CDU + Free Cooling: Redefining Data Center Cooling Efficiency
As AI workloads push data center densities higher, cooling is no longer just an operational challenge; it’s a financial and strategic one. Cutting PUE = cutting utility bills, and the smartest way to do that? Free cooling with CDUs.
Using direct-to-chip liquid cooling paired with dry coolers, we can reject heat to ambient air without mechanical cooling chillers; With a TCS delta of 10°C, 40°C EWT, 2°C CDU HEX approach temp., and 3°C at the dry cooler approach temp.; as long as the dry bulb stays at or below 35°C. That threshold unlocks massive efficiency.
In climate zones 1, 2, and 3 (Hot Humid, Hot Dry, and Mixed Dry) – refer to IECC climate map-, that enables free cooling for over 80% of the year.
The solution? Hybrid cooling systems:
– Free cooling as the default, using CDUs and dry coolers.
– Auxiliary chillers kick in only when ambient DB exceeds 35°C.
– Adiabatic assist extends dry cooler performance in hot-dry zones.
But designing for this efficiency starts with where you build.
Site selection now means more than just power availability, grid reliability, fiber optic infrastructure, and state incentives—climate has become a defining factor. Picking the right ambient conditions can make free cooling viable most of the year, massively improving long-term energy performance.
And the heart of it all is the CDU:
– It must run at low approach temperatures.
– Sustain tight delta-T across the Technical Cooling System (TCS).
– Seamlessly toggle between free and mechanical modes.
AIr2O’s custom CDUs are purpose-built for this kind of hybrid system, maximizing free cooling and minimizing reliance on energy-hungry chillers.
Stop designing around the 20% worst-case conditions. Start designing for the 80% of the year when free cooling wins.
Huge kudos to Billy Roberts / National Renewable Energy Laboratory for developing the US Data Center Climate Map that informs much of this approach.
Follow-Up: Free Cooling, CDUs & AIML Realities
After my recent post on CDUs + free cooling, I had some great public and private conversations. Here are a few takeaways worth sharing:
In China, many hyperscalers are already running 100% free cooling with dry coolers, even at just 4–5°C delta-T. It’s proof that the right climate + design makes it possible.
Challenge raised:
“Maintaining tight delta-T and low approach temps under part-load is tough, especially with AI / ML workloads that spike and shift unpredictably.”
What we’re doing at Air2O:
– Fast-response flow control
– Zone-level sensing
– Oversized HEX with low approach temp, even at part load
Key insight:
Design for the 80% efficient hours, but don’t ignore the 20% worst-case. That’s where hybrid systems (free cooling + auxiliary chiller) offer real flexibility, especially in hot/humid zones.
Future thinking?
AI cooling AI. Expect more adaptive CDU logic, predictive modulation and hashtagsmartmanifolds, and reinforcement learning to handle the complexity of modern workloads.
Thanks to everyone who shared insights, whether in comments or DMs. Let’s keep the conversation open.