In the past, telecom networks were designed to be resilient, scalable, and largely invisible. When something broke, engineers fixed it. When traffic spiked, capacity was added. Intelligence sat mostly at the edge of the system, in operations centers and human decision-making. That’s always been the way of things. But just as it has in other industries, AI is changing the rules of the game. AI-native telcos are moving toward networks that are “born intelligent”, capable of sensing conditions, making decisions, and acting in real time without waiting for comparatively sluggish human hands. Routing paths are adjusted before congestion is noticed, capacity appears where demand is predicted rather than where it has already surged, and traffic paths shift seamlessly between clouds, partners, and regions in milliseconds. The network has grown a brain.
But where does network intelligence come from? Algorithms, yes, but even those are subject to the infrastructure beneath them. In 2025, as online activity surged toward a record-breaking182 zettabytes, with billions of devices and AI-enabled applications added into the mix, the traditional “best-effort” model of moving data has finally had its day. AI doesn’t tolerate guesswork. It needs predictable latency, transparent pathways, and an interconnection environment where you always know who (or what) you’re exchanging traffic with, where, and under what conditions. That level of clarity simply isn’t possible when traffic disappears into opaque transit routes, which also leads to other conversations about sovereignty, control, and exposure.
This is why AI-native telcos are now also pulling interconnection into the spotlight. They’re discovering that network autonomy, above all else, depends on the network’s foundations. They need direct cloud connectivity that keeps data off the open Internet when sovereignty matters; software-defined peering that provides real-time visibility, and distribution that brings Internet Exchange (IX) points closer to the edge where AI inference actually takes place. Modern deployments and tighter integration with cloud and satellite networks will be part of this solution, and as connectivity expands beyond fiber and mobile into LEO constellations and even lunar experiments, the future of communication will be shaped by one generation-defining question: can network intelligence flow as freely as innovation demands?
The Dependency Problem
The promise of autonomous networks rests on the basic assumption that AI systems can observe the network clearly enough to act with confidence. In practice, this is where many architectures begin to strain. Traditional IP networks were built to optimize for reach rather than certainty. Traffic is handed off across multiple autonomous systems, routed dynamically through paths that are efficient but often opaque, and delivered on a best-effort basis that prioritizes the ends rather than the means.
For human users browsing the web or streaming video, this trade-off has been quite tolerable. We barely notice it. But even for mission-critical applications that today run in the cloud, e.g. ERP systems, a best-effort approach is no longer tolerable, and solutions like Cloud Exchanges are now preferred. Looking into the future, AI-driven applications and use cases like autonomously-managed factories that operate in real time will need more assurance. Autonomous decision-making requires precise knowledge of latency, jitter, congestion, and regulatory jurisdictions. When those variables are hidden or shift unexpectedly, AI systems are forced to react after the fact rather than anticipate conditions ahead of time. So, no matter how intelligent the network, it can’t properly flex its capabilities if the underlying infrastructure won’t allow it.
This is the “dependency problem” – AI may be capable of optimizing routing and balancing loads, but only within the boundaries the network allows it to see and control. If the underlay doesn’t share reliable telemetry, deterministic paths, and consistent performance data, autonomy begins to degrade as soon as traffic crosses clouds, partners, or another geographic region. This is the issue AI-native telcos are now grappling with – network intelligence can’t float above the infrastructure as a separate, abstract layer and just “perform”. It needs to be anchored in connectivity that is visible, measurable, and designed to support decision made at machine-speed.
Agentic AI Has Changed the Network-Application Contract
The real driver for this push toward intelligent networks is agentic AI. Agentic AI isn’t limited to just observing and recommending the way ChatGPT or an analytics platform might – AI agents are designed to act independently, negotiating resources, triggering workflows, and adapting their behavior as environments change. In a network setting, this essentially means that they can request bandwidth, select paths, and enforce latency thresholds ahead of demand. In other words, connectivity stops being a static “on/off” service and becomes something that intelligent machines actively use as part of their decision-making loop.
And for that to work, network services must be discoverable, programmable, and dynamically adjustable, rather than locked behind manual provisioning processes or rigid, vendor-locked contracts. An AI agent cannot wait for a ticket to be raised or a configuration window to open. It needs immediate access to the network functions it depends on. This is why AI-native telcos are rethinking their architectures around APIs, automation, and real-time control. As applications are “born intelligent”, the network must evolve into an adaptive platform that intelligent systems can interact with directly and safely.
Sovereignty, Interconnection, and Control
As AI-native telcos move further down this path, the type of network interconnection will matter more and more. Sovereignty, performance, and trust all need to converge at the places where networks meet, which is driving a clear shift away from opaque transit toward intentional traffic flows, where connectivity choices are deliberate, observable, and governed within the software stack. This is why we’re seeing greater geographical dispersion of IXs – both broader footprints in larger hubs and smaller IXs emerging in edge locations, which allow data to move with purpose. Not to remove humans from the loop entirely, but to give intelligent networks the clarity and information they need to act effectively within defined boundaries.
As our need for connectivity grows, this question will loom larger and larger. Wherever AI systems operate, they will demand the same guarantees of visibility, control, and responsiveness. And wherever future AI data centers will be located, the users of AI – be they human or agentic – will be everywhere. Realizing this vision will depend on intelligent networks and AI-optimized IXs. The work we do now will define how possible that future is.
link