Every time you ask an AI tool a question, thousands of chips start working inside a data centre. Those chips get very hot, very fast. Keeping them cool is now one of the biggest challenges in the industry. The answer is liquid cooling, and at its centre is a system called the secondary fluid network, or SFN.
You may never have heard the term. However, without a secondary fluid network, today’s AI data centres simply could not run. This guide explains what an SFN is, how it works and why it matters. We cover the technical details too, but in plain words, so anyone can follow along.
Why AI data centres need liquid cooling
A data centre is a building full of servers. Servers sit in tall metal cabinets called racks. For years, fans and cold air kept these racks cool, much like an air conditioner cools a room.
That worked well when a rack used 5 to 10 kilowatts (kW) of power. For comparison, 10 kW is about what five household geysers use when they all run together. Almost all of that power turns into heat.
AI has changed the picture. A single rack of AI chips, called GPUs, can now use 40 to 150 kW. That is like packing 75 geysers into one cabinet. Air simply cannot carry that much heat away. You would need hurricane-force fans, and they still would not be enough.
However, water and other liquids are far better at moving heat. In fact, water can carry around 3,500 times more heat than the same volume of air. So data centres now bring liquid right up to the chips. Engineers call this direct-to-chip liquid cooling.
In this approach, a small metal block called a cold plate sits on top of each chip. Coolant flows through tiny channels inside the cold plate and picks up the heat. Then the warm coolant flows away to be cooled down again. The secondary fluid network is the piping that makes this journey possible.
What is a secondary fluid network?
A secondary fluid network is the system of pipes that carries coolant between the CDU and the server racks. CDU stands for coolant distribution unit. You can think of the CDU as the pump room and control centre of the cooling system.
An easy way to picture it is the human body. The heart pumps blood through large arteries, then into smaller vessels that reach every part of the body. In a data centre, the CDU is the heart. The secondary fluid network is the network of arteries and vessels. And the cold plates are the organs that need a steady supply.
Most liquid-cooled data centres actually use two separate loops:
- The primary fluid network (PFN) carries water from the CDU out to the building’s cooling plant, such as chillers or dry coolers on the roof. This is the large, building-wide loop.
- The secondary fluid network (SFN) carries a special coolant from the CDU to the racks and back. This is the smaller loop that touches the IT equipment.
The two loops never mix. Instead, a heat exchanger inside the CDU passes heat from one loop to the other. It works like two pipes touching without sharing water. Because of this, the SFN can stay small, clean and tightly controlled. Engineers also call it the technology cooling system, or TCS loop.
The main parts of an SFN
A secondary fluid network is more than just pipes. Each part has a clear job.
Supply and return headers. These are the main pipes that run along each row of racks. The supply header carries cool liquid out. The return header brings warm liquid back. You can think of them as the main road in and the main road out.
Branch lines and rack drops. Smaller pipes branch off the headers to reach each rack. They are like side streets that lead from the main road to each house.
Rack manifolds. Then, inside each rack, a manifold splits the flow between the servers. It works like a power strip, but for coolant.
Quick disconnects. Next come special couplings that let technicians connect or remove a server without spilling coolant. They seal themselves automatically when unplugged.
Valves. Isolation valves let engineers shut off one section for repairs while the rest keeps running. Balancing valves make sure every rack gets its fair share of flow.
Air vents and drains. Air bubbles stop coolant from flowing properly, so vents at the high points let trapped air escape. Similarly, drains at the low points let engineers empty the system for maintenance.
Sensors. Finally, temperature, pressure and flow sensors watch the system all the time. They send alerts if something changes, such as a leak or a blocked pipe.
The technical numbers, made simple
Engineers design a secondary fluid network around a few key numbers. Here is what each one means and why it matters.
Heat load. This is how much heat each rack produces, measured in kW. It is the starting point for everything else. A 100 kW rack needs far more cooling than a 20 kW one.
Temperature rise (ΔT). This is how much warmer the coolant gets as it passes through the rack. Engineers often design for a rise of about 10°C. For example, coolant might enter at 30°C and leave at 40°C.
Flow rate. This is how much coolant must flow through each rack every minute. More heat needs more flow. A useful rule of thumb is about 1.5 litres per minute for every kW. So a 100 kW rack needs roughly 150 litres per minute. That is about the same as eight kitchen taps running at full speed.
Pipe size, pressure and coolant
Pipe size. Once engineers know the flow, they choose the pipe size. If the pipe is too small, the liquid must rush through too fast. That wastes energy, makes noise and wears out fittings. If it is too big, the pipe costs more than it needs to. For our 100 kW rack, a pipe of about 50 mm across (DN50) is a typical choice. The main header feeding ten such racks might be 125 to 150 mm across.
Pressure drop. Every pipe, bend and valve slows the liquid down a little, a bit like friction on a road. The total of all these losses is the pressure drop. The CDU’s pumps must be strong enough to overcome it, with some margin left over.
The coolant itself. Many systems use PG25, a mix of 75% water and 25% propylene glycol. The glycol protects against freezing and germs. Special additives called inhibitors stop the pipes from corroding. However, PG25 is thicker than plain water, so engineers must allow for that when they size pumps and pipes.
Why cleanliness, materials and welding matter
The channels inside a cold plate are incredibly fine, often thinner than a grain of rice. Even a few specks of dirt or metal can block them. When that happens, the chip overheats and slows itself down to stay safe. In the worst case, it shuts down.
That is why a secondary fluid network needs a very high standard of cleanliness. There are three main parts to getting this right.
The right material. Most high-quality SFNs use 316L stainless steel. It resists rust and works well with glycol coolants. The “L” means low carbon, which keeps the metal strong and rust-free around the welds. On the other hand, good builders avoid galvanised steel, because its zinc coating can flake off into the coolant.
Clean welding. Welding joins the pipes together, and the inside of each joint must be smooth. The best method is orbital welding. Here, a machine rotates around the pipe and makes an even weld every time. During welding, an inert gas such as argon fills the inside of the pipe. This stops the hot metal from burning and forming a rough, crusty surface that traps dirt.
Cleaning and a cleanroom. After welding, the pipes go through a chemical clean called passivation. It removes any traces of iron and restores stainless steel’s natural protective layer. Then technicians assemble the pipes in a cleanroom. An ISO Class 8 cleanroom is a controlled space where filters keep dust levels in the air very low. Finally, they cap and seal every open pipe end until installation.
Testing and commissioning
Before any server is connected, the secondary fluid network must prove that it is safe and clean. This happens in several steps.
First, every pipe section is pressure tested in the factory. Engineers fill it and push the pressure well above normal levels, often 1.5 times higher. If nothing leaks, the section passes.
Next, engineers test the full system again after installation on site. Then they flush it with clean water at high speed to wash out any dust from the building work. During flushing, temporary bypass pipes keep the water away from the cold plates.
After that, they fill the system with the final coolant and let out all trapped air. Engineers then measure the coolant’s quality and record it as a baseline. Finally, they check that each rack receives the right flow. All test results go into a handover pack. As a result, the owner has a full record of the build.
What can go wrong with a poor SFN
A badly designed or badly built SFN causes real problems:
- Hot spots. If flow is uneven, racks far from the CDU get too little coolant and their chips slow down.
- Blockages. Dirt from poor welding or cleaning can clog cold plates, which are expensive to replace.
- Leaks. Weak joints can leak coolant near live electrical equipment.
- Wasted energy. Undersized pipes force pumps to work harder every hour of every day.
- Costly upgrades. A network with no room to grow may need to be drained and rebuilt when new racks arrive.
Each of these problems starts with decisions made during design and fabrication. That is why choosing the right SFN partner matters so much.
How to choose an SFN supplier
When you compare suppliers, ask these questions:
- Do they design the SFN from your actual heat loads and cold plate data?
- Do they use 316L stainless steel and orbital welding?
- Do they passivate the pipes and assemble them in a cleanroom?
- Do they pressure test every section and give you the records?
- Can they support flushing, commissioning and future expansion?
A supplier that handles design, fabrication and testing under one roof usually delivers better quality. There are fewer handovers between companies, so there are fewer chances for mistakes or dirt to creep in.