Traffic Growth: an Engineering Challenge for Telecoms in the AI Era

According to the State of Disaggregation Report 2025, more than 80% of operators admit their current infrastructure is unprepared for the exponential traffic growth driven by AI/ML workloads and video streaming. Almost half remain unsure whether their networks can sustain the bandwidth demands of real-time AI. Forecasts only reinforce the concern. By 2027, global telecom traffic is expected to surpass 300 exabytes per month, while the AI segment is set to expand at a compound annual growth rate of 32.6% through 2034.

These figures point to more than simple growth. They mark a shift in traffic profiles, where ”hot spots” emerge within hours and traditional scaling models fail to adapt quickly enough.

Engineering Challenges: from uneven traffic to rigid legacy

Traffic growth is non-linear. At PoPs near major data centres, traffic loads can multiply almost overnight, while upgrade windows are narrowing to weeks. Legacy network stacks compound the problem: procurement cycles stretch into months, vendor lock-in slows the adoption of disaggregation, and the risk of overload rises.

The main challenges are well established:

🔸Uneven traffic profiles: AI/ML workloads create bursty flows that demand predictive monitoring to avoid latency spikes.

🔸Compressed scaling cycles: peaks arrive in hours, while upgrades are scheduled quarterly.

🔸Vendor lock-in: closed ecosystems reduce flexibility and inflate CAPEX.

As GSMA Intelligence notes, the combination of AI and 5G is intensifying these trends, making bandwidth optimisation a strategic priority for operators.

Architectural Responses: Flexibility over Monoliths

The industry is moving away from static models towards integrated architectures where transport, VLAN and IX function as a single system. This brings down both CAPEX and OPEX while giving engineers the ability to adapt in real time.

Typical solution stack:

🔸DWDM transport up to 400G: aggregating bandwidth rapidly without major rebuilds.

🔸Capacity VLAN: redistributing flows dynamically across PoPs.

🔸IX platforms with >10 Tbps capacity: acting as a buffer for bursts and balancing traffic across participants.

🔸Distributed PoPs across Europe: direct cross-links to data centres with sub-10 ms latency - a critical factor for edge AI.

At GNM, this stack is already operational: a 20,000 km DWDM backbone interconnects over 650 networks, GNM-IX provides >10 Tbps of available capacity, and distributed PoPs enable new participants to connect in days, not months.

Practical Insights: Forecasting and Automation

Network engineers are increasingly turning to analytics and ML models to predict traffic growth. At GNM, this translates into:

🔸Predictive planning: analysing AI/ML, CDN and OTT patterns to expand growth zones weeks ahead of peaks.

🔸Capacity VLAN + IX: seamless traffic migration with zero downtime, responding within hours rather than days.

🔸Regional equipment hubs: localised resources reduce the time-to-market for new connections to hours.

🔸AI-driven routing: real-time QoS optimisation based on actual load profiles.

Industry adaptation: from reactive to proactive

AI and video are reshaping the traffic equation: traffic dynamics are now more critical than absolute volumes. Legacy architectures are poorly suited to such dynamics. The path forward lies in modular stacks, disaggregation and real-time orchestration.

GNM’s experience shows that:

🔸Scaling must align transport, VLAN and IX;

🔸Resilience depends less on overbuilding capacity than on redistributing it quickly;

🔸Flexibility and forecasting are now as important as raw throughput.

Traffic growth is no longer a distant threat. It is a daily engineering challenge - and one the industry cannot afford to ignore.