When ChatGPT launched in late 2022, it drew 152 million visits in its first month. Within three months it had crossed one billion, and by the end of 2024 it was approaching four billion monthly visits. That growth, it turns out, was only the opening act.
Speaking at Cisco Live in June 2026, Cisco Chair and CEO Chuck Robbins said AI is on track to triple networking traffic within three years. Analyst Vladimir Galabov, co-host of the AIDC Debate podcast, put an even sharper number on the second wave: a 100x increase in AI traffic by 2030, followed by another 15x jump as agentic AI matures.
Why Networks Are Now Central to AI Infrastructure
Training large language models and running inference both require moving enormous volumes of data across thousands, sometimes hundreds of thousands, of GPUs simultaneously. Sameh Boujelbene, vice president at Dell’Oro Group, framed the issue plainly: “The network is effectively the computer. AI workloads are not just compute-heavy, they are communication-heavy.”
Boujelbene described how AI data centers are moving away from a single network layer toward three distinct fabrics:
- Scale-up networking: Adding more powerful hardware within a rack, so a single rack can handle far greater workloads than before.
- Scale-out networking: Adding more racks to expand total compute capacity across a facility.
- Scale-across networking: Linking multiple data centers so they operate as one unified system for complex AI computations.
Each layer compounds bandwidth requirements. Boujelbene noted that scale-out AI networking can require roughly 10 times the bandwidth of a conventional non-AI network, while scale-up configurations can push that figure closer to 100 times.
Agentic AI Changes the Traffic Profile
Traditional chatbot traffic arrives in peaks and troughs. Agentic AI behaves differently. These systems plan, call external tools, retrieve data, generate outputs, verify results, and act across multiple steps, often continuously. Jeetu Patel, president and chief product officer at Cisco, noted that AI agents consume many times more networking resources than human users do, and that data center operators should treat the resulting pressure as a prompt to review their entire network infrastructure.
As inference workloads begin to outweigh training runs, network pressure also shifts from occasional bursts to always-on production demand, adding yet another layer of sustained load.
Preparing for the Supercycle
Cisco, Arista Networks, Marvell, Nvidia, and Broadcom are all shipping networking equipment aimed at AI-scale workloads. Greg Schulz, founder and senior analyst at StorageIO, acknowledged that upgrading to higher-speed, higher-bandwidth infrastructure carries real cost, but said the expense is becoming increasingly difficult to defer as AI adoption accelerates.
Schulz recommended that operators start with two practical steps. First, build real-time visibility into how networks, compute, storage, and applications are performing together. Second, develop a historical baseline covering workload patterns, resource consumption, and power usage. Both provide the context needed to make informed decisions about when and where to upgrade, before bottlenecks inside or between racks begin to erode GPU utilization and data center economics.
The core message from Cisco Live is straightforward: AI infrastructure planning that stops at GPU counts and cooling capacity is already incomplete. The network fabric is becoming a first-class design consideration, not an afterthought.