Why Global Networks Aren’t Prepared for the AI Surge

This Industry Viewpoint was authored by Michael Reid, CEO at Megaport

AI has moved decisively into the enterprise mainstream; what was once an emerging capability now plays a foundational role in how organizations operate, compete, and respond to shifting market forces. While the world has been focused on the “brain” of AI –  the LLMs and the massive GPU clusters that train them – we have overlooked its “nervous system,” the network infrastructure required to keep AI alive.

As we move through 2026, the scale of this oversight is becoming clear. Estimates show that token usage doubled from 6.4 trillion to over 13 trillion in the first month of this year alone. As AI grows, it consumes more than just power and compute; it demands a level of network capacity and architectural agility that most existing systems simply were not built to provide.

How AI Traffic Differs from Traditional Workloads

Historically, global networks were designed for predictable, asymmetrical workloads. Think legacy video streaming, SaaS applications, and standard cloud migrations, where traffic patterns are relatively stable.

AI traffic behaves like an entirely different species; highly distributed and notoriously latency sensitive. Whether it is a massive training run or a split-second inference request, these workloads strain networks at every layer. As a result, the infrastructure is being pushed toward a tipping point because AI saturates it in ways that traditional architectures cannot sustain.

The Mechanics of Demand: Inference and Scattered “Brains”

When we think about what is actually driving AI’s unprecedented demand, much of it comes down to inference. Every time a model generates a line of code, a medical diagnosis, or a customer service response, it has to reference its training data and calculate a probability-based output. This process occurs billions of times a day.

Each time a user submits an AI prompt, it triggers a cascade of network activity. The “brain” of the model is effectively scattered across the map, requiring systems to retrieve data and coordinate across disparate tools and sources in real-time.

In the AI economy, latency is the enemy of utility. LLMs are products, and maintaining user engagement requires razor-thin response times. When enterprises are competing for the same low-latency channels, supply quickly runs out. Without immediate expansion and optimization, the very speed that makes AI valuable will become its greatest downfall.

Old Networks, New Pressure

As AI traffic proliferates, the constraints of traditional network architectures are becoming clear. Legacy interconnect and Data Center Interconnect (DCI) approaches are increasingly being pushed to their limits, with bottlenecks emerging.

We are also witnessing a fundamental shift in traffic directionality. For decades, North-South traffic (data moving between a data center and an external device) was the priority. Today, it is being outpaced by East-West traffic (data moving between servers or data centers).

Specific sectors are already feeling the heat. For example, in the financial sector, Virtual Cross Connect (VXC) capacity has grown at twice the rate of other industries to keep up with AI-driven high-frequency analysis. Some AI-heavy corridors have even seen traffic grow tenfold in just 24 months. If these trends continue, with some projections suggesting a 700% increase in Wide Area Network (WAN) traffic by 2034, the “token tide” will outpace existing capacity.

A Blueprint for AI-Ready Infrastructure

Enterprises cannot solve this problem by simply throwing more bandwidth at it. Addressing these challenges requires a thoughtful restructuring. Successful organizations are already moving toward a hybrid infrastructure model.

In this new framework, public clouds continue to support general workloads, but sensitive, AI-heavy tasks are transitioning back to specialized bare metal providers or private data centers. This diffused demand model allows for better control over performance and security.

We are also seeing the rise of edge-first architectures. By moving latency-sensitive operations, like inference, closer to the end-user, enterprises can bypass the congested core of the internet. Providers are also beginning to build dedicated connectivity fabrics, which are high-speed optical links that treat distributed GPU clusters as a single logical system rather than isolated silos.

The urgency is compounded by the shift toward AI Agents. Unlike a human asking a single question, agent-to-agent communication involves a constant, high-volume exchange of tokens as autonomous systems coordinate tasks. This agentic future, championed by leaders like Nvidia, will demand even more from the network than the first wave of chatbots did.

Here at Megaport, we have already observed certain channels experiencing over 100% increases in traffic due to early agent integration. This means, for the enterprise, your AI strategy is only as good as your network. To thrive in an era of 13-trillion-token months, organizations must stop viewing the network as a utility and start treating it as the primary competitive advantage.

If you haven't already, please take our Reader Survey! Just 3 questions to help us better understand who is reading Telecom Ramblings so we can serve you better!