In the not so distant future, data centers are going to be launched into orbit, where they’ll be cooled by the vacuum of space. Science fiction becomes reality, the furnace-hot servers will dump heat into the void above Earth while sending compute data down to high-altitude ground stations via fast 2.5 Gbps laser links. It’s all just a matter of time.

Until that day, those LEOs (Low Earth Orbit) data centers floating leisurely hundreds of miles above Earth, we’re stuck with terrestrial facilities. Cue the bad news. These sprawling server farms need Industrial-scale cooling, and those systems demand huge amounts of electrical energy to stop the silicon in those server blades from frying. That means auxiliary equipment filling floors with chillers and pumps, heat exchangers and pipes, raising energy bills.

[Learn more about hydraulics and pneumatics at GlobalSpec]

Bringing the bad news home, where it hits companies right in the wallet, here’s a worrying statistic to consider. The AI revolution has triggered a 415 terawatt hours (TWh) annual energy drain, a figure that represents roughly 1.5% of the entire planet's electricity consumption, as of 2024. To put that in perspective, we are currently burning through enough juice to power Pakistan just to keep our models dreaming. Conclusion: without being overly dramatic, unless changes are made in the existing cooling infrastructure, something is going to give.

The hydraulic shift to liquid-cooled data centers

This has to be more than a stopgap measure. The launch of data centers into space is still in its infancy. Also, thanks to AI training models, regardless of whether the AI revolution is a bubble or here to stay, that energy footprint is growing. Consequently, air cooling won’t cut it anymore, not with thousands of NVIDIA or AMD GPUs at work simultaneously, so full liquid submersion has arrived to save the day.

Now, let’s be clear. This is a different solution compared to the liquid-cooling systems found built into high-end desktop computers. The heat in those is conveyed to a case radiator by cold water pumped in tubes. If that sounds familiar, it’s the same principle used to cool car engines. On a whole other level here, hot server silicon is being submerged in non-conducting fluids, bathing every wattage-hungry component.

Just as a sidebar, even the logistics of finding an adequate fluid medium is headache-inducing, but that’s an off-topic subject for another day. Water glycol is always a good candidate, but proprietary single-phase coolants are more practical as GPUs reach 400W and 500W outputs, like the NVIDIA A100. With that aside stated, the hydraulics and pneumatics considerations continue at a fast pace.

Despite the structural demands and added engineering complexity, liquid cooling offers a far more capable answer than traditional air systems. Where fans struggle with dense, spike-prone AI workloads, fluid-based cooling delivers faster heat removal, tighter thermal control, and better long-term efficiency. The ability to respond instantly to fluctuating compute cycles—without flooding the room with a vibration-inducing roar of energy-hungry airflow-based cooling—makes these systems better suited to modern data center realities. True, while reinforced floors and fluid management add upfront challenges, the payoff comes in stable performance, higher rack densities, and fewer thermal bottlenecks.

Managing the physical mass of liquid computing

Hydraulic and pneumatic solutions involve new, from the ground up equipment builds and legacy retrofits, which replace fans and ducts. The hydraulic backbone, whether radiator cooled or immersion cooled, involves high-efficiency pumps and scores of specially designed server farm fluid fittings. Not least of which would be more advanced, scaled up cold plate cooling units. That’s part of the direct-to-chip tubing method, of course. For immersion technology, unique challenges await. They stem from the fact that these liquid baths are going to get heavy.

It’s not, therefore, a plug and play upgrade path, not for immersion data centers. These tanks require storage space. Floors need reinforcing to handle the extra fluid density. If maintenance needs to be done, time is required first to drain the circuitry housing before the silicon can be handled. A network of pneumatic regulators accompanies the fluid systems, silently controlling dynamic cooling. With LLM model training workloads and dataset fine-tuning causing thermal spikes, pneumatic controls then rapidly offset the limitations of hydraulic latency.

It’s estimated that several things need to happen to make fluid cooling a match for rapid AI uptake. To keep up with the dynamic heat buildups caused by fluctuating compute cycles, stop expensive silicon burnouts and thermal throttling, efficient variable-speed hydraulic pumps are a must, plus the accompanying pressure-compensated flow control valves and relief valves. All of these are essential if latency issues are to be minimized when dealing with instant-call GPU energy spikes. Likewise, expect fast purges and fine-control system buoyancy via powerful state-of-the-art air management, built to instantly regulate flow, keeping fluid surges in check.

The future of frictionless compute is submerged

If AI stressed data centers aren’t yet going into orbit, they’re still undergoing something of an atmospheric change. For the next step in their evolution, moving from on-chip water blocks, they’re dropping under the surface of non-conductive “dielectrics.” There are scores of research papers on the topic. Some are creating real-world answers to this silicon baking problem, while others are confined to the imagination.

Of some interest, instead of space and designed submersion technology, what if the energy sucking digital farms sink underwater? All eyes are on Project Natick, a Microsoft-run venture that’s placing a steel-sealed data center in the cold waters of the North Sea just off the coast of the Orkney Islands in Scotland. It’s a fine feasibility project, but large-scale deployment could still be years away as all of the logistics problems areRobert Beharie, Sooth Clett, Swona, Orkney Islands (24 May 2012), Geograph / Wikimedia Commons, CC BY-SA 2.0. dealt with by engineers with experience in offshore design, perhaps with a pinch of submarine construction as well.

The controlled application of high-technology cooling fluids seems, at least for the moment, the more realistic answer to high processor workload heat. Supported by the latest generation of hydraulics and pneumatics flow control, the equipment consumes far less energy, is quieter, and it even promises to be open to retrofitting. Out with the fans and ducts, in with the hydraulic pumps and tubing, these setups promise to respond quickly to fluctuating AI workloads while keeping thermal stresses in check.