01 Trigger Event

Nokia released its Global Network Traffic Report on November 4, forecasting that by 2035, AI will drive nearly 44% of global network traffic, while total monthly traffic will nearly quadruple to approximately 3,789 exabytes.

The report also makes two more aggressive projections: agentic AI traffic will reach 160 times its 2026 level by 2035, while physical AI traffic will reach 105 times its 2026 level.

Fixed wireless access (FWA), meanwhile, could account for more than one-third of consumer cellular network traffic.

This is not simply another network equipment vendor's routine outlook for bandwidth demand. Nokia is effectively redefining AI from an “application workload” into the primary organizing principle of future network traffic.

02 What It Really Means

The issue is not whether AI will increase traffic, but what will generate the additional traffic.

The first category is multimodal generation. Text, image, audio, and video models consume input tokens and output tokens, but traditional internet traffic measurement is closer to bytes transferred. The two cannot be equated directly. Data movement behind model calls, cache hits, cross-region synchronization, and result delivery ultimately reach the network, even if they do not appear as user-visible streaming.

The second category is agentic AI. An agent may continuously plan, retrieve, execute, and verify tasks. A single user action does not produce a single request; it produces a sequence of requests carrying context, tool calls, and state synchronization. What truly expands is not the size of an individual response, but interaction depth.

The third category is physical AI. Robots, vehicle systems, and industrial equipment require continuous transmission of sensor data, decision states, and telemetry. Physical AI traffic may not be as concentrated as agent traffic, but it could be more continuous and harder to compress.

Nokia uses 2026 as its baseline and forecasts agentic AI traffic at 160 times that level by 2035. This figure is first and foremost an infrastructure investment narrative: if software continues shifting from “answering once” to “executing a workflow,” network workloads will move from request scale toward task duration.

AI network traffic's core variable may not be how many prompts users send, but how many steps agents run in the background.

This is the part of Nokia's forecast that AI builders should pay attention to.

Nevertheless, 160 times is a long-range projection, not an observed market size. The report's summary also does not disclose how AI traffic is converted into network bytes or the precise methodology used for agentic AI traffic. For model API consumers, it is better used as a stress test for capacity planning than as a revenue forecast.

03 Historical Analogies / Structural Comparison

When the iPhone launched in 2007, no one accurately predicted the scale of the industries that would emerge around mobile app stores, mobile internet advertising, and mobile productivity. The key was not the iPhone itself, but the way it combined network connectivity, computing power, and distribution platforms into a new infrastructure layer.

After AWS matured in 2014, IT spending did not disappear; it shifted from on-premises procurement toward on-demand cloud resources. What was fundamentally restructured was budget allocation: fixed capex gave way to ongoing usage.

ChatGPT in 2022 demonstrated a different form of aggregation: foundation model capabilities became concentrated among a small number of providers and were then distributed through APIs. The application layer began building products around standardized interfaces, but the cost structure did not disappear; it migrated from model training to continuous inference.

Nokia's forecast resembles a combination of these three historical shifts: AI is becoming a new source of traffic, agentic AI is increasing the compute density of each task, and FWA and cellular networks are competing for the last-mile distribution layer.

Historical infrastructure inflection points often appear first as “insufficient bandwidth,” followed later by new pricing models and service layers. Before the 2008 financial crisis, operators also saw gaps between traffic forecasts and actual demand. This cycle is different, however, because AI workloads may bring greater compute density rather than merely more video bytes.

I would compare it more closely to the 2014 cloud transition than to the 2007 iPhone launch. The former reshaped cost structures and capacity planning, while the latter was more heavily oriented toward user access and the hardware experience.

04 What It Means for AI Builders

For model API consumers, the immediate priority is to decompose actual costs rather than chase a long-range traffic chart.

First, distinguish traffic, tokens, and GPU time. Downloading an image generates network traffic but may not consume many tokens; a long-context agent may trigger substantial prefill without continuously transferring large volumes of media data. The gains from routing and caching also cannot be estimated solely by response size.

Second, agent products should be modeled around complete tasks. The evaluation target should not be the price of an individual model call, but the total tokens, tool calls, retries, context reuse rate, and human intervention rate required to complete a task. A 160-fold traffic increase assumes that agent execution chains genuinely become longer while unit-task economics remain viable.

For AI startups, network infrastructure costs are not limited to bandwidth. Cross-region deployment, response streaming, real-time voice, vision, and background agents collectively affect latency and bills. Model routing can optimize average costs, but complex fallback chains may also increase cache invalidation and duplicate computation.

There are three concrete actions to take this week:

  • Log model calls, input and output tokens, tool calls, and end-to-end latency for every successful task across major workflows.
  • Incorporate prompt caching, batch APIs, and model-specific routing into the unit-task cost model rather than budgeting from a single input token price.
  • Establish task-level budgets and fallback mechanisms for long-chain agents so that MoE routing or tool loops do not conceal uncontrolled costs in the background.

For model access platforms such as opcx.ai, the opportunity is not merely to “sell more tokens.” The greater value lies in converting call volume into task volume: providing cost observability, provider switching, cache-hit analysis, budget control, and cross-model routing. Otherwise, customers are merely seeing a cheaper token interface rather than a lower cost of task completion.

Nokia also forecasts that FWA will generate more than one-third of consumer cellular network traffic. This is a reminder that infrastructure strategy cannot focus only on GPUs: the final form of AI applications will feed back into last-mile access, edge inference, and data residency. Those who control distribution are more likely to determine where computation occurs.

05 Counterargument / Risks

I may be overestimating how indicative this forecast is for AI infrastructure decisions.

First, 3,789 exabytes and 44% are both 2035 projections covering a horizon of nearly a decade. A change in the technical path could produce an entirely different traffic structure: small on-device models could reduce cloud requests; more efficient compression and sparse MoE could lower the data volume per task; and agents could converge from multi-step calls toward shorter, more stable execution chains.

Second, a 160-fold increase in agentic AI traffic does not imply a 160-fold increase in agent revenue or model provider profit. Tasks may be absorbed internally by platforms, or model efficiency may allow them to be completed with less inference. If most additional traffic comes from low-value telemetry or repetitive data synchronization, its infrastructure value will fall far short of the narrative.

More importantly, AI traffic is not necessarily more “expensive” than conventional traffic. GPU compute, KV cache residency, and database state updates may be scarcer resources than backbone network transmission. Nokia's perspective naturally leans toward network equipment: the company may see bytes while underestimating the constraints imposed by the compute and memory hierarchy.

I have not operated Nokia's internal model or reviewed its complete forecasting assumptions, so I cannot confirm whether its definitions of agentic AI and physical AI map directly to API call volume. I may be wrong on that point.

The most important elements of Nokia's thesis to verify are therefore not the headline figure that traffic will grow 160 times by 2035, but three trackable indicators: whether AI's share of traffic continues to rise, whether agentic workloads move from demos into production, and whether the gross margin generated by each AI task can cover the increased inference and network costs.

Only if the third condition holds will the traffic inflection point become an industry inflection point. Otherwise, it is likely to be little more than the story cloud vendors are most eager to tell.