How Networks Evolved to Enable AI Infrastructure
Much of the discussion around artificial intelligence focuses on models, computing power, and datacentres.
But there is another layer of infrastructure that is just as critical.
Networks.
The systems that move data.
Without the evolution of telecommunications networks over the past several decades, the AI systems we see today simply would not be possible.
From Voice Networks to Data Infrastructure
Telecommunications networks were not originally designed for data.
They were designed for voice.
Early digital transmission systems, such as Plesiochronous Digital Hierarchy (PDH), were built to carry consistent, predictable traffic, primarily voice circuits.
These systems were:
rigid
hardware-defined
difficult to scale
complex to manage
Each part of the network operated with slightly different timing, requiring additional mechanisms just to keep data aligned.
For the use case at the time, this was acceptable.
For modern data driven systems, it is not.
The Shift to Synchronous Networks
The introduction of Synchronous Digital Hierarchy (SDH) marked a significant step forward.
By synchronising network elements to a common timing source, SDH enabled:
more reliable transmission
simpler multiplexing
improved fault management
large scale backbone network design
SDH networks also introduced self-healing capabilities, allowing traffic to be rerouted automatically, typically within 50 milliseconds, in the event of a failure.
This level of resilience made SDH a foundational technology for national and international telecommunications infrastructure.
However, SDH was still fundamentally designed for constant, predictable traffic.
As data usage began to grow, limitations became apparent.
The Move to Packet-Based Networks
Modern networks have shifted away from circuit based systems toward packet based transport.
This includes technologies such as IP, MPLS, Ethernet, and DWDM.
Rather than allocating fixed bandwidth to specific circuits, packet based networks:
divide data into packets
route traffic dynamically
optimise for variable demand
scale more efficiently
This shift enabled networks to support:
internet traffic
video streaming
cloud computing
and eventually AI workloads
It also marked a broader transition:
From hardware defined systems
To software driven infrastructure
The Reality of Transition: QoS and Legacy Systems
This transition, however, was not seamless.
Earlier technologies such as SDH provided highly predictable, constant bit rate transmission, which was well suited to voice and latency sensitive services.
Packet based networks introduced variability.
Latency could fluctuate.
Packets could arrive out of order.
And jitter became a real issue.
To address this, Quality of Service (QoS) mechanisms were introduced.
QoS allows networks to:
prioritise critical traffic
manage congestion
reduce delay variation
maintain service reliability
This capability is fundamental in modern IP networks and is widely defined in standards such as those developed by the Internet Engineering Task Force (e.g., DiffServ and traffic classification frameworks).
Even with QoS, not all services transitioned easily.
Many legacy systems were originally designed for stable, circuit switched environments.
These included:
lift emergency phones
building alarm systems
traffic control and road network systems
industrial and safety critical communications
These systems relied on:
consistent timing
guaranteed bandwidth
minimal delay variation
Moving them onto packet based networks required:
redesign of service architectures
additional buffering and timing controls
emulation technologies (such as circuit emulation over packet)
or, in some cases, continued use of legacy transport layers
This highlights an important point.
Infrastructure evolution is not just about new capability.
It is about maintaining reliability for existing systems while enabling new ones.
Many of these challenges still exist today, particularly in environments where legacy systems coexist with modern IP based infrastructure.
Why This Matters for AI
Artificial intelligence is fundamentally a data movement problem.
Training clusters involve thousands of GPUs exchanging data continuously.
Inference workloads move data between users, edge systems, and centralised computing environments.
This creates:
high bandwidth demand
unpredictable traffic patterns
massive east to west data movement between datacentres
None of this aligns with the design principles of earlier network generations.
AI requires networks that are:
flexible
scalable
high capacity
dynamically managed
Modern packet based networks provide this foundation.
The Evolution of Infrastructure
The progression from PDH to SDH to modern IP networks reflects a broader infrastructure shift.
Then | Now |
Voice-centric | Data centric |
Fixed circuits | Dynamic routing |
Hardware defined | Software driven |
Predictable traffic | Variable traffic |
Limited scalability | Near unlimited scalability |
This is not just a telecommunications story.
It is an infrastructure story.
A System, Not a Technology
AI infrastructure is not a single system.
It is the combination of multiple interconnected layers:
electricity
datacentres
fibre networks
transmission systems
computing environments
Networks sit at the centre of this system.
They enable data to move between every other layer.
Without them, the system does not function.
A Practical Perspective
Having worked across multiple generations of telecommunications technology, from early circuit based systems through to modern IP networks, I have seen firsthand how these changes have unfolded.
What stands out is not just the technology.
It is the pattern.
Infrastructure evolves in response to demand.
And when demand changes fundamentally, infrastructure must change with it.
The Next Phase
AI is now driving the next stage of network evolution.
This includes:
higher capacity fibre systems
more distributed architectures
tighter integration between computing and network layers
increased reliance on software defined control
Global internet traffic continues to grow at approximately 20–30% per year, driven by cloud computing, video, and AI workloads (as reported by organisations such as Ericsson and industry exchange data).
The networks that supported early internet growth are now being pushed to support AI scale data movement.
The Strategic Question
The question is no longer whether networks can support AI.
It is whether they can scale fast enough.
Because in the same way that electricity is emerging as a constraint.
And datacentres are becoming industrial scale infrastructure.
Networks will determine how efficiently data can move between them.
Closing Thought
AI is often described as a breakthrough in computing.
But it is equally a result of decades of infrastructure evolution.
From voice circuits to global data systems.
From fixed networks to dynamic platforms.
From hardware to software.
The systems behind AI did not appear overnight.
They were built over time.
And they will continue to evolve.
Footnote
This article is part of a series exploring the physical infrastructure behind the AI economy.
Disclaimer
The views expressed in this article are my own and are intended for general information and discussion purposes only. They do not represent the views of any employer, organisation, or client.
© 2026 Rodney Terry – Digital Backbone. All rights reserved