Overcoming The Energy Tax on AI

This brief assesses the industry's efforts to reconcile the gigawatt-scale power requirements of artificial intelligence with the physical and economic constraints of existing telecom and data center infrastructure. We use news analysis, briefings, and technical papers from last month's OFC 2026 event in Los Angeles to fuel our summary and trend analysis.

Energy consumption has transitioned from a secondary operational concern to a primary strategic risk for both telcos and data center operators. For telcos, energy represents 3 to 5 percent or more of total operating expenditure, while in the data center world, that figure is significantly higher and rising. With the acceleration of climate change and continued energy market volatility driven by regional conflicts, such as the Iran war, the focus of network engineering has shifted from peak capacity to performance per watt. Sustainability targets are no longer discretionary commitments; they are binding constraints imposed by investors, senior leadership, and network engineers. And both telecom networks and data centers matter. While hyperscaler energy consumption is growing rapidly even as telco use has plateaued, the telecom sector consumes significantly more energy on an absolute basis: 340 terawatt hours in 2024 vs. 189 TWh for hyperscalers. The telecom sector's carbon footprint also exceeds that of hyperscalers: telco greenhouse gas emissions across all three scopes totalled 342 million metric tons of CO2-equivalent in 2024, compared with 229 million for hyperscalers.

The latest data from OFC 2026 makes it clear that power consumption is the biggest challenge for new networks. Companies such as NVIDIA, Cisco, Broadcom, Marvell, and Coherent are exploring co-packaged and linear architectures because traditional ways of handling optics use too much energy. While some new tech is hitting sub-picojoule efficiency levels or saving 97 percent on power in specific network types, other areas are struggling. Ultra-wideband systems currently have a 48 percent energy penalty, and new designs are still very complex. For anyone building large AI clusters, efficiency is no longer a choice; it is a requirement for survival.

Organizations mentioned:

3M
AMD
Arista
Broadcom
China Telecom
China's State Key Laboratory of Photonics and Communications
Ciena
Cisco
Coherent
Eoptolink
Ericsson
Google
Hyperlight
Keysight
Lightmatter
LightXcelerate
Lumentum
Marvell
Meta
National Tsing Hua University
Nokia Bell Labs
NTT
NVIDIA
OIF
Semtech
Telefonica
TIM
TSMC
UMC
University College London
University of Melbourne
University of Tokyo
Wavetek

Product Code: TES-10042026-1

Summary

Recommendations for industry players

For vendors
For operators

Key themes and announcements from the event

OIF interoperability showcase: energy efficient interfaces
Cisco: Silicon One G300 and the 100T barrier
Coherent: Pioneering waste heat recovery and CPO versatility
NVIDIA: Securing the optical supply chain and the road to Feynman
Broadcom: Taurus and Tomahawk 6
Marvell and Arista: Linear-receive and transmit-retimed optics
Hyperlight and Lightmatter: next-generation materials and fabrics
Semtech, Lumentum, & Eoptolink: low-power, high-speed optical for AI scaling

Implications for market development

Energy efficiency as a primary KPI in data centers
Energy consumption in telecom: The push for dynamic networks
The shift toward liquid cooling and XPO
The diversification of the interconnect roadmap
From heat recovery to electricity generation

Energy-related research findings at OFC 2026

Data centers
Telecom networks

Conclusion: Managing the energy bottleneck

Appendix