Data center heat loads are rising faster than at any point in the industry’s history. A traditional enterprise rack drew 5 to 10 kW. Today’s AI training and inference racks, packed with GPUs and high-bandwidth memory, draw 40 to 150 kW, and roadmaps already point higher. At these densities, air cooling runs out of headroom: there is simply not enough airflow or fan power to move that much heat through a rack.
Direct-to-chip liquid cooling has become the answer. Cold plates mounted directly on processors pull heat into a circulating coolant, which carries it to a Coolant Distribution Unit (CDU) for rejection. Between the CDU and every cold plate in the hall sits the secondary fluid network (SFN): the headers, branches, valves, and connections that deliver coolant to each rack and bring it back.
The secondary fluid network rarely gets the attention that CDUs and servers do, yet it decides whether the whole system works. If flow is not even some GPUs will throttle some will be overcooled. If the piping is dirty, particles clog cold plate micro-channels. If a joint leaks, coolant reaches live IT equipment. Each of these failures starts with decisions made during SFN design and fabrication.
This guide walks through those decisions in the order an engineer would make them. It covers where the SFN sits in a liquid cooling architecture, how to size flow and pipework, how to lay out and balance the network, which materials and fabrication methods keep it clean, and how to test and commission it before the first server goes live.
What is a secondary fluid network?
A liquid-cooled data centre does not run a single coolant loop from the chiller to the chip. It runs two separate loops that meet at the CDU’s heat exchanger. Understanding that split is the starting point for designing a secondary fluid network.
The first is the Facility Water System (FWS), also called the primary loop. It carries heat from the CDUs out to chillers, dry coolers, or cooling towers. This loop is large, runs across the building, and often uses water whose quality is harder to control. In RexEdge terms, the piping on this side is the Primary Fluid Network (PFN).
The second is the Technology Cooling System (TCS), or secondary loop. It carries a treated, inhibited coolant, commonly PG25 (25% propylene glycol in water), from the CDU to the IT equipment and back. Because this coolant flows through cold plates sitting on processors worth thousands of dollars each, it must stay clean, chemically stable, and free of air.
The secondary fluid network is the physical pipework of that TCS loop. It typically includes supply and return headers running along the rows, branch lines and rack drops, isolation and balancing valves, high-point air vents and low-point drains, and the quick-disconnect couplings that connect to each rack manifold. In AI halls, it is often integrated into overhead structures alongside busways and cable trays.
Separating the two loops is what protects the IT hardware. The CDU’s heat exchanger transfers heat without mixing fluids, so the secondary fluid network can be kept as a small, closed, tightly controlled circuit. That separation only works, however, if the SFN itself is designed and built to the same standard of cleanliness and reliability as the equipment it serves.
Sizing the sFN
Every secondary fluid network design begins with two numbers: the heat load of each rack and the temperature rise the coolant is allowed to rise across it. Together they set the flow rate through the basic energy balance, where Q is heat load in kW, ṁ is mass flow in kg/s, cp is the coolant’s specific heat, and ΔT is the difference between supply and return temperature:
Q = \dot{m} \cdot c_p \cdot \Delta T
Consider a single 100 kW AI rack cooled with PG25, which has a specific heat of roughly 3.9 kJ/kg·K and a density of about 1,025 kg/m³. With a design ΔT of 10 K, the required mass flow is 100 ÷ (3.9 × 10), or about 2.56 kg/s. That equals roughly 2.5 liters per second, or about 150 liters per minute. This lines up with the common planning figure of around 1.5 L/min per kW for cold-plate cooling, though the final value should always be confirmed against the server OEM’s cold plate specification.
With flow known, pipe diameter is set by velocity. Pipes sized for very low velocity are larger, heavier, and more expensive than they need to be. Pipes sized for high velocity suffer erosion at fittings, generate noise, and demand more pump energy. Typical practice keeps rack branches around 1.0 to 1.5 m/s and main headers around 1.5 to 2.5 m/s. For our 100 kW rack at 1.5 m/s, the rack drop needs an internal diameter of about 46 mm, so DN50. A row header serving ten such racks carries 1,500 L/min, which at 2.0 m/s needs roughly 126 mm internal diameter, so DN125 to DN150.
The next check is pressure drop. Straight-pipe losses are calculated with the Darcy–Weisbach equation, then losses through elbows, tees, valves, strainers, quick disconnects, rack manifolds, and the cold plates themselves are added. Glycol mixtures matter here: PG25 is roughly twice as viscous as water at typical TCS temperatures, so water-based friction charts will underestimate the loss. The total must sit within the available head of the CDU’s pumps, with margin for filters that load up over time.
Supply temperature is the final variable. ASHRAE defines liquid cooling classes, from W17 to W45 and W+, by the facility supply temperature. Running warmer, in the 30 to 45°C range, allows more hours of free cooling and lower energy costs but leaves less thermal margin at the chip. That means flow rate, ΔT, and supply temperature have to be chosen together rather than one at a time.
Getting these numbers right early saves expensive rework later. Undersized headers cannot be fixed without draining the loop, and oversized pumps waste energy for the life of the facility. A careful sizing exercise, checked against real cold plate data, is the foundation for everything that follows.
SFN topology and hydraulic balancing
The aim of secondary fluid network layout is simple to state and harder to achieve: every rack should receive its design flow, whether it is the first rack on the header or the last. The topology chosen at the start largely determines how much effort that takes.
The simplest layout is direct return, where supply and return headers run side by side and each rack connects across them. It uses the least pipe, but racks closest to the CDU see a shorter path and therefore more flow, while distant racks are starved. Direct return can work well, but only with properly selected balancing valves at every rack drop.
Reverse return solves the problem with geometry. The first rack to be supplied is the last to be returned, so every path through the network has roughly the same length and pressure drop. Flow divides naturally and evenly, at the cost of an extra run of return pipe. For long rows of identical high-density racks, that extra pipe is often a good trade.
Ring or looped headers add resilience. Because coolant can reach any rack from two directions, a section of the header can be isolated for maintenance or modification while the rest of the hall keeps running. This approach is common where concurrent maintainability is a design requirement, and it pairs well with CDUs installed in an N+1 arrangement.
Whatever the topology, two principles keep flow even. First, size headers generously so that the pressure drop along the header is small compared with the drop across each rack. Second, fit pressure-independent control valves or flow-limiting valves at rack drops so flow stays steady as racks are added, removed, or throttled. Finally, plan for growth with valved, capped tap-off points for future racks, and place air vents at every high point and drains at every low point, because trapped air is one of the most common causes of poor cold plate performance.
Materials, cleanliness, and fabrication
Cold plates contain micro-channels that can be a fraction of a millimeter wide. A few particles of weld slag, grinding dust, or mill scale are enough to restrict them, and a restricted cold plate means a throttling GPU. That is why material selection and fabrication cleanliness matter as much in a secondary fluid network as hydraulic design does.
316L stainless steel is the preferred material for SFN piping. It resists corrosion in glycol-based coolants, remains compatible with copper cold plates when the coolant’s inhibitor package is maintained, and its low carbon content reduces the risk of corrosion in the heat-affected zone around welds. Galvanized steel, zinc-bearing components, and uncontrolled combinations of dissimilar metals should be avoided because they can react with inhibitors and shed particles into the loop. Every wetted component, including gaskets, valve seats, seals, and flexible hoses, must also be rated for the coolant chemistry and operating temperature.
How the pipe is joined is just as important. Automatic orbital TIG welding produces repeatable, full-penetration joints with smooth internal weld beads, which manual welding struggles to match consistently. Each weld must be made with an inert gas back-purge. Without it, the inside of the joint oxidizes into a rough, crusty surface often called “sugaring,” which traps debris, sheds particles, and becomes a starting point for corrosion.
Once fabricated, spools are degreased, pickled, and passivated, for example, in line with ASTM A380 and A967. Pickling removes heat tint and embedded iron from welding and handling, while passivation restores the thin chromium oxide layer that makes stainless steel corrosion-resistant. Skipping this step leaves the inside of the pipe vulnerable from the day the loop is filled.
Final inspection and assembly should take place in a controlled environment. An ISO Class 8 cleanroom, as defined by ISO 14644-1, limits airborne particles while spools are inspected, fitted and sealed. Every open end is then capped and bagged until installation, so the pipe that arrives on site is as clean as the pipe that left the factory. Weld logs, material certificates and inspection records travel with each spool, giving full traceability.
At RexEdge, every secondary fluid network is designed, fabricated, passivated and assembled in-house, with ISO Class 8 cleanroom assembly. Keeping each step under one roof removes the hand-offs between fabricators, cleaners and installers where contamination and quality gaps usually creep in.
Testing, flushing and commissioning
A secondary fluid network should be proven leak-tight and clean before it ever meets a server. Testing is not a single event at the end of the project but a sequence that starts in the factory and finishes with a balanced, documented loop at handover.
In the factory, each spool is pressure tested above its design pressure, commonly at 1.5 times, under the governing piping code. Hydrostatic testing with clean water is the usual method, with pneumatic testing used where water cannot be introduced. For critical joints, helium leak detection or pressure-decay testing can find leaks far smaller than any visual inspection. Every result is recorded against the spool’s unique ID, so the test history follows the pipe to site.
Once installed, the assembled network is pressure tested again as a complete system. Rack connections are isolated or bypassed during this stage, so that a leak or overpressure event cannot reach IT equipment. Any joint made on site, such as a flanged or grooved connection between factory spools, is checked with particular care.
The loop is then flushed through temporary bypass jumpers rather than through the cold plates, at velocities higher than normal operation, to carry out any debris introduced during installation. Flushing continues until temporary strainers stay clean and water samples meet the agreed particle limits. Only then is the loop filled with pre-mixed, inhibited coolant and vented at every high point to remove trapped air.
Before handover, a coolant baseline is recorded: pH, conductivity, inhibitor concentration and particle count. Flow to each rack is measured and balanced against design, and the CDU’s own filtration, often 50 microns or finer, and monitoring take over for day-to-day operation.
The handover pack should bring all of this together: as-built drawings, pressure and leak test certificates, weld records, material certificates and the coolant baseline. This documentation is what makes later expansion, troubleshooting and audits fast and low-risk, and it is the clearest evidence that the network was built right.
SFN design checklist
- Heat load per rack and design ΔT agreed with the IT and server teams
- Flow per rack confirmed against the cold plate OEM’s specification
- Pipe sizes checked for velocity limits on branches and headers
- Total pressure drop within CDU pump head, with filter margin
- Topology and balancing method chosen for even flow and maintainability
- 316L stainless steel and coolant-compatible seals for all wetted parts
- Orbital welding with back-purge, plus weld records
- Passivation, ISO Class 8 assembly, and capped, bagged spools
- Factory and site pressure tests, flushing and coolant baseline documented
Conclusion
The secondary fluid network is a small share of a data centre’s capital cost, but it carried every watt of heat from the chip to the CDU. When it is designed and built well, nobody notices it. When it is not, the consequences show up as throttled GPUs, clogged cold plates, leaks and unplanned downtime.
The principles are straightforward. Size flow from real heat loads and cold plate data. Keep velocities and pressure drops within sensible limits. Choose a topology that delivers even flow and allows maintenance without shutting down the hall. Build in 316L stainless steel with orbital welding, passivation and cleanroom assembly. Then prove the result with pressure testing, flushing and a documented coolant baseline.
As rack densities keep climbing, the margin for error in the secondary fluid network keeps shrinking. Facilities designed today for 100 kW racks may need to support far more within a few years, so building in spare capacity, tap-off points and clear documentation is as important as meeting today’s load.
RexEdge designs, fabricates and tests secondary fluid networks in-house, with ISO Class 8 cleanroom assembly and full traceability for every spool. If you are planning a liquid-cooled deployment, talk to our engineering team about your SFN requirements.