Enterprise networks are facing a dramatic shift in traffic patterns as generative AI (genAI) applications proliferate across organizations. While traditional network architectures were designed around asymmetric data flows, with small requests going up and large responses coming down, the rise of genAI workloads is completely inverting this model. Network architects and administrators need to adapt quickly to this new reality or risk significant performance issues.

“GenAI applications are fundamentally chatty, and users expect quick responses,” explains Kishan Ramaswamy, director of product management for Broadcom’s VeloCloud division. “This creates two distinct characteristics that impact network design: data flows and encryption.”

GenAI creates new traffic patterns

The first major shift comes from how genAI applications handle data flows. Traditional web applications follow a predictable pattern: users send small requests upstream and receive larger responses downstream. This asymmetric model has shaped network design for years. However, genAI-powered productivity tools like code generators are dramatically different.

“When using code generators, we start to send much larger amounts of data upstream,” Ramaswamy notes. “You might send a substantial chunk of code to a generative AI application for completion or optimization, and it will return that amount of data back or even more.” This shift toward symmetric or even inverted traffic patterns has serious implications for network infrastructure, particularly in terms of bandwidth planning and quality-of-service (QoS) policies.

The challenges don't stop at traffic patterns. As Ramaswamy points out, genAI traffic is encrypted. “With encryption, it becomes difficult from the start to detect what the application is. The traditional methods of looking at source address, IP address, destination address, and then making inferences is not good enough anymore. And if I don't know what that application is, I can't provide it with the appropriate quality of service.”

This opacity requires a new approach to traffic management. Machine learning techniques can analyze traffic patterns, flow characteristics, and application behaviors to identify genAI workloads with high confidence, essentially using AI to detect AI.

The invisible AI challenge

Security remains a crucial consideration in this new landscape. Modern solutions must incorporate both cloud-based and on-device security features, including next-generation firewall (NGFW) functionality, URL filtering, and AI-powered intrusion prevention systems. These security measures must work seamlessly with traffic optimization features to ensure both performance and protection.

What network administrators face today is what some are calling the “invisible AI challenge” — the need to effectively manage encrypted AI traffic that can't be directly observed. Success requires analyzing context and behavior patterns to make intelligent routing and prioritization decisions, ensuring optimal performance for business-critical applications while maintaining network efficiency.

Organizations that want to prepare their networks for increasing genAI workloads should focus on solutions that provide detailed visibility into genAI application usage and can predict future network needs. Key features should include dynamic traffic optimization, support for network slicing, advanced analytics, and integrated security capabilities.

The VeloCloud advantage

Solutions such as the VeloRAIN (VeloCloud Robust AI Networking) architecture, which underpins VeloCloud SD-WAN and the VeloCloud portfolio, are central to managing these new traffic patterns. VeloRAIN will enhance VeloCloud Dynamic Multipath Optimization (DMPO), which helps organizations continually analyze and optimize traffic based on application requirements and network conditions, by incorporating AI intelligence.

Network slicing, particularly important in 5G environments, allows organizations to create dedicated pathways for different types of genAI workloads. Communication service providers can now offer multiple service levels, or “slices,” optimized for different types of traffic.

“Part of the VeloRAIN architecture is Dynamic Application-Based Slicing, or DABS,” Ramaswamy explains. “We identify applications and intelligently route them across different network links based on type and priority. When combined with carrier-provided network slicing in 5G networks, this creates a powerful solution for managing complex traffic patterns.”

Intelligent application profiling through advanced analytics tools helps organizations understand how AI applications behave on their networks and adjust accordingly. With VeloCloud, network administrators receive detailed telemetry data about link performance and application impact, helping them identify which critical applications are affected during network degradation. Predictive analytics help forecast potential network issues and capacity constraints, particularly important given the dynamic nature of AI application traffic.

Rethink your network design

As genAI applications continue to proliferate across enterprise networks, the ability to effectively manage their unique traffic patterns will become increasingly crucial for maintaining business operations. Organizations that invest in forward-looking network management solutions today will be better positioned to handle the demands of tomorrow's AI-driven workplace.

The impact of AI on network traffic patterns is likely to grow as more organizations adopt AI-powered tools and applications. “Many vendors have not really thought through how AI applications will transform the way networks operate,” Ramaswamy observes.

“Companies designed networks based on the apps running today, but they also need predictive analytics to understand new consumption and traffic patterns, and new tools to optimize genAI app performance. Enterprise IT teams need to rethink their networks to accommodate the genAI wave.”