores once more.
One of the approaches being explored is so-called hot-water cooling, using higher-temperature loops, for example around 45°C input and 60°C output. The mathematics here lead us to large gains with such an approach. Heat rejection for dry coolers follows a cubic relationship, De Bock points out, and thus doubling the temperature differential between coolant and ambient air reduces required fan power by a factor of eight. Rejecting heat at 60° Celsius could eliminate the need for evaporative cooling towers entirely, saving water. At that temperature, the waste heat also becomes viable for municipal district heating networks, aligning with existing European directives.
Measuring the right thing
Not just the instruments need to change, however, as we also need different yardsticks. Replacing PUE as the industry benchmark is, for De Bock, as important as upgrading the physical infrastructure. His proposed alternative is tokens per watt. Rather than measuring how efficiently a facility delivers power to IT equipment, tokens per watt measures actual compute output, or the AI outputs performed, against the grid connection. The metric is difficult to standardize today, but any practical measurement of it would enable IT operations to move towards the kinds of systemic improvements required for the future of AI compute.
The clear issue here is that various engineering teams will invariably be siloed nowadays. For large-scale AI infrastructure, an integrated solution is commonplace thanks to reference designs and hyperscale-level expertise on scaling the infrastructure. For your everyday data center, such a smooth experience is unlikely to occur just yet, especially when the end customer isn’t set in stone for a particular facility. Working out the logistics and finding common ground among those building the physical data centers and those working out the IT side is going to be a challenge.
Conclusion: the many gains to be made
The efficiency gains from a well-designed liquid-cooled AI setup should be worth overcoming such pain points. De Bock puts the figure at 20 to 30 percent more output from the same power envelope. Given that power availability is the primary bottleneck on AI deployment, that difference carries enormous weight. Not to mention the fact that current societal norms around data center expansion focus on exactly the total energy demand. A tricky element here is that the industry risks hiding its own enormous efficiency gains and steady achievements by scaling up as power is available. We’re not exactly sure how that gets solved, but De Bock at least provides us with a type of thinking that may well change how data centers are perceived if communicated clearly to the outside world.
Other challenges are less philosophical and purely based on results. Efficiency on paper without reliability serves little purpose in production environments. It explains why some prefer conservative data center designs even now, even as the gains from a novel approach become apparent. De Bock is categorical about what liquid cooling must borrow from aerospace engineering. There should be structured failure management across three dimensions. Severity is one of them: backups must be inherent to the system design. Occurrence, the failure rates in other words, must be strictly quantified. Detectability is the final cornerstone in De Bock’s view. The predictive monitoring must flag potential issues weeks or months in advance, enabling scheduled maintenance rather than emergency response.
De Bock does not see liquid cooling as a niche reserved for hyperscalers and HPC clusters. Once the economics and space efficiency advantages are clear, and he argues they already are, adoption in standard enterprise environments becomes a matter of when, not whether. That ‘when’ is still critical, given the rapid development of AI hardware, but the requirement to change the ongoing data center paradigms will become clearer with every passing day.