For years, service providers could plan network growth with a reasonable sense of predictability. More people came online. More devices connected. More video moved across the network. The pattern was familiar: plan for higher demand, more downstream traffic and the usual peak-time pressure.
AI changes the pattern, forcing service providers to rethink how network infrastructure can best support all of the new user cases and opportunities.
The next wave of demand will not come only from people clicking, watching or downloading. It will come from software agents working on behalf of people and businesses. They will search, reason, retrieve, compare, transact and repeat. Not only when a user is staring at a screen but also continuously.
This also creates a different planning challenge. It is more than just traffic – the very behavior of that traffic changes and becomes more closely tied to the intelligence of the services customers use.
AI traffic does not look like web traffic
Cisco’s AI Impact on Networks report shows that while AI inference traffic is small today, it is set to grow at a pace we’ve never seen before. Token consumption grew nearly 10x year over year, and service provider measurements showed roughly 4x AI inference traffic growth in eight months.
Inference is the moment when a model responds to a prompt, generates an answer or helps an agent decide what to do next. As AI moves into enterprise applications, and connected devices, inference traffic becomes a bigger part of the network mix.
The issue is not just volume. AI flows are different. They last about twice as long as regular web transactions. They are smoother and more sustained, rather than short and bursty. They also challenge the old assumption that most traffic is downstream. About 9% of AI inference flows carry more upstream than downstream traffic, compared with about 0.5% for web traffic.
That is a signal for service providers. Networks optimized for human-paced consumption now have to support machine-paced interaction.
Agents will change the demand curve
The real inflection point is agentic AI.
A traditional application usually waits for a human to act. An agent does not. It can break work into steps, call tools, collect context, query models, interact with systems and keep going until a goal is complete.
In Cisco’s analysis, an agent completing a task generated up to 450% more total traffic than a human performing the same task manually. Around 70% of that incremental traffic was AI inference.
That is why the connection between the agent and the model becomes critical. If that path slows down, the agent slows down. If it fails, the agent fails. The network is no longer just carrying the experience. It becomes part of the experience.
For service providers, this has implications across capacity, assurance, security and architecture. AI traffic will need to be identified, measured, prioritized and protected differently.
Capacity planning needs new assumptions
For service providers, the opportunity is not only to rethink their own networks. It is also to help enterprise customers understand what agentic AI will mean for their applications, traffic patterns, and user experience. As more businesses embed agents into customer service, operations, software platforms, and digital workflows, they will need partners who can help them plan for performance, resilience, and scale. That puts service providers in a stronger position to move from connectivity provider to strategic infrastructure partner.
Therefore, the planning question is becoming more urgent: what happens when demand is driven not only by people, but by agents that work continuously?
That shift changes the baseline. Without agentic AI, enterprise network traffic is projected to grow about 2.5x over the next decade. With agentic AI, that growth could reach about 9x.
Consumer traffic changes. AI and agentic AI are projected to drive total network traffic to 6.6x today’s levels by 2035, with AI inference expected to account for 25% of total traffic.
And this is not limited to enterprise networks. As AI becomes embedded in consumer apps, connected devices, search, shopping, entertainment, and personal assistants, service providers will see the same pattern show up across broader internet traffic.
These are not just bigger numbers – they’re also architecture signals. Planning models built around downloads, peak viewing hours and downlink-heavy traffic will need to adapt to longer flows, more upstream context, distributed inference and critical paths.
Latency is part of that story too. Today, model processing dominates AI response time, while network latency is a smaller component. But as inference hardware improves, the network will become more visible in the experience. Measuring AI-specific performance will become essential for service providers.
The opportunity ahead
This is also an opportunity for service providers to lead.
The providers that understand AI traffic early can build better capacity plans, improve assurance, offer AI-ready connectivity and become trusted partners for enterprises adopting agents at scale.
The future internet will not simply connect more users. It will connect more digital workers, autonomous workflows and intelligent services running all the time.
That is the network supercycle ahead.
The Cisco AI Impact on Networks report provides a better understanding of the data, projections and planning implications for service providers.
