Why failure is not an option for CDUs

In the second part of this CDU series, we look at why resiliency still matters as much as ever — and why Cooling Distribution Units (CDUs) have become vital to next-generation data centers.

Missed part one? Read it here for a primer on how CDUs work.

The data center industry has historically been defined by one word: resiliency.

Efficiency, sustainability, and even performance were important but design, build and operation were geared around facilities avoiding downtime at all costs.

Recently however, it has looked like the AI-era was driving a switch in that focus. The attention on ever more powerful GPUs, and denser racks, seemed to have shifted the industry’s balance in favour of outright performance. But as AI becomes more mission critical, and AI inference is embedded into more and more applications, the pendulum looks likely to swing back to resiliency once more.

Why Is Resiliency Making a Comeback in the AI Era?

The fact that resiliency is still critical is exemplified by outages still being a fact of life for operators. According to Uptime Institute’s Annual Outage trends report, outage rates on a per-site basis are declining but the pace of improvement has actually slowed. Power is still the main culprit but thermal problems also cause downtime. A recent high-profile outage at an AWS data center was traced back to an issue with its cooling system.

What Is a CDU and Why Is It Critical?

When it comes to liquid cooling, one of the most critical components is the cooling distribution unit (CDU). Likened to the heart in a human body, the CDU is vital to direct-to-chip liquid cooling at scale. Depending on the type of CDU, it usually sits between the facility’s chilled water loop and the secondary sealed loop that runs coolant directly to server cold plates – a heat exchanger transfers heat between the two circuits without ever mixing the fluids.

That separation matters: facility water is often less pure, less chemically controlled, and circulated through greyspace hardware never designed to sit six inches from a GPU.

The CDU’s secondary loop keeps a tightly monitored, filtered, corrosion-controlled fluid in direct contact with the compute hardware. Its pumps, valves, and controls regulate flow and temperature precisely enough to keep chips within their thermal limits under variable AI workload spikes.

Why Are Data Center Outages Becoming More Complex?

In other words, the CDU is a potential single point of failure for an entire liquid-cooled cluster. If it goes down, so does every rack downstream of it. A single point of failure is obviously not ideal but Uptime Institute believes future failures will also be driven by increasing complexity. “We believe that over time, failures will increasingly not be the result of a single point of failure, but instead be linked to complex interactions between systems, including software, networks, and external dependencies,” says Andy Lawrence, executive director, Uptime Intelligence.

Those existing and emerging causes of downtime are why the entire CDU category has come to be defined not just by raw cooling capacity but by redundancy, serviceability, and control intelligence. Those three factors are also the criteria around which Airsys has designed and built its FluidCool-X CDUs.

What Are the Technical Specs of Airsys’ FluidCool-X CDU?

Airsys’ FluidCool-X CDU line, available in 1000kW and 2000kW models (with a 500kW variant also in the family), is rated for 99.999% uptime and a partial PUE as low as 1.01. Those headline numbers come from a set of engineering decisions rather than a single feature, and it’s worth walking through where the differentiation actually sits.

• Redundancy that’s layered. FluidCool-X runs backup pumps and drives that kick in automatically if one fails, alongside two separate sensor systems watching temperature and pressure at the same time. Add an optional second power feed with automatic switchover, and you get several layers of backup instead of one, which is exactly what the strictest data center uptime standards require. Widespread use of stainless steel helps to prevent any deterioration alongside turbidity and pH sensors. “We put a lot of focus on ensuring that the quality of the fluid inside of the CDU is always maintained at the highest level and purity,” says Sean Corcoran, Airsys’ vice president of sales – S. PowerOne.
• A pump built for headroom. The pump can push up to 65 m³/h of coolant at 4 bar of pressure, noticeably more muscle than comparable competitor CDUs. That extra push supports long pipe runs and complex multi-rack setups without strain, and it leaves real spare capacity, up to 60% extra on the 1MW unit and about 25% on the 2MW unit, so a sudden spike in server load doesn’t push the system to its limit.

How Does FluidCool-X Handle Serviceability and Thermal Control?

• Smarter efficiency. The heat exchangers, sourced from ALFA LAVAL and SWEP, are tuned to pull heat out within a tight 4°C margin, so they transfer more heat for the same amount of water flow and need less pumping energy to do it. On top of that, the pumps continuously adjust their speed to match actual demand rather than running flat-out all the time, cutting energy use by roughly 30%.
• Repairs without a shutdown. This is where a lot of competing designs fall short in practice. A backup filter keeps things running while the active one is swapped out, valves on either side of every major part let technicians isolate and remove it without draining the whole system, and quick-release fittings mean parts come apart in minutes rather than hours. Ninety percent of the unit can be serviced from the front, and it’s also fully accessible from the back, making it easy to install side-by-side with server rows.
• Monitoring that catches problems early. Beyond the standard temperature, pressure, and flow sensors (all doubled up for backup), FluidCool-X keeps a constant watch on the coolant’s pH and conductivity, an early warning sign of corrosion or contamination well before it can damage expensive cold plates. A sensor strip running the length of the drip tray catches leaks early. It’s all managed through a touchscreen control panel which uses Modbus, BACnet and SNMP.
• Built to grow the way AI infrastructure actually grows. The unit is sized to fit standard server rows without any retrofitting, multiple units can be linked so a spare automatically kicks in if needed.

Where Are FluidCool-X Units Manufactured and Supported?

None of this technology exists in a vacuum. The FluidCool-X line is supported by three geographically distributed manufacturing sites, in the U.S., Europe and Asia providing over 2GW of combined annual capacity and supply-chain resilience against regional disruption. That’s paired with 24/7/365 technical support, spare parts warehouses across all three global regions, and factory-trained field service technicians, the kind of operational backbone that matters as much as the hardware itself.

What’s Next for CDU Design in High-Density Data Centers?

What’s next for CDU design depends on a range of factors but bigger units seems to be one direction with the potential for Facility Distribution Units (FDUs) built out in the grey space. However, as Sean Corcoran from Airsys describes, it very much depends on the customer’s needs. “People want to customize CDUs to fit exactly what they want to accomplish and that ultimately depends on the customer,” he says.

As AI and HPC deployments keep pushing rack densities higher, the CDU is no longer a supporting component sitting quietly in a mechanical room. It is functionally as mission-critical as the servers it’s cooling. Failure can’t be ruled out, but Airsys CDUs are designed to make it an increasingly unlikely option.

Frequently Asked Questions

What does a CDU do?

A cooling distribution unit (CDU) sits between a data center’s facility chilled water loop and the secondary sealed loop that runs coolant directly to server cold plates. A heat exchanger transfers heat between the two circuits without mixing the fluids, keeping less-pure facility water away from direct contact with compute hardware.

Why is a CDU considered a single point of failure?

Because every rack downstream of a CDU depends on it for cooling, if the CDU goes down, the entire liquid-cooled cluster it serves goes down with it. That’s why redundancy, serviceability, and control intelligence — not just raw cooling capacity — have become the defining criteria for CDU design.

What uptime and efficiency does Airsys’ FluidCool-X CDU deliver?

FluidCool-X is rated for 99.999% uptime and a partial PUE as low as 1.01.

What sizes does FluidCool-X come in?

FluidCool-X is available in 500kW, 1000kW, and 2000kW models, with spare pump capacity of up to 60% on the 1MW unit and about 25% on the 2MW unit to absorb sudden load spikes.

How does FluidCool-X reduce energy use?

Its heat exchangers pull heat within a tight 4°C margin, and its pumps continuously adjust speed to match actual demand instead of running at full output, cutting energy use by roughly 30%.