Cisco released a new research report warning that the rapid rise of AI inference and Agentic AI will fundamentally alter traffic behavior across wide area networks, forcing operators and enterprises to rethink capacity planning, resiliency, Quality of Service (QoS), observability, and upstream bandwidth assumptions. The report, titled “AI Impact on Wide Area Networks: Cisco Report 2026,” combines live service provider traffic measurements, Cisco Crosswork Assurance telemetry, and empirical testing of AI agents to analyze how AI-generated traffic differs from traditional web traffic.
Cisco argues that the next phase of networking will center on what it calls the “spinal cord” connecting AI agents to large language models (LLMs). Unlike conventional web traffic optimized for human-driven interactions and bursty downstream video delivery, Agentic AI creates persistent machine-to-machine communication patterns operating continuously at software speed. Cisco projects that AI inference traffic could account for approximately 25% of all network traffic by 2035. The company says AI inference flows already exhibit materially different transport characteristics: they last roughly 2x longer than conventional web transactions, generate lower but steadier throughput, and increasingly shift traffic symmetry toward upstream-heavy flows because AI prompts contain expanding contextual state.
One of the most significant findings concerns the network effects of Agentic AI. Cisco tested an Open Deep Research-style AI agent and found that agent-driven task execution generated approximately 450% more traffic than the same activity performed manually by a human user. About 70% of the incremental traffic came from AI inference communications between the agent logic and the AI model itself. Cisco describes this inference connection as the “agent’s spinal cord,” arguing that latency, congestion, or outages on these paths directly impair agent functionality. The report suggests that as enterprises deploy autonomous software agents for research, workflow orchestration, customer operations, ERP systems, and autonomous decision-making, AI inference paths will become strategic infrastructure assets requiring differentiated treatment, resiliency engineering, and continuous telemetry.
Cisco also identified a structural change in traffic asymmetry. Roughly 9% of AI inference flows already carry more upstream traffic than downstream traffic, compared with only about 0.5% for traditional HTTP web flows. The company expects this imbalance to intensify as AI agents accumulate larger contextual histories, continuously exchange state information, and coordinate with multiple external tools and services. This trend could have major implications for mobile radio planning, broadband access architectures, enterprise WAN design, and peering economics, where networks historically optimized for downstream-heavy consumer traffic.
The report also concludes that AI traffic will force changes in security and traffic engineering infrastructure. AI inference traffic increasingly uses QUIC in addition to TCP, with QUIC already accounting for 57% of measured AI inference traffic volume in Cisco’s data set. Cisco warns this creates visibility challenges for Deep Packet Inspection (DPI) systems because of encryption and transport-layer evolution. Meanwhile, flow-aware devices such as firewalls and intrusion detection systems will need to scale significantly because AI inference sessions persist longer and maintain more state than traditional web sessions.
Cisco projects that enterprise traffic could grow roughly 9x by 2035 if Agentic AI adoption accelerates as expected, compared with approximately 2.5x growth without widespread AI agents. On the consumer side, Cisco expects AI and Agentic AI to drive total internet traffic growth to 6.6x current levels by 2035, representing roughly 63% additional growth compared with non-AI projections. The report further predicts that the most aggressive growth phase for AI-driven traffic will occur between 2029 and 2032 as Agentic AI adoption enters mainstream enterprise and consumer workflows.
Cisco notes that network latency is not yet the dominant bottleneck for AI experiences because inference processing time still far exceeds transport latency. The report estimates AI inference responses commonly range from hundreds of milliseconds to several seconds, whereas network latency often remains in the 20–50 millisecond range. However, Cisco expects network latency sensitivity to rise sharply as inference hardware accelerates and token generation speeds improve, increasing the importance of edge inference placement, path optimization, and WAN assurance.
“This report marks the first step in ongoing research on the impact of AI on network traffic,” Cisco wrote. “As AI adoption deepens across enterprises and consumers, future research work will refine projections, validate emerging patterns, and guide operators in evolving their networks for an AI-driven world.”
🌐 Analysis: Cisco’s report provides one of the industry’s earliest quantitative looks at how Agentic AI could transform traffic engineering assumptions across WANs and service provider networks. Much of the AI infrastructure discussion over the past two years has focused on GPU clusters, optical interconnects, Ethernet fabrics, and AI data center scaling. Cisco shifts attention toward the WAN and access-network consequences of autonomous AI systems operating continuously at machine speed. The findings align with a broader industry shift toward persistent AI workflows, where agents communicate not only with users but with APIs, databases, enterprise tools, vector stores, cloud services, and other agents.
🌐 The report also reinforces why networking vendors increasingly position observability, telemetry, and assurance as critical AI-era infrastructure layers. As AI inference traffic becomes longer-lived, upstream-heavy, encrypted, and latency-sensitive, traditional “best effort” traffic engineering models may become insufficient. Cisco’s emphasis on the “agent spinal cord” concept parallels emerging discussions across hyperscalers and AI infrastructure providers around inference fabrics, distributed AI orchestration, and persistent low-latency connectivity between AI reasoning engines and external tool ecosystems.
Download the full Cisco report: https://www.cisco.com/c/dam/en/us/solutions/collateral/artificial-intelligence/mass-scale-infrastructure/ai-network-traffic-report.pdf
Readers are encouraged to download the full report for detailed traffic models, enterprise and consumer AI adoption projections, flow analysis, and WAN engineering implications.
