Every major shift in business communications places new demands on the network. In this two-part series, we first look at how AI may change traffic patterns, expose hidden bottlenecks and test the limits of established techniques such as Quality of Service and prioritisation. In Part Two, we examine the main response options in more detail, from targeted upgrades and additional fibre through to CWDM, DWDM and multi-generation PON architectures.
VoIP introduced concerns around delay, jitter and call quality. Video conferencing increased bandwidth consumption. Cloud services changed where applications and data were hosted, and how users accessed them.
AI is now prompting a similar discussion.
Can existing networks cope with the additional demand, or will organisations need to invest in more capacity?
The answer will not be the same for everyone. Nor should the first response always be a major network upgrade. In many cases, the starting point is to make better use of the capacity already available.
However, AI may also require organisations to look beyond headline bandwidth figures. The volume of traffic matters, but so do its direction, duration and behaviour.
How networks have traditionally adapted
When new applications become widely used, they begin competing with existing traffic.
Some services are more sensitive than others. Voice uses relatively little bandwidth, but is highly sensitive to delay, packet loss and jitter. Video needs more capacity and consistent performance. A large file transfer may consume significant bandwidth but can usually tolerate taking longer to transit the network.
This is where Quality of Service, or QoS, can help.
QoS can protect delay-sensitive traffic. Prioritisation can favour critical applications. Traffic shaping can control how bandwidth is consumed during busy periods. VLANs, VRFs and other forms of logical separation can also isolate services or user groups.
These are valuable techniques. They help organisations use their networks more efficiently and maintain service quality as demand increases.
However, they have an important limitation. They manage the capacity that already exists. They do not create more of it.
Average utilisation can hide the problem
Network capacity is often assessed using average utilisation. This is useful, but it may not tell the whole story.
A link may average only 40 or 50 per cent utilisation across the day yet still experience short periods of congestion. These peaks can cause packets to queue, increase latency and affect application performance.
The issue can be particularly difficult to identify when traffic is bursty. A large volume of data may arrive over a short period rather than being spread evenly throughout the day.
Monitoring intervals can hide this. A five-minute average may smooth out a congestion event lasting only a few seconds. Those few seconds may still be enough to affect a voice call, disrupt a video session or slow a time-sensitive application.
Capacity planning should therefore consider peak utilisation, queue behaviour and latency under load, rather than relying on a single average figure.
How AI may change the traffic profile
AI will not affect every organisation in the same way.
Employees accessing a cloud-hosted AI assistant may generate very little additional traffic. That type of adoption is unlikely to transform an organisation’s optical networking requirements.
The picture can be very different when AI relies on large datasets, distributed storage or specialist computing resources.
Data may need to move between servers, storage platforms, cloud environments and data centres. These transfers can be large, frequent and sustained. A small number of high-volume transfers, sometimes described as “Elephant flows”, may consume a significant share of the available capacity.
Traditional enterprise networks have often been designed around north-south traffic. This is data moving between users and applications, or between the organisation and the internet.
Modern data-centre and AI environments can generate much more east-west traffic. This is traffic moving between servers, storage systems and processing environments.
That distinction matters. A network designed mainly around user access may not be optimised for large volumes of data moving between infrastructure platforms.
AI may therefore change more than the amount of traffic. It may change where traffic flows, how long transfers last and where capacity is needed.
Look beyond the obvious link
A bottleneck is not always found on the most visible connection.
An external circuit may have adequate capacity, while congestion exists within the switching infrastructure behind it. The issue may be an oversubscribed uplink, an inefficient path or several access links feeding into a lower-capacity aggregation point.
Oversubscription is not inherently poor design. Most users and systems do not transmit at full capacity at the same time. It becomes a problem when traffic patterns change and simultaneous demand exceeds the assumptions made when the network was designed.
This is why increasing the capacity of one connection may not solve the problem. Capacity needs to be considered across the complete path.
What are the options?
Once the real constraint has been identified, there are several possible responses.
The first is to optimise the existing network. QoS, traffic shaping, routing changes and logical separation may improve performance without major investment. This is cost-effective where the issue is contention or poor traffic management, but less effective where capacity is genuinely exhausted.
The second option is a targeted upgrade. Increasing the speed of a constrained interface or replacing an oversubscribed uplink may remove a specific bottleneck. The risk is that the bottleneck simply moves elsewhere.
A third approach is to add more capacity through higher-speed circuits or additional fibre. This can be straightforward where infrastructure is available, but costly or impractical across longer distances.
The final option is a more scalable optical platform. CWDM and DWDM can increase capacity over existing fibre and provide a clearer route for future growth. PON architectures may also allow several generations of active technology to share the same passive fibre infrastructure.
No single option is automatically right. The correct response depends on traffic behaviour, fibre availability, distance, resilience and expected growth.
AI is encouraging organisations to review the infrastructure supporting their applications and data. That is worthwhile. However, decisions should be based on evidence rather than headlines.
In Part Two, we look more closely at the strengths, limitations and trade-offs of the available options.
Discuss your network requirements
To discuss your current network requirements, capacity challenges or future expansion plans in more detail, contact the FTL technical sales team.
Telephone: 01344 752222
Email: sales@fibre.co.uk
