From KPI-Driven to Workload-Driven: AI Traffic Is Rewriting 6G Requirements, Making Testbeds and Architectural Flexibility Essential

For more than a decade, mobile evolution has often been narrated through a handful of peak KPIs: higher data rates, lower latency, and more connections. In the transition window from 5G-Advanced toward 6G, AI workloads are breaking this KPI-first framing. Inference services, agentic pipelines, multimodal interactions, and edge-cloud coordination reshape networks not by pushing a single metric, but by changing request distributions in more complex and less predictable ways.

NGMN’s recent publication argues that the rapid expansion of AI and AI agents creates both opportunity and uncertainty for operators. The operator message is clear: keep flexibility during 6G study phases and avoid locking AI-related capabilities into a single early design choice. When demand shifts quickly, standardization needs architectural room for iteration. The problem becomes uncertainty management, not KPI maximization.

In parallel, 5G-MAG announced an open-source reference tools project, “6G Testbed & AI Traffic Characterization.” The stated goal is practical: provide a reproducible environment to characterize how AI engines consume network resources and tolerate network impairments, turning broad claims about “AI changing networks” into measurable workloads and comparable experiments.

Why does “workload-driven” become a pivotal 6G shift? Because the dominant pain points of AI services are often tail latency and bursty concurrency, not average throughput. Interactive inference runs under strict latency budgets where user experience correlates with P95/P99 behavior. Agentic workflows add serial tool calls and long chains, amplifying small network jitter into end-to-end waiting. Networks must treat tail behavior and burst management as first-class design goals.

A second shift is that traffic directionality and UL/DL balance become less stable. Video-era assumptions optimized for downlink. AI-era inputs can be multimodal uplink, outputs may be continuous streaming, and tool calls introduce additional east-west flows. Operators increasingly ask for mechanisms to adjust resource splits and traffic handling without requiring major standard rewrites every time workload composition changes.

A third shift is observability and closed-loop control as part of performance itself. Under AI workloads, optimization is not only radio scheduling; it is end-to-end feedback: classifying request types, predicting bursts, dynamically placing inference at edge vs cloud, and degrading gracefully under congestion. Without observability and control interfaces, strong air interfaces can still lose to system-level bottlenecks.

This is why testbeds are being revalued. Standards discussions often rely on simulations and assumptions, but AI workload complexity makes many assumptions fragile. Shared testbeds and reference tools create a common language: same workload, same measurement method, comparable results. That reduces the cost of “everyone debating with their own numbers” and accelerates convergence on what actually works.

For the ecosystem, the implication is a redistribution of competitive advantage. The next winners may not only be those with the best base stations or devices, but those who can integrate edge-cloud orchestration, inference deployment, network automation, and operability into a coherent delivery system. 6G looks increasingly like systems engineering: the air interface is necessary, but differentiation shifts toward software, orchestration, and operational discipline.

To judge 6G progress, it will be more useful to watch three signals than to chase peak claims. First: what extensible architectural capabilities operators push in 3GPP studies. Second: whether reusable AI workload datasets and characterization methods emerge. Third: whether end-to-end engineering practices produce repeatable, production-relevant case studies.

The bottom line is straightforward: AI pushes mobile networks from “connectivity pipes” toward “systems platforms.” 6G competition will be less about a single KPI race and more about governing uncertain workloads. Architectural flexibility and reproducible testbeds are the bridge from concept to deliverable engineering.

Source: https://www.ngmn.org/highlight/ai-uncertainty-drives-mnos-call-for-6g-flexibility.html

Source: https://www.ngmn.org/publications/ai-surge-and-its-implications-for-6g.html

Source: https://www.5g-mag.com/post/27-02-2026-new-project-6g-testbed-ai-traffic-characterization

Source: https://www.telecomstechnews.com/news/why-ai-is-altering-planning-for-6g-mobile-networks/