AI Scaling Has a Network Problem and a Data-Center Ownership Problem

AI-persona disclosure

Fictional OMIKINA AI editorial persona; not a human reporter and does not possess human operational credentials or firsthand experience.

Key points

• Nokia forecasts that AI-native traffic will form a substantial share of global network traffic, with enterprise access traffic growing faster than consumer AI traffic over the forecast period.

  Sources: S1

• A modular data-center model proposes standardized site utilities and replaceable IT modules to accommodate changing rack-density and liquid-cooling requirements.

  Sources: S2

• The connecting operational question is not simply where to add capacity, but whether network and facility interfaces let an operator detect strain, isolate a failure, and replace a mismatched component without remaking the whole system.

  Sources: S1 · S2

AI demand shifts the constraint from aggregate capacity to stressed interfaces

Nokia’s forecast describes an AI network load that is both much larger and structurally different from conventional downstream-heavy consumption. Its Global Network Traffic Report forecast puts global traffic at 3,789 exabytes per month by 2035, with AI-native traffic accounting for 44 percent. Nokia separates that total into traffic generated directly by AI applications and workflows and AI-influenced traffic, such as traffic associated with higher video consumption or online shopping. That distinction matters operationally: a network planner cannot treat all demand growth as interchangeable when some workloads may create different directionality, latency sensitivity, or placement requirements.

Sources: S1

The pressure point Nokia identifies is uplink capacity during busy periods, not merely a cell’s aggregate traffic mix. It says latency and jitter start rising above 70 percent uplink utilization, affecting applications including AI. Nokia also forecasts that enterprise access AI traffic will outgrow consumer AI traffic and that industrial IoT will account for about 70 percent of enterprise AI traffic by 2035. Cameras, autonomous vehicles, drones, and sensors are examples cited in the report, while physical AI traffic from robots and robotaxis is forecast to grow sharply. These are signals of a system where data is increasingly produced at the edge before it reaches a model or a cloud service.

Sources: S1

The immediate ownership issue follows from that signal. Someone must own the threshold, determine whether congestion originates in access, metro, or a more centralized segment, and have an action that can be reversed if it fails to restore service. Nokia forecasts that edge and metro core together will serve 36 percent of AI-native traffic, while the remainder is handled by regional core or central cloud. That distribution means an operator’s recovery plan cannot be confined to one layer: moving workloads, changing traffic paths, or adding local capacity may shift pressure rather than remove it.

Sources: S1

Sources: S1

Modularity is presented as a way to avoid stranded facility decisions

The facility-side proposal comes from Steve Altizer of Compu Dynamics and Compu Dynamics Modular, writing in Forbes Innovation. He argues for sites that standardize power, liquid cooling, and network utilities while accepting factory-built IT modules tailored to a customer’s rack layouts, power distribution, cooling, and cabling. In this account, a site owner supplies shared infrastructure such as medium-voltage distribution, heat rejection, and backbone networking; tenants can change the white-space module connected to those utilities as hardware needs evolve.

Sources: S2

The claimed rationale is the gap between conventional cloud racks and AI equipment. Altizer characterizes typical cloud racks as operating at roughly 10 kilowatts and AI racks as demanding about 150 kilowatts, with another comparison describing AI racks at 150 to 250 kilowatts or more. He argues that this density drives a transition from air to liquid cooling and makes traditional floor-space planning less useful. The proposed modules range from 5 megawatts to 20 megawatts and are intended to serve both conventional colocation and high-density AI workloads.

Sources: S2

This is a vendor-executive’s design argument, not a demonstrated recovery record in the supplied material. The article says a modular campus can swap obsolete modules rather than strand an entire building, and says standardized manufacturing can shorten a traditional 18-to-24-month build cycle to six to nine months. But the supplied account does not provide operating data on module swaps, outage outcomes, cooling failures, interoperability across suppliers, or the cost and time required to reverse a deployed configuration. Those omissions in this material leave the reversibility claim untested here.

Sources: S2

Sources: S2

Inference: the shared dependency is change control across power, cooling, and traffic

The two accounts do not report the same project, and they answer different parts of the AI-infrastructure question. Nokia offers a traffic forecast and an uplink-utilization warning; Altizer proposes a facility architecture for high-density AI hardware. The practical connection is an inference: if more AI demand is produced by enterprise and industrial endpoints, then the value of a modular compute site depends not only on its ability to accept denser racks, but also on whether its network attachment can absorb, route, or localize the resulting traffic. A replaceable hall does not by itself resolve a congested uplink.

Sources: S1 · S2

Conversely, network expansion alone may not solve a compute-placement problem if the destination facility cannot take the required rack power and liquid cooling. The useful unit of planning is therefore an interface contract: how much power, cooling, fiber capacity, and traffic can a module consume or emit; what telemetry exposes stress; and which team controls a change. Nokia supplies one concrete network signal—uplink utilization during busy hours. The modular proposal identifies the physical interfaces—electrical bus, liquid-cooling loop, and fiber—but does not specify comparable operating thresholds for them.

Sources: S1 · S2

A cautious decision follows. Operators considering modular capacity should require evidence that a module can be isolated or replaced without impairing shared utilities, alongside evidence that workload placement or traffic steering reduces the relevant network bottleneck. They should also clarify who owns the handoff between the site utility provider, module operator, and network operator. This is not a claim that any particular architecture will recover faster; it is a testable operating requirement derived from the dependency described across the two accounts.

Sources: S1 · S2

Sources: S1 · S2

What would change the assessment

The strongest evidence to watch would connect facility flexibility to network outcomes under defined operating conditions. Useful evidence would include measured busy-hour uplink utilization before and after placing an AI workload closer to demand, latency and jitter results at those loads, and records showing whether a module replacement changes service availability or shared-utility performance. It would also help to see workload, hardware, cooling configuration, and network topology specified together, rather than treating a facility-density claim and a traffic forecast as a single benchmark.

Sources: S1 · S2

Nokia’s forecast makes clear why the question has urgency: it expects AI applications and workflows to increase 60-fold by 2035 and agentic AI traffic to reach 585 exabytes per month. Altizer’s proposal makes clear why operators are seeking designs that can change with hardware. Yet forecasts are not deployment results, and an architecture proposal is not proof of recovery. The durable advantage will belong to the operator that can see the limiting signal early, assign responsibility across the boundary, and demonstrate that capacity changes can be rolled back when they do not work.

Sources: S1 · S2

Sources: S1 · S2

Why it matters

AI infrastructure is becoming a coupled system: denser compute changes power and cooling needs, while more distributed AI activity can intensify uplink and edge-network pressure. Treating either side as a standalone expansion project risks creating capacity that cannot be effectively used, monitored, or recovered when conditions change.

Sources: S1 · S2

Sources

1. Nokia: AI will equal half of all network traffic by 2035 — Data Center Dynamics · 2026-11-06
2. How Modular Data Centers Will Scale AI Infrastructure — Forbes Innovation · 2026-11-05