When Success Becomes the Risk: Dependency Formation in AI-Native Network Evolution

Abstract

AI-native networks are being developed toward a vision in which communication, computing, sensing, and physical systems are jointly optimized within a single execution environment, extending from data centers to the edge, to on-device AI, and into industrial and physical AI. This paper does not question whether that vision will be achieved. It takes the field's stated goals at face value and asks what follows if they are. We argue that in such a system, technology dependency need not arise from integration failure, supply disruption, or adversarial action, and can instead arise from integration success. Three structural properties of the vision, together with one feature of how such systems are deployed, can generate this outcome. Deep cross-layer co-optimization means that a component superior in isolation may be inferior in integration and fail to enter the system without any exclusionary act. Extreme heterogeneity among the coupled subsystems means that fast-evolving layers can impose repeated revalidation demands on slower ones wherever their interfaces, state, or safety assumptions are coupled; and where staged migration is unavailable, the accumulation of such couplings can sharply reduce the availability of replacement opportunities. Scope expansion is intrinsic to the vision and aligns with the short-term interest of nearly every actor, but the party that decides on expansion and the party that subsequently bears the dependency are not the same. Finally, where adoption is clustered in time and one option is clearly the most mature, independently rational choices may coincide on it, carrying the result across adopters and sectors. Most of the building blocks on which the argument draws are established or observed in adjacent literatures. Their conjunction in this setting, and several of the consequences we derive from it, are not, and we grade each claim accordingly. An analogous concentration outcome has already been observed in an adjacent setting: sovereign AI programs, explicitly designed to reduce dependency, have been reported to remain heavily concentrated on a single accelerator ecosystem, with hosting localized but procurement not. The mechanism requires no actor to behave irrationally or opportunistically. We further argue that the resulting exposure is not an explicit or standardized object of measurement at the firm or the sector level, and that it aggregates above the level of the individual adopter. Meanwhile the indicators that are measured move in the direction of improvement throughout: interfaces are opened, vendor counts rise, conformance and interoperability tests pass. This is not a defect of those mechanisms. Interoperability testing asks whether a system still works, not how much more tightly it now works. What is required is therefore not a limit on integration but a distinct capability, which we term AI Network Evolution Assurance, whose object is the trajectory rather than the state: how upgrades and expansions accumulate dependency, and what substitutability will remain after a decade of such evolution. We suggest that its effect may be permissive rather than restrictive, and that it may lower rather than raise barriers to adopting globally supplied technology, since dependency concerns that cannot be grounded in evidence may manifest as delay. This is a conceptual synthesis and research agenda rather than an empirical test; we state the argument's limitations and the observations that would weaken it.