Reference Design #505: Eaton 1 GW Power Architecture Meets Trane Thermal Management

Overview

The paper opens with the rack-power trajectory: GB200 NVL72 about 120 kW, GB300 about 140 kW, Vera Rubin NVL72 190-230 kW, Rubin Ultra NVL576 about 600 kW and a Kyber-era SuperPOD about 5 MW. A 100 MW hall of GB200 NVL72 racks means about 830 racks at roughly 120 kW each. Cooling grows with it: CDU capacity rose about 100x in five years, from ~100 kW in 2018 to 1-2 MW in 2025, with 5-10 MW modules next.

Mechanical load is now its own electrical problem. A 100 MW compute block can impose about 20 MW of motor load, with pumps up to 1,500 HP. Each cooling block holds 10 chillers of about 3,000 tons (10-11 MW thermal) and about 20 large pumps. Feeding these motors at 4.16 kV is about current: a 1,500 HP pump draws 1,420 A at 480 V but 164 A at 4.16 kV. CDU pods are treated as UPS-backed critical loads.

On the rack side the paper argues for 800 V DC. Its table shows near-equal currents at 480 V AC and 800 V DC: 152 A versus 150 A at 120 kW, 1,267 A versus 1,250 A at 1 MW, 12,668 A versus 12,500 A at 10 MW. The real gain is conductor count: two-wire DC is claimed to cut copper weight by up to 34% and raise efficiency by 3-5%. This is a joint marketing-style reference design by two manufacturers, so the 3-5% efficiency and 34% copper figures should be read as claims without independent verification. Still, architecture decisions such as MV motor feeds and putting CDUs on UPS give a sound framework for calculations on very large AI campuses.

Key points

1. Rack power: GB200 NVL72 ~120 kW, GB300 ~140 kW, Vera Rubin NVL72 190-230 kW, Rubin Ultra NVL576 ~600 kW; Kyber-era SuperPOD ~5 MW
2. Current at 480 V AC vs 800 V DC: 120 kW 152 A vs 150 A; 1 MW 1,267 A vs 1,250 A; 10 MW 12,668 A vs 12,500 A
3. 800 V DC claimed to raise efficiency 3-5% and cut copper weight up to 34% (2-wire vs 3-phase)
4. CDU capacity grew ~100x in five years: ~100 kW (2018) to 1-2 MW (2025) to 5-10 MW modules
5. A 100 MW compute block can impose ~20 MW of motor load, pumps up to 1,500 HP
6. 1,500 HP pump: 1,420 A at 480 V vs 164 A at 4.16 kV
7. Each 100 MW cooling block: 10 chillers of ~3,000 tons (10-11 MW thermal each) plus ~20 large pumps
8. 100 MW hall of GB200 NVL72 racks needs ~830 racks at ~120 kW
9. CDU pods placed on UPS as critical loads

Using it in design

• Voltage-selection rule: motors above about 200 kW (large chillers and pumps) default to MV (4.16/6.6/11 kV) feeders; compute I = P/(sqrt(3) x V x PF) to show cable savings.
• For liquid-cooled AI blocks default the mechanical load ratio to motor load of about 20% of IT.
• Always put CDUs on UPS.
• Offer 800 V DC rack power as an extra option flagged as emerging.

Who should read it

Designers and investors defining power and cooling architecture for AI campuses of hundreds of MW.

Related standards

• IEEE 1584
• NEC Article 708/645
• OCP Open Rack / 800 VDC
• ASHRAE TC9.9 Liquid Cooling Guidelines

Publication details

Authors

Eaton; Trane

Publisher

Eaton / Trane

Year

2025

Document number

Reference Design #505

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