Transforming Colocation Facilities to Meet Demand for AI Compute

Overview

The guide starts with density: average multi-node AI pod density is projected to rise from 40-100 kW to 100-250 kW. On power it covers BESS and grid-support UPS modes to ease grid constraints, UPS behaviour under GPU step loads, and higher-voltage and 800 VDC distribution delivered by busway. Data-centre BESS is typically Li-ion sized for 1 to 6 hours at full load. GPU step changes commonly range from below 30% to 100% of load, with peaks up to 150% of UPS rated capacity within milliseconds.

The cooling section splits heat between direct-to-chip liquid and residual air. Air-cooled facilities typically run PUE 1.5-1.7, meaning up to 30% of power goes to non-compute functions. Direct-to-chip cooling typically removes over 70% of rack heat; the remaining ~30% (memory, NICs, PSUs) needs air. With rear-door heat exchangers, room air must be able to carry 100% of the rack load while the exchanger is serviced. For CDU placement, in-rack, in-row or gallery options are covered, with row- or suite-level N+1 and liquid-to-air CDUs for operators with few liquid customers.

Commissioning and operation steps are concrete: fluid quality verification, air removal from all pipe runs, sensor and alarm validation and staged ramp-up with simulated heat loads, with leak sensors at every critical connection and loop separation practices. Factory-integrated prefabricated modules can cut deployment time by up to 50% against traditional builds. The text is a vendor guide, and its numbers are projections and typical values rather than measured results.

Key points

1. Average multi-node AI pod density is projected to rise from 40-100 kW to 100-250 kW.
2. Data-centre BESS is typically Li-ion sized for 1 to 6 hours at full load, paired with grid-support UPS modes.
3. GPU step changes commonly range from below 30% to 100% of load, with peaks up to 150% of UPS rated capacity within milliseconds.
4. Air-cooled facilities typically run PUE 1.5-1.7, meaning up to 30% of power goes to non-compute functions.
5. Direct-to-chip cooling typically removes over 70% of rack heat; the remaining ~30% (memory, NICs, PSUs) needs air cooling.
6. With rear-door heat exchangers, room air must carry 100% of the rack load while the exchanger is serviced.
7. Liquid commissioning: fluid quality verification, air removal from all runs, sensor/alarm validation and staged ramp-up with simulated heat loads; leak sensors at every critical connection.
8. Factory-integrated prefabricated modules can cut deployment time by up to 50% versus traditional builds.

Using it in design

• Give the rack/pod model a liquid fraction (default about 70% for direct-to-chip) so residual air load is computed automatically.
• For rear-door exchanger racks check room air capacity covers 100% of the rack during door service, and require N+1 CDUs at row or suite level.
• Require leak detection points at CDU and manifold connections, and flag UPS overload capability below the expected GPU peak (up to 150%).
• Offer CDU placement options (in-rack, in-row, gallery, liquid-to-air) and a liquid-cooling commissioning checklist.

Who should read it

Operators and designers converting existing colocation halls for AI pods.

Related standards

• ASHRAE TC 9.9 Liquid Cooling Guidelines
• OCP ORV3
• IEC 62040-3

Publication details

Authors

Vertiv

Publisher

Vertiv

Year

2026

Document number

SL-80478

Document language

English

Source type

Manufacturer

Access

Free to read

Licence

Copyright retained; citation only — the publisher’s terms apply to the original document

Citation