Free shipping on orders over $99 Track OrderHelp Center

Nokia Forecasts AI Will Drive 44% of Network Traffic by 2035

A new traffic model points to greater uplink demand and distributed AI processing, with important limits on what its figures mean for optical infrastructure.

Conceptual fiber connections between a central AI data center and metro or edge nodes

A traffic forecast with a wider infrastructure message

Nokia released its Global Network Traffic Report on November 4, projecting that monthly global traffic within its defined scope will approach 3,789 exabytes in 2035, nearly four times the 2026 level. AI is expected to drive 44 percent of that total. The official announcement presents a forecast based on application use, connected devices and AI workflows, rather than a measurement of traffic already flowing through operators’ networks.

For optical communications, the significant point is the changing distribution of demand. More upload activity and processing closer to users would alter where capacity needs to be installed. A single global growth figure cannot specify which metro route, access aggregation link or data center connection will fill first. Goilips’ interpretation is that the report offers useful scenarios for infrastructure planning, while local measurements remain essential to investment decisions.

The publication also separates different kinds of AI activity. Autonomous software agents complete workflows, while physical AI encompasses machines interacting with their surroundings. Nokia expects both categories to expand rapidly from small starting points. These applications should therefore be evaluated by their traffic behavior and service requirements, rather than grouped into one uniform demand profile.

What the headline includes, and what it leaves outside

The original report’s definitions distinguish AI-native traffic from AI-influenced traffic. The former includes traffic generated or managed through AI applications and cloud workflows; the latter covers additional access traffic caused by AI recommendations changing consumer behavior. The 44 percent headline combines these categories. Describing the entire share as AI-native would erase that distinction and overstate what the figure says about direct AI processing.

The forecast also has explicit boundaries. Internal GPU communications, transfers of training datasets between data centers, and several forms of replication or backup sit outside its scope. Consequently, the total is not a complete inventory of every byte crossing an AI computing fabric. It should not be used as a direct forecast of scale-up or scale-out optical bandwidth, nor as a way to estimate transceiver shipments without a separate deployment model.

Goilips’ planning assessment is to retain these definitions when comparing the study with an operator’s own counters. Teams should establish whether a measured volume represents access exchanges, a cloud workflow or internal fabric communication. Combining values measured at different boundaries can produce misleading totals. Keeping the traffic boundary visible makes an otherwise broad market forecast more useful to network engineers.

Uplink demand and the geography of AI processing

Nokia’s report overview projects that edge and metro core infrastructure will serve 36 percent of AI-native traffic by 2035. It also expects more than 90 percent of physical AI traffic to originate in the uplink. These percentages describe different populations: the first concerns where AI-native processing is served, while the second concerns the direction of physical AI traffic. They cannot be substituted for each other.

The engineering implication, in Goilips’ view, is to examine a service path in both directions. A workload can involve a sustained upload and a small response, or bursts separated by quiet periods. Planning only around average download consumption could miss the busy-hour conditions that matter to an interactive application. Measurements should therefore include directional load, delay variation and the time available to recover from a failure.

Moving processing toward metro and edge sites also makes connection geography important. A planner can compare central processing with a distributed scenario, then identify the aggregation points and fiber routes affected by each. That exercise does not assume every new workload needs a separate optical wavelength. It asks which existing links have margin, which need upgrades and which require a different routing or resilience design.

The optical decision comes after that workload and route analysis. Required reach, host support, available fiber, optical loss and an agreed latency budget should guide the choice of terminal equipment. A global AI traffic share does not establish that a specific 800G or 1.6T module is required on a particular link. Those choices need evidence at the service and system level.

Turning a scenario into a measured optical network plan

Independent coverage by SDxCentral on November 5 corroborates the report’s central forecast and its emphasis on uplink pressure. DCD subsequently carried the same author’s story, explicitly identifying the original publication. The two media pages therefore provide one editorial reporting chain, alongside Nokia’s primary material. This article relies on the primary definitions when interpreting the headline figures.

For an operator assessing the forecast, a practical next step is to build several demand scenarios against the same network inventory. Each can record expected workloads, busy-hour traffic, endpoint locations and protection requirements. Comparing those scenarios with measured link utilization can identify capacity that should be added now and capacity that can wait for clearer customer demand. This is Goilips’ suggested method, not a deployment announced by Nokia.

Optical procurement should follow those findings with a system qualification process. Relevant checks include host interoperability, optical margin, operational monitoring and service performance during recovery. Qualification records should state the tested configuration, because a result from one route or host does not establish suitability for every installation. The report is a reason to revisit planning assumptions, rather than a substitute for those checks.

The remaining uncertainty is how quickly the modeled workloads will appear in individual markets and how operators will carry them. The public forecast does not announce an optical module order, a commissioned network or a guaranteed equipment revenue stream. Its value is a clearly framed view of possible demand. Forecast figures above are attributed to Nokia; infrastructure implications are Goilips’ analysis. The AI-generated cover is a conceptual fiber network illustration.

‹ Back to News & Blog