Enter the Network Supercycle: Preparing Data Center Networks for AI’s Next Wave

retained text

Enter the Network Supercycle: Preparing Data Center Networks for AI’s Next Wave
At Cisco Live 2026, experts warned: AI’s growth demands a networking supercycle, making networks as vital as GPUs and power.
The release of ChatGPT in late 2022 marked the beginning of an unprecedented surge in AI traffic. Despite still being a basic generative AI chatbot at the time, monthly visits soared immediately to 152 million in November of 2022. From there, it hit one billion visits within three months and almost four billion by the end of 2024. But this was only the first wave of AI.
“The second wave of generative AI unleashed in 2025 will [drive] 100x traffic by 2030,” said Vladimir Galabov, data center expert and co-host of the AIDC Debate podcast. “It will jump by another 15 times due to the growth of agentic AI.”
This amount of traffic requires a great number of graphics processing units (GPUs) and enormous quantities of power. But another area receiving scant attention is the network backbone needed to handle the volume and velocity of AI workloads.
“AI is going to triple networking traffic within three years,” said Cisco Chair and CEO Chuck Robbins during his keynote address at the Cisco Live conference in June of 2026.
For data center operators and designers, the message is clear. AI infrastructure planning can no longer stop at GPUs, power, and cooling. As AI workloads grow larger and more distributed, networking is becoming a larger, more central part of the overall design equation.
Part of the reason for the predicted AI-driven surge in networking traffic is a shift in the kind of AI workload. Requests into chatbots tend to result in highly variable network demand – sudden peaks and troughs of queries and responses. Agentic AI, on the other hand, tends to work more regularly. These agents typically maintain high operational levels over time.
According to Jeetu Patel, president and chief product officer at Cisco, AI agents consume many times more networking resources than humans.
“This networking supercycle means data centers should review their network infrastructure as their volume of AI workloads expands,” Patel said.
Compute Heavy, Communication Heavy
Training of large language models (LLMs) and AI inference both depend on moving huge amounts of data across thousands, and in some cases hundreds of thousands, of GPUs. If the network cannot keep up, utilization falls, latency rises, and the data center's economics break down.
“The network is effectively the computer,” said Sameh Boujelbene, vice president, Dell’Oro Group. “AI workloads are not just compute-heavy, they are communication-heavy.”
She pointed out that AI data centers have fundamentally shifted the playing field – the network is no longer a single layer. It is becoming three distinct fabrics:
Scale-up: Scale-up networking deals with adding more powerful equipment and GPUs within the rack. Thus, one rack can do far more than before.
Scale-out: Scale-out networking means adding more racks to increase the amount of compute performed.
Scale-across: Scale-across networking means setting it up so multiple data centers work in tandem – effectively operating as one huge computer to tackle complex AI computations.
Boujelbene added that AI traffic will not grow linearly. Due to these network layers and the nature of evolving AI applications, traffic will compound.
“Each layer adds a major bandwidth multiplier versus traditional data center networks,” she said. “Scale-out AI networking can require roughly 10X the bandwidth of a conventional non-AI network, and scale-up can push that requirement closer to 100x.”
Bigger models, for example, create more east-west traffic inside clusters. As inference workloads begin to dominate training workloads, network pressure shifts from occasional training runs to always-on production workloads.
Agentic AI adds yet another multiplier, because agents do not just answer once; they plan, call tools, retrieve data, generate, verify, and act across multiple steps. For instance, network load rises with model size, user demand, and workflow complexity.
Networking Refresh Considerations
Cisco recently rolled out a series of new networking, switching and routing products that are designed with AI in mind. Vendors such as Marvell, Arista Networks, Nvidia, and Broadcom are also offering AI networking equipment.
Greg Schulz, founder and senior analyst with the StorageIO consulting firm, noted that such upgrades to raise network speed and bandwidth will come at a price. But as AI usage grows, it is an expense that is increasingly difficult to postpone.
“Networking the many IT resources together for AI is a bottleneck for traffic that is both within racks, clusters, or data centers as well as between racks, clusters, and data centers,” said Schulz. “In addition to a greater volume of network traffic, there will also be increased traffic velocity as well as tremendous variance in the size and type of traffic associated with AI applications and workloads.”
How to Prepare for the Network Supercycle
What should data center designers and operators be thinking about on the networking side as they make plans to add highly dense servers and racks packed with GPUs?
Monitor: Schulz recommended that they gather timely insight and awareness into how their networks, compute/GPUs, storage, and software/apps are performing.
Build historical visibility: This also means having a historical perspective on app workloads, their resource usage and power/cooling demands.
Prepare for agent-to-agent and private AI: “Data centers should assume they will be seeing a lot more agent-to-agent activity, more models being trained and deployed privately and locally, as well as an increase in other AI activity and traffic across cloud and resource providers,” Schulz said.
Use AI-enabled management tools: This will require not only having more power, cooling, and networking resources, Schulz said, but also having an AI-enabled network and data infrastructure with resource management tools to detect anomalies, load balance traffic, implement prioritization and optimizations, and forecast future resource needs.
Matching the Network to AI Needs
Currently, considerable attention is being paid to obtaining and using as many GPUs as possible. That inevitably leads to a discussion about how to gain access to more power and how to keep those GPUs cool. Attention is now turning to networking, which has received far less consideration.
“The real question is not just how many GPUs can fit in a rack,” said Boujelbene. “It is whether the network fabric can keep those GPUs highly utilized, synchronized and productive at scale.”
Her advice to data center designers is to shift their thinking from the traditional view of networking as mere back-end plumbing.
“In dense GPU clusters, the network is part of the AI system,” she said. “The focus should be on bandwidth, latency, congestion control, topology, optics, telemetry, and resilience from day one.”

document fetched 2026-07-08 08:32 · canonical https://datacenterknowledge.com/networking/enter-the-network-supercycle-preparing-data-center-networks-for-ai-s-next-wave