Modular Technology Cooling System for "Cloud Scale": Design & Delivery of Liquid to Rack Distribution Systems, Version 1.0 (25 August 2025)

Overview

The paper starts from the point where AI factories, with uncertain rack power densities, shifting delta-T ranges and fluid-quality demands, outgrow traditional job-site construction. Its answer is DfMA (design for manufacture and assembly), with off-site, BIM-enabled modular pipe systems and BIM content at LoD 350+ as the starting point. It is a crowd-sourced, non-normative guidance paper; product images are examples and equivalent products are equally recommended. It builds on earlier OCP papers on connection methods and on propylene-glycol and water-based fluids.

The modularity chapter lists benefits: adaptability to delta-T and kW-per-rack changes, cleanliness that makes sub-25 micron contamination levels attainable, lower embodied carbon, supply-chain resilience, fewer defects and fewer skilled labour hours on site. It recommends fewer joints, standard connection methods (BSPP for threads), alignment tolerance, and flexibility for vibration and seismic loads. A cloud-scale data centre can have 20,000 to 40,000 or more pipe connections, which is why FMEA and connection-reliability analysis are central. Temperatures from job-site conditions of 15-20 C to peaks above 70 C are stressed.

The fluid and pipe-material chapter states that microchannel exchangers need filtration of 25 micron or finer, glycol additives reduce specific heat by under 10% with proportional capacity loss, and corrosion-prone materials such as carbon steel are generally avoided. The pipe design table gives rear-door loops Tmax below 37 C and 6.5 bar, and liquid-to-chip loops Tmax 70 C, design range 50 C and at most 4 bar (stainless 304/316 Sch 5/10). High-performance systems may exceed 3 m/s but lower velocities are preferred because head loss rises sharply. Thermal movement across 20-70 C is a key material-selection driver.

The capacity-alignment chapter ties TCS capacity to electrical capacity. Cold-plate systems capture 70-80% of IT heat, so TCS capacity may be about 80% of electrical load, against close to 100% for immersion. At a typical branch velocity of 2.3 m/s and 10 C delta-T, a DN25 branch carries about 56 kW and a DN50 branch about 214 kW; DN50 carries about four times the flow at the same velocity but head loss rises with the square of velocity. The same CDU piping delivers 2,422 kW at 10 C delta-T and 4,845 kW at 20 C. DN25 uses the OCP large quick connector standard; DN50 uses double ball valves or proprietary drip-less connectors.

The appendices cover reliability and FMEA, calculation tools, OVHcloud experience (25 micron filtration in the TCS and makeup water filtered to 1 micron where possible), flow control, module size and valving (blast radius set by isolation valve location), structural support, commissioning for cleanliness (no pipe tape or dope on site; passivation, flushing, clean delivery), thermal ride-through and SLA topics. Ride-through time is mass times specific heat times temperature margin divided by heat load; the example uses 100 litres of water, a 10 kW load, 25 C initial and 40 C critical temperature. Tool outputs are indicative and must be verified by an engineer.

Key points

1. Target scale is roughly 10 MW to over 300 MW; the approach is DfMA with off-site fabricated modular pipe systems and BIM at LoD 350+.
2. A cloud-scale facility can have 20,000-40,000 or more pipe connections; reliable operation of 20,000 connections over 25 years needs an MTBF above 4.4 billion hours.
3. Microchannel cold plate heat exchangers need filtration of 25 micron or finer; glycol additives cut specific heat by under 10%.
4. Design table: rear-door loops Tmax below 37 C and 6.5 bar; liquid-to-chip loops Tmax 70 C, design range 50 C, max 4 bar, stainless 304/316 Sch 5/10.
5. High-performance TCS may exceed 3 m/s, but lower velocities are preferred because head loss and pump energy rise sharply.
6. Cold plate systems capture 70-80% of IT heat, so TCS capacity may be about 80% of electrical load, versus close to 100% for immersion.
7. At 2.3 m/s and 10 C delta-T a DN25 branch carries about 56 kW and a DN50 branch about 214 kW; the same CDU piping delivers 2,422 kW at 10 C and 4,845 kW at 20 C delta-T.
8. Most IT equipment is designed for a 10 C coolant delta-T, with some designs considering 20 C.
9. OCP cold plate fluid guidance keeps chloride below 25 ppm, enabling 304 stainless steel; OVHcloud uses 25 micron TCS filtration and 1 micron makeup water where possible.
10. Pipe tape and dope are prohibited on site; threaded connections should be limited to factory or shop work, with mechanical connections for repeated rack connections.

Using it in design

• When sizing TCS capacity against electrical capacity, state the cold-plate capture fraction (70-80%) and target delta-T (10 or 20 C) explicitly, and use the roughly doubled delivery at 20 C (4,845 versus 2,422 kW) as design headroom.
• Compare DN25 and DN50 branch lines on capacity (56 versus 214 kW), connection method, supply chain and pressure drop; for racks above 100 kW, write the DN50 drip-less connector or double-valve procedure into the specification.
• Select pipe material by 20-70 C thermal movement, pressure rating (the 6.5 and 4 bar table), loss of internal diameter in non-metallic pipe above 40 C, corrosion and embrittlement data; passivate after welding and minimise dead legs.
• Specify filtration at 25 micron or finer, makeup-water and chemistry limits (chloride below 25 ppm), require modules to be flushed and delivered clean from the factory, and ban pipe tape and dope on site.
• Plan module size and isolation valves around the acceptable blast radius and concurrent maintainability; dual valves between modules allow valve maintenance.
• Calculate thermal ride-through from mass, specific heat, temperature margin and heat load and confirm it with CDU and IT thermal inertia included.

Who should read it

Mechanical engineers designing TCS piping for large liquid-cooled data centres, contractors, prefabricators and commissioning teams.

Document contents

1. Version History
2. Executive Summary
3. Introduction
4. Compliance with Open Compute Project Tenets
5. Modular TCS @ Cloud-Scale - Design & Delivery
6. Key Benefits for Modular TCS Design & Delivery
7. Requirements for Modularity in TCS Pipe Systems
8. Reliability Considerations
9. TCS Water-Based Fluid Requirements
10. TCS Flow Considerations
11. TCS Pipe Selection Considerations
12. Loop & Branch Line Alignment to Electrical Capacity
13. Conclusion
14. Glossary
15. Appendix A. Reliability & Risk Considerations in TCS Pipe Systems
16. Appendix B. Calculations & Selector Tools for TCS Planning
17. Appendix C: 25um Filtration in TCS - Insights from OVHcloud
18. Appendix D: Flow Control Considerations
19. Appendix E: Considerations for TCS Module Size & Valving
20. Appendix F: Structural Considerations for TCS Modules
21. Appendix G: Commissioning for Cleanliness Considerations
22. Appendix H: Thermal Ride-Through Considerations of the TCS
23. Appendix I: Service Level Agreement Considerations

Read the document

Publication details

Authors

Don Mitchell; Hussameddine Kabbani; Vali Sorell; Stuart Lawrence; Ali Chehade; Vivek Prasad; David Quirk; Matt Hatley; Dariusz Raczkowski; Chris Campbell; Pardeep Shahi

Publisher

Open Compute Project Foundation

Year

2025

Document language

English

Source type

Industry body

Access

Open licence — downloadable

Licence

CC BY-SA 4.0 — share with attribution under the same licence

Licence evidence

www.opencompute.org

Citation