How 6 AI Attributes Change Data Center Design

Overview

The paper identifies six attributes: workload type, accelerator network communication, accelerator TDP, accelerator peak power, synchronous computation and cluster size. Together they push rack densities past 100 kW and create peaky, synchronous electrical loads. Most traditional data centres support peak racks of about 10 to 20 kW; AI clusters use tens or hundreds of racks above 40 kW, and density trends past 100 kW toward 1 MW.

The GPU TDP trend is V100 300 W, A100 400 W, H100 700 W, B200 1,000 W and B300 1,400 W. Accelerators can exceed steady-state TDP (for example by about 50%) for tens of milliseconds, and synchronous training swings rack power sharply. A packing example: 10U air-cooled 8-GPU servers give 32 GPUs per 42U rack, while 2U liquid-cooled 8-GPU servers give 168 GPUs. In an inference example, a 600 GB model needs 1 server (about 10 kW per rack) and a 2.4 TB model needs 4 servers (about 40 kW per rack).

Recommendations fall in three areas. Power: larger power blocks, higher voltages, arc-flash and transient handling. Cooling: a move to direct-to-chip liquid cooling with CDUs. Racks: bigger, stronger cabinets, at least 750 mm wide, 1,200 mm deep and 48U, with static load above 2,270 kg, dynamic load above 1,600 kg and a slab rated above 3,000 kg per rack. The paper also covers energy procurement, supply chain and software tools such as EPMS and DCIM. Rack weight is stated as over 1,300 kg in the key takeaways but over 1,800 kg in the rack section (an inconsistency); use the higher value.

Key points

1. Six AI attributes: workloads, accelerator network communication, accelerator TDP, accelerator peak power, synchronous computation, cluster size.
2. GPU TDP trend: V100 300 W, A100 400 W, H100 700 W, B200 1,000 W, B300 1,400 W.
3. Accelerators can exceed steady-state TDP by about 50% for tens of milliseconds; synchronous training swings rack power sharply.
4. Traditional data centres support peak racks of about 10 to 20 kW; AI clusters use tens or hundreds of racks above 40 kW, trending past 100 kW toward 1 MW.
5. Packing example: 10U air-cooled 8-GPU servers give 32 GPUs per 42U rack; 2U liquid-cooled ones give 168 GPUs.
6. Recommended AI racks: at least 750 mm wide, at least 1,200 mm deep, 48U; static load above 2,270 kg, dynamic above 1,600 kg; slab above 3,000 kg per rack.
7. Rack weight is stated as over 1,300 kg in key takeaways and over 1,800 kg in the rack section (inconsistent); use the higher value.
8. Inference example: a 600 GB model needs 1 server (about 10 kW per rack); 2.4 TB needs 4 servers (about 40 kW per rack), at 8x80 GB GPUs per server.

Using it in design

• Add an AI cluster design profile with rack defaults of 750x1,200 mm, 48U, static 2,270 kg and dynamic 1,600 kg.
• Floor-loading check: at least 3,000 kg point load per rack.
• Make liquid cooling the default above 40 kW per rack.
• In electrical design, flag synchronous load swings: size UPS and generator for transient peaks of about 1.5x TDP over tens of milliseconds, and recommend energy storage or filtering.
• Keep a GPU TDP catalog (H100 700 W, B200 1,000 W, B300 1,400 W) so rack kW can be estimated from GPU counts.

Who should read it

Engineers and infrastructure planners designing power, cooling and racks for AI clusters.

Related standards

• ASHRAE TC9.9 Liquid Cooling Guidelines
• OCP Open Rack
• NFPA 70E
• IEC 60364

Publication details

Authors

Wendy Torell; Maria A. Torres Arango

Publisher

Schneider Electric

Year

2026

Document number

WP 110

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