AI infrastructure is scaling at a pace most facilities aren’t ready for. Rack densities are climbing. Power demands are spiking. But the real choke point, the issue that could cripple performance, delay deployments, and erode margins – isn’t compute. It’s cooling.
And more specifically, it’s liquid cooling.
We’ve seen firsthand how this plays out. After 400,000+ unit hours operating a 100% liquid-cooled AI facility, here’s the truth: most operators and vendors still treat liquid cooling as an IT project, not infrastructure. That’s a mistake.
The Mismatch Between Cooling and AI
Modern AI workloads: training, inference and real-time processing, create intense, uneven heat profiles. A 70kW rack today might need 120kW tomorrow, depending on the workload. But most data centers are built to move air, not water.
Air cooling was built for 5-10kW racks. AI racks are now pushing 50-120kW. And nearly 80% of data centers are running on air cooling systems that were never designed for this kind of heat. Performance throttling, thermal shutdowns, and energy waste are already showing up.
And retrofitting for liquid isn’t just about adding CDUs. It’s about rethinking everything:
- Design and installing the hydronic system
- Pressure management across the hydronic loop
- Leak containment and prevention
- Coordination between IT and mechanical teams
- Control systems that can respond to real-time load fluctuation
We’ve seen projects with GPUs on order and nowhere to put them, because cooling was an afterthought. In AI infrastructure, cooling delays are deployment delays.
Vendor Assumptions Are Failing in the Field
Too many liquid cooling systems on the market are still designed like IT gear: in-rack, localized, self-contained. They assume:
- Perfect installs
- Ideal operating conditions
- Zero operator error
But in real-world deployments, that’s not reality. We’ve had customers leave fill valves open for hours. We’ve seen air introduced mid-operation. We’ve had to support facilities with no mechanical gallery and 100-foot runs between CDUs and racks. These aren’t exceptions—they’re standard field conditions.
A cooling strategy that can’t handle these variables isn’t just fragile, it’s a liability.
Cooling Is Now Part of Your Risk Profile
Here’s the shift no one wants to talk about: liquid cooling doesn’t just deliver performance, it introduces risk. It touches power, IT, controls, facility design, operations, and serviceability.
We’ve learned that you can’t treat CDUs and liquid cooling like a checkbox. They need to be engineered into the system from the start. And operators need:
- Real commissioning support
- Training based on actual failure scenarios
- Designs that prioritize maintainability over marketing specs
If you’re not building for failure modes, you’re not building for AI.
Modern cooling systems need to do more than reject heat. They must:
- Support 100kW+ racks
- Scale facility-wide, not just rack-by-rack
- Be deployable in weeks, not quarters
- Minimize energy and water waste
- Integrate into existing spaces without full gut-renovations
Our Point of View on Liquid Cooling and AI Infrastructure
Liquid cooling isn’t the weak link because it doesn’t work. It’s the weak link because it’s being deployed without the rigor of infrastructure thinking.
At Nautilus, we’ve spent the last four years operating, iterating, and evolving our cooling platform in production. We’ve designed around air entrapment, transient thermal loads, and operator mishandling. We’ve built CDUs that are part of the mechanical plant, not an afterthought bolted to the rack.
If we want AI to scale, we have to elevate cooling to the same level as power and compute. Not just in specs, but in accountability.
Cooling isn’t secondary. It’s the first domino. You can’t scale AI compute without scalable AI cooling.
Because no one brags about PUE if their racks are melting.
