The Cooling Imperative for AI Data Centers
Modern AI workloads have rewritten the rules of data center thermal management. GPU clusters powering large language models and high‑performance computing (HPC) now routinely exceed 100kW per rack, a density that air cooling alone cannot handle. For any facility scaling AI, a high‑performance CDU for AI Data Center is no longer optional—it is the core of a viable thermal strategy. Gerchamp’s Bluearth Series answers this challenge with a liquid to air (L2A) sidecar design that removes the single biggest obstacle to liquid cooling: facility reconstruction.
Bluearth CDU: Supporting NVIDIA GB300 NVL72, Vera Rubin NVL72, and Ultra HPC Workloads
Selecting a CDU for AI Data Center that aligns with current and future GPU roadmaps is critical. The Bluearth CDU4175 L2A 175kW is purpose‑built for this exact challenge. A single unit operates independently to support a fully loaded NVIDIA GB300 NVL72 at peak power, making it an ideal thermal partner for current‑generation AI clusters. When your workloads graduate to the higher thermal demands of NVIDIA Vera Rubin NVL72, two Bluearth units deploy in parallel, doubling the heat rejection capacity and extending headroom all the way to the planned Vera Rubin Ultra. Throughout every configuration, the ATD=10 specification ensures that even with a 35°C hot aisle, coolant is delivered at the precise temperature required to keep Rubin GPUs below throttling thresholds during sustained HPC training. This thermal precision allows an AI data center to run the most demanding workloads without performance degradation, backed by a modular CDU for AI Data Center architecture that scales in lockstep with your GPU roadmap.
Sidecar CDU for AI Data Center: Zero Construction Deployment
The Bluearth Sidecar CDU4175 L2A 175kW for AI Data Center delivers liquid cooling without any infrastructure modification. Traditional approaches demand chilled water plants, floor reinforcement, and months of pipework. This liquid to air unit eliminates all of that. It rolls into place beside a rack, connects to the cold plates, and begins rejecting heat directly into the existing room air. No facility water, no outdoor condensers, no construction. Full deployment to 175kW/52U of cooling at ATD=10 is completed in a few days, making this CDU for AI Data Center a practical tool for HPC environments that need to scale now.
How L2A Differs from L2L
| Dimension | Liquid‑to‑Air (L2A) sidecar | Liquid‑to‑Liquid (L2L) in‑row |
|---|---|---|
| Heat sink | Room air via integral radiator | Facility chilled water loop |
| Facility water required | No | Yes |
| Cooling tower / dry cooler | Not required | Required |
| Construction scope | Roll in and connect | Pipework, pumps, floor work |
| Deployment time | Days | Weeks to months |
| Best fit | Hybrid and retrofit halls | Purpose‑built high‑density halls |
For retrofit and hybrid sites, this is the difference between a project that starts this quarter and one that starts next year. Rack‑scale system vendors list L2A sidecar options specifically for facilities without a cooling tower or water supply, confirming the use case is mainstream.
Hybrid Deployment: Air and Liquid in the Same Room
Servers without a liquid cooling requirement still need air handling, so whole‑hall conversion is rarely the right first move. A CDU that rejects heat into room air lets liquid‑cooled GPU racks and conventional air‑cooled servers share one room and one thermal envelope. Capacity can then be added incrementally as GPU racks arrive, existing air handling is not stranded, and construction comes off the critical path of getting compute online.
Redundancy, Hot‑Swap, and Intelligent Control
A CDU for AI Data Center must protect the uptime of the entire GPU cluster. The Bluearth 175kW/52U unit features dual redundant pumps in a lead‑lag configuration with automatic failover. If one pump degrades, the backup assumes full flow instantly, preventing any interruption to HPC jobs. Both pumps and filters are hot‑swappable, enabling replacement without shutting down the cooling loop. The integrated controller supports two operating modes—constant flow for steady‑state AI training and constant pressure for fluctuating workloads—while logging all parameters and providing remote visibility. This combination of reliability and control makes the liquid to air sidecar a trustworthy asset for any AI data center.
Extreme Density in a Compact Footprint
Space in an AI data center is valuable. At just 900mm wide, this CDU for AI Data Center delivers 175kW/52U of cooling at ATD=10, or 125kW/52U at ATD=5, without consuming additional white space. This density allows operators to fully populate GPU racks to their electrical limits, eliminating stranded capacity and maximizing every square foot. A single unit handles a fully loaded NVIDIA GB300 NVL72 rack; for Vera Rubin NVL72 and beyond, two units scale the cooling linearly. The liquid to air design means no outdoor heat rejection equipment is required, preserving valuable facility real estate.
The Practical Bridge from Air to Liquid
The Bluearth CDU4175 L2A 175kW is built for the real‑world data center that was never designed for liquid. Its liquid to air architecture enables hybrid deployments where liquid‑cooled AI supercomputer racks operate alongside traditional air‑cooled servers within the same room. This approach eliminates the need to overhaul existing air handling infrastructure, transforming the transition from air to liquid into a phased, low‑risk evolution. For HPC teams, this means more time spent on compute and less on construction. The Bluearth Series proves that a CDU for AI Data Center can be both high‑performance and practical, unlocking the full potential of AI supercomputing without asking operators to rebuild their facilities.
FAQ: Choosing a Coolant Distribution Unit for AI Data Center
Does a CDU require facility water to operate?
Not if it is liquid‑to‑air. An L2A sidecar CDU rejects heat into room air through an integral radiator, so it needs no chilled water connection, cooling tower, or outdoor condenser. Liquid‑to‑liquid units do require a facility water loop.
How much cooling capacity does one AI rack need?
Plan against the platform’s maximum, not its average. NVIDIA’s GB300 NVL72 Enterprise Reference Architecture specifies up to 142 kW for a full rack. Because each generation typically raises rack power again, sizing to current draw alone tends to produce stranded capacity.
What does ATD mean, and why does it matter?
ATD is the gap between coolant supply temperature and the temperature of the medium rejecting the heat. Capacity ratings are comparable only when ATD is stated — the same unit can be rated 175 kW at ATD = 10 °C and 125 kW at ATD = 5 °C.
The Bottom Line
Specify a coolant distribution unit for AI data center racks on four criteria: capacity at a stated ATD, redundancy that survives a pump failure without operator action, a footprint that fits the hall you actually have, and a heat‑rejection path that does not depend on construction you cannot schedule. L2A sidecar designs meet all four for hybrid and retrofit sites, which is why they have become the practical entry point for liquid cooling in buildings that were never designed for it.
Compare the full specification on the Gerchamp Bluearth CDU4175 L2A 175 kW product page.