← Back to Blog
Technical 19th November 2026

How Many Valves Go Into a 1 GW Liquid-Cooled Data Center?

Illustration: 1 GW of IT load leads to 7,576 racks and 9,000 to 26,000 valves, with rows of racks joined to blue supply and orange return pipes that carry valves

Key Takeaway

About 9,000 to 26,000 valves for every 1 GW of IT load, counting only the liquid loop that serves the racks (132 kW per rack).

We found no public source for this number, so we worked it out from the rack count and one public example. A handy rule: about 3.4 valves per rack.

The biggest group is the isolation valve on each rack’s supply and return pipe. Vertiv says manufacturers typically use “an isolation valve (such as a ball valve)” there, so that is the first place to look for ball valves in the count.

How many valves is that?

Between 9,000 and 26,000. The low end leaves out the valves on each rack’s pipes.

WhereValves per 1 GWHow we got it
Rack drops (the pipes into each rack)about 15,0002 per rack × 7,576 racks. Vertiv names a valve at the rack inlet; the second, on the return, is our assumption
CDU and row manifoldsabout 9,000 to 11,000Scaled up from one public example
Facility-water sidenot countedNo public figure found
Total9,000 to 26,000Low end has no rack-drop valves

Per rack, that is 2 valves on the rack’s own pipes plus about 1.4 at CDU and row level, so about 3.4 valves per rack.

What does “1 GW” mean here?

1 GW of IT load, which is the power the racks themselves use. Cooling and electrical losses come on top, so 1 GW at the grid connection holds fewer racks.

H2X Engineering gives 132 kW for a GB200 NVL72 rack. At that density 1 GW holds about 7,576 racks (1,000,000 kW ÷ 132 kW).

Where do the valves sit?

At the CDU, along the row manifold, and on the pipes into every rack. The Open Compute Project calls this the technology cooling system (TCS) loop: the row manifold, the rack manifolds and the secondary side of the CDUs.

Where the valves sit in the liquid loop that serves the racks A coolant distribution unit feeds a supply and a return row manifold. Each rack hangs off the row manifold on two drops, one supply and one return, each with an isolation valve. Valves also sit at the CDU connections and along the row manifold. The facility-water side of the CDU is not counted. At 1 GW of IT load and 132 kW per rack, the rack drops hold about 15,152 valves and the CDU and row level about 9,100 to 10,800. Facility-waterside, not counted CDU secondary side Row manifold, supply Row manifold, return 1 Rack 1 rack manifold Rack 2 rack manifold Rack N rack manifold ··· 2 3 1 CDU and row-manifold valves: about 9,100 to 10,800 per GW, scaled from one published example. 2 Rack drops: two per rack, 15,152 per GW at 132 kW per rack. An assumption, not a published count. 3 Facility-water side (chillers, towers, primary headers): not counted, no public figure found. Valve symbol = isolation valve. Original diagram.
The TCS loop and the three groups of valves this page counts. Original diagram, simplified.

None of these sources gives a quantity. They show where valves are, not how many.

Where do the 9,000 to 11,000 come from?

From one example bill of materials on the TCS reference page at intelaccelerator.com. It lists 5 CDUs of 2.5 MW (12.5 MW installed), 80 racks and 114 valves at CDU and row level: 26 control or balancing valves and 88 isolation, check or drain valves.

That is 9.1 valves per installed MW, or 1.4 per rack. Scaled to 1 GW, it gives about 9,100 to 10,800. The page says “All figures shown are demo data for structural demonstration” and does not say whether rack-level valves are included, so we treat them as excluded.

Why does rack density matter?

Denser racks mean fewer valves per gigawatt. Rack-drop valves follow the number of racks, and 1 GW of 40 kW racks is 3.3 times as many racks as 1 GW of 132 kW racks.

Rack densityRacks per 1 GWRack-drop valves (2 per rack)Total with CDU and row (about 9,100)
40 kW (H100-class, per H2X)25,00050,000about 59,000
132 kW (GB200 NVL72, per H2X)7,57615,152about 24,300

The 40 kW row assumes every rack is liquid-cooled, which a real hall may not be.

Why is a single valves-per-megawatt number hard to trust?

Because it swings with three things.

When you see a valves-per-megawatt figure, ask which rack density and which scope are behind it.

How many of them are ball valves?

No public source gives a ball valve count, so we do not give one. What we can show is where ball valves are most likely to sit.

Vertiv’s white paper says: “Typically, manufacturers use an isolation valve (such as a ball valve) placed between the secondary fluid network piping and the in-rack manifold’s inlet.” That is the rack-drop group, about 15,000 per gigawatt under our two-per-rack assumption. Read that figure as the size of the group where ball valves usually go, not as a ball valve count.

Other positions are less clear. Vertiv gives typical header piping as 150 to 300 mm (3 to 6 in) and branch piping as 25 to 50 mm (1 to 2 in), so valve sizes in the loop span a wide range. Which valve type suits each size is a design decision, and no source we read splits them.

Material is the other half of the question. OCP’s row-manifold guideline expects pipework in “stainless steel, minimum grade 304”, and its propylene glycol guideline prefers 304L, 316L or higher grades for wetted metals. A stainless steel ball valve in those grades fits both. For the reasoning, see brass or stainless in a glycol cooling loop and why data centers choose 3-piece ball valves.

What does this count leave out?

How do you count your own project?

  1. Divide the IT load by kW per rack to get the rack count.
  2. Count the valves on each rack’s supply and return pipes, plus any inside the rack manifold.
  3. Count the CDU connections from the datasheet, and note which valves come inside the CDU.
  4. Count the isolation points, drains and vents on each row manifold.
  5. Count the facility-water side separately. Then group every valve by size and duty (isolating or controlling) to get your own ball, butterfly and other split.

Frequently Asked Questions

How many ball valves are in a 1 GW liquid-cooled data center?

No public source says. The rack drops are the largest group, about 15,000 per GW at two per rack, and Vertiv says an isolation valve “such as a ball valve” typically sits there. That is the size of the group, not a ball valve count.

How many valves per megawatt does a liquid-cooled data center need?

About 9 per MW at CDU and row level, plus 2 per rack on the rack pipes. At 132 kW per rack that is roughly 9 to 26 per MW in total. These are derived figures, not published ones.

Does 1 GW mean grid power or IT load?

This page uses IT load. Grid power also covers cooling and electrical losses, so the same connection supports fewer racks and fewer valves.

Does the CDU include its own valves?

Airedale’s does: it lists isolation valves for supply and return. Other vendors differ, so confirm what is inside each CDU before counting field valves.

What are liquid cooling ball valves made of?

OCP’s row-manifold guideline expects stainless steel pipework, minimum grade 304. OCP’s propylene glycol guideline prefers 304L, 316L or higher grades for wetted metals. A stainless steel ball valve in those grades fits both.

Is 132 kW per rack a fixed figure?

No. H2X Engineering gives 132 kW for a GB200 NVL72 rack and about 40 kW for an H100-class rack. Rack power rises with each generation, and the valve count per gigawatt falls with it.

Published by LINS Valve Industrial Co., Ltd., Taichung, Taiwan. Last Updated: 2026-11-19. The cover is a generated illustration, not a photograph of a specific product or facility.

Referenced sources: Open Compute Project, Guidelines for Pre-Commission Preparation of Technology Cooling System (TCS) Row Manifolds in Liquid Cooled Data Centers, version 1.0; Airedale, 400kW - 2MW+ Coolant Distribution Unit (CDU); H2X Engineering, Data Center Liquid Cooling Design for Direct-to-Chip Racks; Vertiv, Deploying liquid cooling in the data center: a guide to high-density cooling (white paper, 2024); TCS / Liquid Cooling Loop reference page, intelaccelerator.com.

OCP and Vertiv quotations were read from the PDF text layer, so check them against the documents. The other pages were read on 19 November 2026. All per-gigawatt figures here are arithmetic on these sources, not published counts.