For most of the history of the data center, cooling meant moving air. Cold air in, hot air out, with the room designed around keeping that cycle steady.
That model still works well for a large share of IT. But it was built for a range of rack densities that AI infrastructure has left far behind.
According to Kyle Geib, Application Engineering Manager at Motivair by Schneider Electric, traditional air cooling is sufficient for loads under 20 kW per rack. The latest AI racks, he notes, are reaching “the 200 kW range or more,” with NVIDIA’s Vera Rubin platform cited at 227 kW per rack.
That’s not an incremental step. It’s a tenfold jump past the point where air stops being enough on its own.
The physics behind the limit
The reason comes down to how much heat a fluid can carry.
Schneider Electric puts it in two numbers: water is about 23.5 times more efficient than air at transferring heat, and it absorbs about 3,500 times as much heat as an equivalent volume of air.
To remove the heat of a 200 kW rack with air alone, a facility would need to move enormous volumes of air through a single rack footprint, with the fan power, noise and pressure that come with it. At some point, it stops being a matter of adding more airflow and becomes a matter of using a different medium.
That’s what direct-to-chip liquid cooling does. Cold plates sit directly on the hottest components (GPUs, CPUs and, increasingly, memory and networking) and carry the heat away in liquid, before it ever reaches the room.
Air doesn’t disappear. It changes role.
One of the most useful details in Schneider Electric’s guidance is where the heat actually goes in a liquid-cooled rack.
Direct-to-chip cooling captures 80% or more of the total rack heat load. The rest still leaves the rack as warm air, from components the cold plates don’t cover, such as power supplies, drives and other board-level parts.
At 200 kW, even 20% is a significant amount of heat, comparable to an entire high-density air-cooled rack from just a few years ago.
That’s why many high-density designs pair cold plates with a rear door heat exchanger (RDHx). The cold plates take the bulk of the load, the rear door captures the residual air heat, and the result is what Schneider describes as a “room neutral” solution, where all the heat from the system is rejected to water.
For planning, this means a liquid-cooled AI hall is rarely a 100% liquid environment in practice. It’s a hybrid, and both halves need to be designed, installed and verified.
Coolant temperature is changing the facility, too
The other shift is happening on the facility side of the loop.
In a September 2026 white paper, Water Usage at AI Scale: Insights from a 100 MW Comparative Analysis, Schneider Electric compared four 100 MW designs in Paris and Dallas, from traditional air cooling to liquid cooling at 32°C (about 90°F) and 45°C (113°F).
Two findings stand out:
- In Dallas, switching from air to liquid cooling reduced water use by at least 50%.
- Running liquid cooling at 45°C brought substantial additional savings, because warmer coolant extends the hours in which outside air can reject heat without mechanical chilling.
Warmer coolant means the facility relies less on chillers and more on dry coolers and free cooling. That changes what gets installed outside the white space, not only inside it.
What this means on the floor
When rack densities move into the 200 kW range, the cooling scope of a project changes in a few concrete ways:
Every rack carries two cooling systems
Cold plate connections and, where used, a rear door heat exchanger. Each has its own supply, return, controls and test requirements.
The secondary fluid network becomes core infrastructure
CDUs, manifolds, hoses and quick disconnects are now as critical to uptime as power distribution, and they need the same level of planning, labeling and documentation.
Residual air heat still needs a plan
Whether it’s captured at the rear door or handled by room air units, the 20% that stays in the air has to be accounted for in the design, not discovered at commissioning.
Cooling, power and IT have to land together
A 200 kW rack can’t go live until its liquid, power and network connections are all ready and verified. The schedule has to treat them as one system.
The bottom line
Air cooling isn’t obsolete. It still serves most of the world’s IT, and it still has a role inside every liquid-cooled AI hall. What has changed is that, at 200 kW per rack, air alone can no longer carry the load.
The facilities that handle this transition well are the ones that design for both mediums from the start, and install them with the same rigor.
How is your team approaching residual heat in high-density deployments? We’d like to hear what’s working in your halls. Reach out through weinstall.ai or connect with us on LinkedIn.
Sources
Schneider Electric. Water Usage at AI Scale: Insights from a 100 MW Comparative Analysis. September 2026, as reported by The Register, September 21, 2026.
Geib, K. “Data center thermal management: Top cooling questions answered.” Schneider Electric Blog, August 24, 2026.