Request Consultation
Liquid Cooling for AI Data Centres · Part 3

Who Owns the Liquid Cooling Loop?

A direct-to-chip cooling loop runs from the cold plates on each GPU, through the rack and the coolant distribution unit (CDU), out to the building’s chillers or dry coolers. On most projects, four or five companies supply and install its parts. So who owns the liquid cooling loop as a whole?

By Team@MCE November 10, 2026 6 min read
Who Owns the Liquid Cooling Loop? One coolant loop through five parts, each with a different owner, with the four connections between them marked.
On this page
  1. 01 How the Loop Is Usually Split
  2. 02 How the Loop Is Built, and Who Does What
  3. 03 What Is at Stake
  4. 04 Where Responsibility Tends to Fall Through
  5. 05 Three Questions for Your Team
  6. 06 Settling Who Owns the Loop, Early

Each part usually has a clear owner. The loop as a whole often does not. The easiest way to see why is to follow the loop as it is built: who does what, in what order, and where one company hands over to the next.

In the last post we showed why a direct-to-chip hall has seconds, not minutes, when its cooling stops. This post looks at a question that comes before that one: when the loop has a problem, who is responsible for it? In an air-cooled hall the split was simple. The server ended at its fans, and the building looked after the air. A liquid cooling loop runs from inside the server to the roof.

In an air-cooled hall the split was simple. A liquid cooling loop runs from inside the server to the roof.

01

How the Loop Is Usually Split

At the chip

The cold plates inside the server belong to the server maker. They ship with the server and sit under its warranty.

In the rack

The rack manifolds and quick-disconnect couplings belong to the server maker, the rack integrator or the facility, depending on the contract. Often nobody is sure which.

The pipework to the racks

The pipework between the racks and the CDU, called the technology cooling system (TCS), is designed by the mechanical engineer and installed by the mechanical contractor. In a colocation hall it may be the tenant’s.

The CDU

The coolant distribution unit is supplied by its maker, installed by the contractor and connected by the controls integrator.

The facility side

The building’s cooling water, called the facility water system (FWS), belongs to the building: designed by the engineer of record and run by the facilities team for the life of the hall.

One loop, five parts Each part has an owner. Who owns the connections between them, and what runs through all of them? Server cold plates on each GPU Server maker Rack manifolds, couplings Depends on the contract TCS piping between racks and CDUs Mechanical contractor CDU pumps, heat exchanger, controls CDU maker, contractor, integrator Facility (FWS) chillers or dry coolers Building owner or landlord ? ? ? ? ? a connection between two scopes: often on nobody’s drawing or test cool supplywarm return Runs through every part, and needs a named owner Coolantspecified, filled and run by three parties ControlsCDU, servers and BMS, set up apart Leak responsealarm, isolate, keep running? End-to-end testthe whole loop, as one system Illustration for the concept, not a detailed design. © Mission Critical Engineers™ One loop, five parts Each part has an owner. Who owns the connections between them, and what runs through all of them? Servercold plates on each GPUServer maker Rackmanifolds, couplingsDepends on the contract TCSpiping between racks and CDUsMechanical contractor CDUpumps, heat exchanger, controlsCDU maker, contractor, integrator Facility (FWS)chillers or dry coolersBuilding owner or landlord ? ? ? ? ? a connection between two scopes:often on nobody’s drawing or test cool supplywarm return Runs through every part, and needs a named owner Coolant: specified, filled and run by three parties Controls: CDU, servers and BMS, set up apart Leak response: alarm, isolate, keep running? End-to-end test: the whole loop, as one system Illustration for the concept, not a detailed design. © Mission Critical Engineers™
What the picture shows. One coolant loop runs from the cold plates on each GPU to the coolers on the roof, through five parts, each supplied by a different party. The connections between the parts, and the things that run through all of them (the coolant, the controls, the leak response and the end-to-end test), need a named owner too.
02

How the Loop Is Built, and Who Does What

The order varies from project to project, but on a typical project the loop goes in roughly like this.

  1. 01DesignThe engineer designs the building’s cooling water and the pipework to the racks, to suit what the server maker says its servers need. The servers themselves are often still being chosen.
  2. 02The cooling plantThe mechanical contractor installs the chillers or dry coolers and the building’s cooling water pipework.
  3. 03The CDUsThe CDU maker supplies the units, and the mechanical contractor sets them in place and connects them to the building’s water.
  4. 04The pipework to the racksThe mechanical contractor runs the pipework out to each rack position, checks it for leaks and cleans it out.
  5. 05The coolantThe loop is filled with the coolant the server maker asks for. Who fills it, and who supplies the coolant, varies.
  6. 06The controlsThe controls integrator connects the CDUs and the leak detection to the building’s control system.
  7. 07The racks and serversOften months later, the server maker or rack integrator brings in the racks and connects them to the pipework.
  8. 08Running as oneThe whole loop runs with the servers’ heat on it.
  9. 09HandoverThe facilities team takes over the loop and runs it for the life of the hall.
How the loop is built, and who does what Each stage is done by a different company, and each one hands over to the next. designhandover the loop waits,often for months 1Design Engineer 2Building cooling plant Mechanical contractor 3CDUs set and connected CDU maker + contractor 4Pipework to the racks Mechanical contractor 5Coolant fill Contractor (varies) 6Controls connected Controls integrator 7Racks and servers Server maker or integrator 8Whole loop runs as one All the parties 9Handover to operations Facilities team 1 2 3 Where one company hands over to the next 1 Who looks after the loop while it waits? 2 Who checks the connection when the racks arrive? 3 The first time the whole loop runs together Illustration for the concept. Order and timing vary by project. © Mission Critical Engineers™ How the loop is built, and who does what Each stage is done by a different company, and each one hands over to the next. designhandover the loop waits, often for months 1 DesignEngineer 2 Building cooling plantMechanical contractor 3 CDUs set and connectedCDU maker + contractor 4 Pipework to the racksMechanical contractor 5 Coolant fillContractor (varies) 6 Controls connectedControls integrator 7 Racks and serversServer maker or integrator 8 Whole loop runs as oneAll the parties 9 Handover to operationsFacilities team 1 2 3 Where one company hands over to the next 1 Who looks after the loop while it waits? 2 Who checks the connection when the racks arrive? 3 The first time the whole loop runs together Illustration for the concept. Order and timing vary by project. © Mission Critical Engineers™
What the picture shows. The loop is built in stages, each by a different company. The pipework is usually finished and filled well before the racks arrive, and the racks are connected by the last company on site. The three numbered points are where one company hands over to the next, and where responsibility is easiest to lose.
03

What Is at Stake

When something goes wrong with the loop, the first question is often whose part it was. A leak in a rack, a blocked cold plate or a CDU that cannot keep the coolant cool enough can take time to sort out between the companies involved, and the servers wait meanwhile.

Warranty is one place it matters. Server makers usually expect the coolant to be kept to their own standard. If nobody was in charge of the coolant from the start, that can be hard to show later. And at handover, the operations team takes on the loop as a whole, so someone needs to have checked it as a whole.

04

Where Responsibility Tends to Fall Through

1

The wait

The pipework is often finished and filled long before the racks arrive. Someone has to look after the loop while it waits, and it is often not clear who.

2

The last connection

The racks are connected by the last company to arrive, to pipework installed by a company that may already have left the site. That connection can end up in nobody’s test.

3

The coolant

The server maker says what the coolant must be, the contractor fills the loop, and the facility team runs it for years. Unless one party is in charge of the coolant from the start, it is easy for it to slip between them.

4

The controls and the leak alarms

The CDU, the servers and the building’s control system are set up by different people at different stages. Who answers a leak alarm, and who decides whether the servers keep running, is often not written down.

5

The first run as one

Each piece passes its own test along the way. The whole loop may first run as one system on the day the servers go live.

None of this tends to show in the contracts, because each one covers its own scope and its own stage. It usually shows when something goes wrong.

What must be proven

Every part of the loop, and every handover between companies, has a named owner. And the loop is tested as one system before the first server relies on it. A colocation hall needs extra care here: the lease sets a line between landlord and tenant, and that line falls somewhere along the loop.

Three Questions for Your Team

  1. Who looks after the loop between the coolant fill and the day the racks arrive?
  2. Who checks the connection when the racks are plugged in?
  3. Who has run the loop from the chip to the roof as one system?

If the answers differ, they are worth settling before construction.

Share the Post:

Related Posts

October 27, 2026

When Liquid Cooling Fails: The Seconds That Matter

When a coolant pump or CDU stops, the GPUs in a direct-to-chip liquid-cooled AI data centre have seconds. How the…

Read More »
October 8, 2026

Liquid Cooling, Explained: Five Technologies and Where Each Fits

Five liquid cooling technologies for AI data centres: rear doors, single- and two-phase cold plates, single- and two-phase immersion. What…

Read More »
July 20, 2026

Load Bank Testing in Data Centre Commissioning: Six Problems We See on Almost Every Project

Six load bank testing problems that derail data centre commissioning — generator tests that prove nothing, cooling validation that misses…

Read More »

Settling Who Owns the Loop, Early

At MCE, our commissioning starts with a design review. It looks at who does what at each stage and where one company hands over to the next, so responsibilities are settled while the contracts can still be changed.

As Commissioning Authority (CxA), we follow the loop through every stage and test it end to end before it is handed over.

Planning or building a liquid-cooled hall? We would be glad to help.

Request Consultation