Technical guideFree to read2026

Designing for AI Workloads (ASME Digital Infrastructure column)

ASME (Digital Infrastructure column) · Robert Sty

A column by Robert Sty, global data center practice director at HDR, outlining mechanical design issues for liquid-cooled AI halls.

On this page

Overview

This column appears in ASME's Digital Infrastructure series and is written by Robert Sty, global data center practice director at HDR. Its subject is the topics met in the mechanical design of liquid-cooled AI halls. It starts from the jump in density, noting that air-cooled racks were typically 10-20 kW while AI cabinets can exceed 100 kW.

The author describes coolant distribution units (CDUs) as the boundary between the facility water system (FWS) and the technology cooling system (TCS). CDUs regulate flow, temperature and pressure to each cabinet and are typically installed N+1, with variable-speed drives on the pumps. On temperature classes, the ASHRAE W classes cover facility water temperatures and the S classes cover TCS supply to IT. This distinction clarifies which loop should be described with which temperature range.

AI training and inference workloads create large and fast power swings, so thermal transients matter. To handle them, TCS piping volume and sometimes small buffer or thermal storage tanks are used, and solutions are validated with CFD. In piping, the advice is to keep cooling water above the room dew point, to use looped piping for redundant paths, to select joints and materials carefully, and to install leak sensors.

The column cites ASME B31.9 for building services piping and ASHRAE 90.1, 90.4 and 189.1 as applicable references. As a limit, it is a column; it gives no detailed calculation method or numerical sizing criteria. Even so, it presents in compact checklist form which topics mechanical design should watch in high-density liquid-cooled halls, and it should be read alongside the ASHRAE TC 9.9 liquid cooling guidelines.

Key points

  1. Air-cooled racks were typically 10-20 kW; AI cabinets can exceed 100 kW.
  2. CDUs separate FWS from TCS and regulate flow, temperature and pressure to each cabinet; typically installed N+1.
  3. AI training and inference cause large, fast power swings; CDU pumps use variable-speed drives.
  4. TCS volume or small thermal storage tanks buffer temperature rise, validated with CFD.
  5. ASHRAE W classes cover facility water temperatures and S classes cover TCS supply to IT.
  6. Keep cooling water above room dew point; use looped piping for redundant paths, careful joint and material selection and leak sensors.
  7. Applicable references: ASME B31.9 for building services piping, ASHRAE 90.1/90.4/189.1.

Using it in design

  • Default CDUs to N+1 per row or pod.
  • Warn when TCS supply temperature is set below the room dew point.
  • Provide a TCS loop-volume and buffer-tank check against a user-defined load step (kW over seconds).
  • Add looped layout, leak detection and material selection to the piping design checklist.

Monthly newsletter · free

Keep your reading list current

Once a month: new library sources, changes to Turkish regulations and upcoming data centre events.

SoSo CAD

Check cooling capacity alongside your design

Review zone heat loads, available capacity and the capacity balance after the largest unit is lost in SoSo CAD.

Request a demo
Zone heat loads and capacity balance in the SoSo CAD cooling workspace

Who should read it

Engineers designing the mechanical systems of high-density liquid-cooled AI halls.

Related standards

  • ASME B31.9
  • ASHRAE 90.4
  • ASHRAE 90.1
  • ASHRAE TC 9.9 Liquid Cooling Guidelines

Publication details

Authors
Robert Sty
Publisher
ASME (Digital Infrastructure column)
Year
2026
Document language
English
Source type
Industry body
Access
Free to read
Licence
Copyright retained; citation only — the publisher’s terms apply to the original document

Citation

Robert Sty (2026). Designing for AI Workloads (ASME Digital Infrastructure column). ASME (Digital Infrastructure column). https://www.asme.org/topics-resources/content/blog-designing-for-ai-workloads

All content and rights belong to the respective authors and publishers. SoSo CAD provides only the citation, its own summary and a link to the source, and is not responsible for the sources' content.