The short version: Most guides to data center cooling systems read like a menu—air, rear-door, direct-to-chip, immersion—and tell you to pick one. At AI rack densities that choice is already made for you by the silicon, and the real project is the facility-side plumbing the menus skip: coolant distribution, water chemistry, leak risk in a live hall, and floor loading. The cooling technology is the easy part. The loop behind it is where deployments stall.
The hardware picks the cooling, not the operator
For two decades, choosing a cooling approach was a genuine design decision. You sized CRAC or CRAH units, laid out hot and cold aisles, maybe added containment, and air did the job. That era is closing at the high end—not because liquid is fashionable, but because air physically runs out of room.
Benchmarking from Energy Solutions Intelligence puts numbers on the ceiling: above roughly 15–20 kW per rack, airflow demand exceeds what standard perforated floor tiles can deliver. Active rear-door heat exchangers stretch that to about 25–35 kW per rack, and high-capacity in-row systems to 35–50 kW. Past 50 kW per rack, the same analysis finds immersion and direct-to-chip economics pull ahead. Modern AI training racks built around dense GPU accelerators sit well into—and beyond—that top band, which is why the industry-wide pivot to liquid is happening on the hardware vendors’ timeline, not the facility team’s.
The practical consequence: for an AI hall, “air vs. liquid” is not a question you answer. The accelerator answers it for you. What’s left to engineer is everything behind the cold plate—and that is exactly the part the popular explainers gloss over.
What data center cooling systems guides leave out
Search “data center cooling systems” and the top results are catalogs. TechTarget’s widely cited rundown, for instance, walks through CRAC/CRAH units, in-row cooling, containment, free and adiabatic cooling, direct-to-chip, immersion, rear-door heat exchangers, and even geothermal and thermal-wheel (“Kyoto”) designs. It is a thorough taxonomy of what the technologies are.
What it does not cover—and what nearly every ranking page omits—is the facility infrastructure that actually makes a liquid deployment work: coolant distribution units (CDUs), the split between primary and secondary loops, water quality and filtration, leak detection inside a populated room, and the floor loading a filled immersion tank imposes. The menu tells you immersion hits a great PUE. It does not tell you what you have to build to get there.
That omission is the whole problem, because the technology choice is the cheap, fast part of the project. The loop is the slow, expensive, risk-bearing part.
The loop behind the rack is the real project
Drop a direct-to-chip manifold or an immersion tank into a hall designed for air, and the rack is the least of your worries. A CDU has to sit between the facility water system and the IT loop, sized to the new heat load and isolated so a problem on one side cannot contaminate the other. The secondary loop needs its own pumps, redundancy, and commissioning.
Then there is the fluid itself. Industry coverage of liquid cooling consistently flags water and coolant management as a top adoption barrier: improper maintenance leads to sediment buildup, contamination, and even bacterial growth, and wetted components often have to be specified in corrosion-resistant materials—bronze, cupronickel, titanium—to stay compatible with the coolant. None of that appears on a “types of cooling” list, yet all of it determines whether the system runs for years or fouls in months.
Leak detection moves from nice-to-have to mandatory. You are now running liquid above and around live, energized equipment; a slow leak you cannot see is a different category of risk than a warm aisle. Floor loading matters too—immersion tanks full of dielectric fluid are heavy, and plenty of raised floors were never rated for it. These are mechanical and civil-engineering problems, not IT-procurement problems, and they are where schedules actually slip.
You’re going to run a mixed hall for years
Almost no operator flips an existing facility to liquid wholesale. The realistic path is coexistence: a liquid-cooled zone for the AI racks alongside the air-cooled hall that still runs everything else. That hybrid state is harder to operate than either pure mode.
You end up maintaining two cooling control regimes, two failure modes, and two maintenance skill sets under one roof. Some of the air-cooled footprint becomes stranded—floor space and even power you cannot fully use because the thermal envelope was built for a density the new workloads blew past. And the liquid loop has to be commissioned and brought online without taking the production air-cooled hall down, which constrains when and how you can do the work. The transition, not the end state, is the engineering challenge most teams underestimate.
Warmer water quietly rewrites the economics
There is a genuine upside that makes the effort worth it, and it also lives in the loop rather than the rack. NVIDIA says its newest AI servers accept coolant temperatures up to 45°C (113°F). That sounds like a footnote; it is the opposite. Warmer return water can reject heat through dry coolers in many climates instead of mechanical chillers—NVIDIA describes chiller-less operation that can cut a facility’s cooling-related water draw from roughly 2.6 million gallons per megawatt per year toward near zero, and notes that raising chiller-plant temperature by even one degree trims cooling energy by about 4%.
The efficiency ladder is real, too. Energy Solutions Intelligence’s deployment benchmarking puts traditional air cooling at a PUE of about 1.50–1.80, rear-door exchangers at 1.20–1.40, direct-to-chip at 1.15–1.30, and single- and two-phase immersion as low as 1.02–1.08. But every one of those numbers assumes the facility loop can actually supply and absorb water at the right temperature and flow. Warm-water operation is a property of the plant you build, not the cold plate you buy.
What operators should plan for
The takeaway is a reframe. Stop shopping for a cooling “type” and start engineering a transition. For an AI buildout, the cold plate or tank is close to a commodity decision; the differentiated, schedule-driving work is the CDU strategy, the secondary-loop design, the water-quality program, leak detection across a live hall, floor-loading verification, and a credible plan for running air and liquid side by side while you migrate.
Adoption data supports treating this as an unfolding transition rather than a switch. Energy Solutions Intelligence pegs immersion at under 2% of the global cooling market by IT load in 2025, projecting 12–18% by 2030—rapid growth, but a long coexistence period in absolute terms. The operators who come out ahead will not be the ones who picked the “best” cooling technology off a list. They will be the ones who treated the loop, the water, and the mixed-mode transition as the actual project from day one.