Powering the AI Revolution: Building the Infrastructure That Makes Intelligence Possible (Manja Thessin)

Overview

The article appears in BICSI ICT Today, Vol. 47, No. 1 (Jan/Feb/Mar 2026), p. 30. The author argues that AI data centers must be planned as an integrated system of compute, data movement and energy rather than as siloed disciplines.

It notes that training compute grew about 1.5x per year from 1950 to 2010 and roughly 4.3-4.6x per year since 2010 (Epoch AI data). 150 kW racks are described as standard deployments, with roadmaps beyond 200 kW, affecting airflow, floor loading and service access. Meta and others report 30-40% of AI system time lost to network wait states, with tail latency driving utilization.

On energy, data centers used about 536 TWh in 2025 (about 2% of global electricity), projected to reach 1,065 TWh by 2030, which makes power procurement strategic. Later pages address fiber density: some AI clusters need 2x, 4x or even 10x the fiber of previous hyperscale sites, with per-rack counts reaching several thousand and some sites exceeding one million terminations.

Hollow-core fiber can cut propagation delay by up to 40%, while single-mode is winning new builds for reach and coherent-optics compatibility. The related source is IEEE 802.3dj. The article is an industry viewpoint; it gives trends and scale rather than detailed design rules.

Key points

1. Training compute grew about 1.5x per year 1950-2010 and roughly 4.3-4.6x per year since 2010 (Epoch AI data).
2. 150 kW racks are described as standard deployments, with roadmaps beyond 200 kW, affecting airflow, floor loading and service access.
3. Meta and others report 30-40% of AI system time lost to network wait states.
4. Data centers used about 536 TWh in 2025 (about 2% of global electricity), projected to reach 1,065 TWh by 2030.
5. Some AI clusters need 2x, 4x or 10x the fiber of previous hyperscale sites; per-rack counts reach several thousand and some sites exceed one million terminations.
6. Hollow-core fiber can cut propagation delay by up to 40%.
7. Single-mode is winning new builds for reach and coherent-optics compatibility.

Using it in design

• For AI templates, default to single-mode fiber, high-count trunks and dedicated back-end fabric pathways.
• Size floor loading and service clearances for 150-200 kW racks.
• Show fiber termination counts per rack in the network BOM.
• Run compute, network and power planning as one integrated process.

Who should read it

Designers and decision makers planning power, cabling and rack density together in AI data centers.

Related standards

• IEEE 802.3dj

Publication details

Authors

Manja Thessin

Publisher

BICSI ICT Today

Year

2026

Document number

Vol. 47, No. 1 (Jan/Feb/Mar 2026), p. 30

Document language

English

Source type

Standard

Access

Free to read

Licence

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

Citation