Data Center Evolution: Powering the AI Revolution Sustainably (OCP EMEA Summit 2025 keynote, Madhusudan Iyengar)

Overview

This document is the deck for the keynote given by Madhusudan Iyengar at the OCP EMEA Summit in Dublin on 29 April 2025. Google explains how synchronous AI training loads change rack power and cooling requirements and what the company is bringing to OCP in response. The deck's starting point is that AI training loads run many accelerators at once and so multiply rack power and heat.

The power section argues that 48 V rack power has been widely adopted for about a decade, but ML rack power is climbing through steps of about 50, 120 and 150 kW where per-rack AC-to-48V shelves plus BBUs no longer scale. In the proposed disaggregated power rack, 416 VAC feeds an AC-to-DC sidecar that distributes +/-400 VDC, scaling delivered power more than 10x versus 48 VDC and supporting racks above 500 kW. The long-term intent is to remove the sidecar and supply the DC bus from building infrastructure.

On cooling, Google reports more than 1 GW of liquid-cooling infrastructure across about 20 data centres with roughly 99.999% rack uptime. A fourth-generation CDU is deploying, a fifth generation targets 1 MW racks, and the Project Deschutes CDU design is being contributed to OCP. The design-for-fungibility priorities named are common flow, pressure and temperature envelopes across IT suppliers and the quality and reliability of wetted materials. The deck also touches briefly on low-carbon concrete and clean alternatives to diesel backup. It is a vendor presentation; the numbers are specific to Google's scale and should be read as an indication of architectural direction, not as general industry data.

Key points

1. 48 V rack power has been widely adopted for about a decade; ML rack power is climbing through ~50, 120 and 150 kW steps where per-rack AC-to-48V shelves plus BBUs no longer scale.
2. Proposed disaggregated power rack: 416 VAC feeds an AC-to-DC sidecar distributing +/-400 VDC; more than 10x delivered power versus 48 VDC; racks above 500 kW.
3. Long-term intent is to remove the sidecar and supply the DC bus from building infrastructure.
4. Google reports more than 1 GW of liquid-cooling infrastructure across about 20 data centres with roughly 99.999% rack uptime.
5. Fourth-generation CDU deploying; fifth generation targets 1 MW racks; Project Deschutes CDU design contributed to OCP.
6. Design-for-fungibility priorities: common flow, pressure and temperature envelopes across IT suppliers and wetted-material quality/reliability.
7. Related: OCP Open Rack V3, OCP Diablo 400 (Mt Diablo) power rack.

Using it in design

• Rack catalogue entries should carry a distribution-voltage attribute (48 VDC, +/-400 VDC, AC); flag 48 V in-rack designs above roughly 150 kW per rack as needing a sidecar or higher-voltage DC architecture.
• Liquid-cooled rack templates should expose CDU capacity and flow/pressure/temperature envelopes as first-class fields.
• In high-density AI halls, reserve space and busway margin for DC distribution and sidecars.

Who should read it

Design engineers planning power and cooling architecture for high-density AI racks, and teams selecting OCP-based hardware.

Related standards

• OCP Open Rack V3
• OCP Diablo 400 (Mt Diablo) power rack

Publication details

Authors

Madhusudan Iyengar

Publisher

Open Compute Project Foundation

Year

2025

Document number

OCP EMEA Summit 2025 keynote, 29 April 2025

Event

OCP EMEA Summit 2025, Dublin

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