From voice to video to AI-shaped traffic: Why network architecture must evolve at software speed
Connectivity has become essential to everyday life, and networks are the backbone that makes modern digital society work. Over time, that connectivity has meant voice, messaging, mobile internet, and streaming video. Now, we are entering a new phase where the key shift is not simply “more traffic,” but a change in the “shape” of traffic, increasingly driven by AI.
The new traffic reality: Interactive and bursty
AI-driven experiences are interactive and bursty, replacing stable sessions with event-driven spikes that can compress traditional downlink-to-uplink ratios as users shift from “consuming” to “creating and sensing.” Global averages may drift only modestly, but local variance is widening across markets and high-demand areas. The planning challenge isn’t a single new ratio; it’s that “busiest places,” “busiest times,” and the traffic mix are moving targets, driven by device adoption, mobility, and event-triggered bursts, so networks must be engineered for tail behavior and adaptability rather than a static busy-hour snapshot. This same dynamism spills into the control plane, where interactive AI increases signaling load and policy churn through frequent, context-driven service interactions.
“Physical AI” takes this further. Robots, autonomous vehicles, and industrial sensors perceive through the uplink and act on decisions in real time. In this world, the network becomes less like a content pipe and more like a nervous system. Safety-critical loops are moving from tens of milliseconds toward single-digit milliseconds, and in tightly bounded local deployments, toward sub-millisecond regimes. In that environment, a slip in reliability or latency is no longer a minor buffering issue; it can become a safety event.
This is the beginning of an AI supercycle: as AI and physical AI traffic evolve rapidly, the bar for adaptability, reliability and responsiveness rises. Meeting it requires coordinated intelligence end-to-end across RAN, core, access, transport, and the automation, agentic control, intent, and governance layers above, not siloed optimization. AI-native applications increasingly need intent-aware decisions, which in practice requires coordinated orchestration of RAN, core, access, and transport rather than domain-by-domain optimization.
While this is an end-to-end challenge, we start where these changes first collide with physics and determinism: the mobile network, and especially the RAN. Every generation of mobile networks has been shaped by a dominant traffic pattern. 2G and early 3G were built for voice and messaging. 4G turned the network into the delivery fabric for the application economy, and 5G took that further as video and streaming became the dominant consumer load. While video drove bandwidth demand sharply upward, its traffic profile was relatively predictable: long sessions and consistent flows that operators learned to engineer for. As we discussed earlier, AI and physical AI are changing that profile and as this traffic evolves rapidly, it is now putting architectural pressure across the stack.
The hidden constraint: Architectural pivots are generational
Mobile networks have changed architecture before, but they tend to do so only at generational inflection points. When data became dominant, we began the shift from circuit-switched foundations to packet-switched architectures and progressively moved toward all-IP. When broadband demand and spectrum expansion became decisive, we moved from CDMA to OFDM to make wideband operation and flexible spectrum use practical at scale. These were architectural pivots that redefined what the system was optimized for.
That pace was workable when the dominant traffic reality was stable for long periods. It is much harder when traffic becomes interactive, event-driven, and serves real-time intelligent control, especially when applications and interaction patterns are evolving rapidly. This creates a critical mismatch: a demand profile evolving at software speed colliding with an architectural model that moves at generational speed.
Mobile core evolution matters here as well. As the core becomes more cloud-native and API-exposed, it must support more dynamic policy, steering, and slicing behavior aligned with rapidly changing application needs. Just as importantly, it must be able to incorporate insights from RAN, transport, and access to make end-to-end decisions that are intent-aware and responsive, not static and domain-isolated.
A first-principles reset
We are at an inflection point. Across recent separate writings by industry luminaries, Martin Cooper, Bob Gallager, Andrew Viterbi, Arogyaswami Paulraj, and John Cioffi, a consistent warning is emerging: we are piling on complexity for progressively smaller gains and approaching practical limits where “optimizing harder inside the old box” stops being the right answer. The industry needs an end-to-end rethinking from first principles.
We write this as leaders and technologists who have lived on both sides of the boundary: telecom networks and large-scale AI systems. That experience makes one thing clear: when workloads become highly dynamic and compute-intensive, sustained innovation requires a programmable platform with the right acceleration, and an operating model that can evolve at software speed.
This is not about importing “cloud practices” wholesale into the Radio Access Network (RAN), but about building AI-native networks under carrier-grade constraints. The RAN has far stricter timing and safety constraints; it is about making continuous innovation feasible under those constraints.
Three core architectural shifts
-
Programmability and performance at Scale
To evolve behavior at software speed, the network requires inherent programmability. To handle the resulting complexity and deliver timing, energy, and throughput requirements at scale, it requires accelerated compute for workloads that are fundamentally parallel and math-heavy, especially as optimization becomes more cross-coupled and high-dimensional. This is no longer about relying only on fixed-function pipelines, but an architecture that can absorb continuous improvement while remaining carrier-grade.
-
Decoupling innovation from hardware
If innovation must move at software speed while hardware continues to evolve, software must absorb that change without constant, costly rewrites. This is the role of a Hardware Abstraction Layer (HAL). A well-designed HAL establishes stable execution semantics so that new hardware generations can be absorbed through focused kernel work and platform tuning rather than wholesale re-architecture of the software base. The server industry proved this model: virtualization provided stable execution semantics and portable abstractions that decoupled software from hardware refresh cycles and fast-tracked software-led innovation. The HAL preserves long-lived investment while allowing for continuous platform differentiation.
-
Intelligence in the decision loop
When traffic is bursty and interactive, adaptation cannot live only in long-cycle planning tools; it must participate directly in the loops where decisions such as resource allocation and scheduling are made. Once the optimization space becomes high-dimensional, dynamic, and cross-coupled, classical heuristics stop scaling. At this point, learning becomes increasingly necessary for parts of the decision space to manage these complex search and decision spaces at runtime.
This described shift has precedence in our industry. Computing systems decoupled software from hardware and accelerated the innovation cycle through programmability and abstraction. Networking went through similar shifts through SDN and programmable control. The RAN is reaching the point where a comparable architectural evolution becomes necessary, now with the added discipline of determinism, and intelligence in the decision loop.
Carrier-grade governance and openness
As networks become more adaptive, they must remain a “glass box,” not a black box. This glass-box requirement is central to how we think about AI-native networks at Nokia, and we will cover it in depth in a follow-up blog where we discuss intent-based operations, agentic automation with guardrails, lifecycle and policy management, and autonomous operations that are observable and auditable end-to-end. Operators need observability and lifecycle control to understand what was deployed, where, and with what outcome, and to roll back safely if needed.
This is where CI/CD (Continuous Integration/Continuous Deployment) becomes the essential engine for delivering innovation at software speed. It is not about reckless change; it is the mechanism by which new capabilities are governed by validation gates, canary deployments, and safety-case-bounded cadences. In the AI-native networks era, CI/CD extends beyond code to the model-and-data lifecycle, automating training and finetuning pipelines and enforcing validation, controlled rollout, and rollback of models and policies used across RAN and automation loops in production. Critically, this must be done end-to-end: testing, rollout, and rollback need to account for coupled behavior across RAN, transport, core, and automation domains, not just a single network function in isolation.
Finally, innovation at this scale requires open interfaces, within the RAN and the core, toward management and orchestration, and for “intelligence operations." This is why support for O-RAN and 3GPP is critical; openness is how interoperability and innovation move faster together.
Conclusion
When traffic changes in character, not just in volume, the network cannot remain a system optimized at design time and upgraded in long cycles. The next era demands that we decouple innovation from hardware, bring intelligence into the decision loops, and operationalize continuous improvement with carrier-grade guardrails. That is how we keep networks reliable and deterministic while making them adaptive enough for the AI-shaped future.
