The AI Traffic Surge Is Rewriting the Infrastructure Playbook

For decades, telecommunications network planning followed a relatively predictable rhythm. Carriers could model traffic growth with reasonable accuracy, plan their capital expenditures accordingly, and build to meet demand on a rolling multi-year cadence. That era may now be over. The explosive growth of artificial intelligence — from large language model inference to real-time machine learning workloads — is introducing a fundamentally different traffic profile, one that is reshaping where, how, and how fast telcos must build physical network infrastructure.

What’s emerging isn’t just an upgrade cycle. Industry analysts and network architects are increasingly describing it as a structural transformation — one that simultaneously demands more capacity at the ocean floor, in the stratosphere, and at every interconnection point in between. The pressure is being felt from hyperscale data center campuses all the way to the last-mile connection, and carriers are finding that traditional build timelines simply don’t align with the pace of AI adoption.

Submarine Cables: The Invisible Backbone Under Siege

Undersea fiber optic cables carry approximately 95% of international internet traffic, and that load is intensifying rapidly. AI model training and inference require massive cross-continental and transoceanic data transfers between hyperscale facilities — workloads that are uniquely bandwidth-hungry and latency-sensitive in ways that previous generations of video or cloud traffic were not.

In response, both traditional telcos and hyperscale players like Google, Meta, and Microsoft have dramatically accelerated submarine cable investment. New cable systems are being commissioned with spatial division multiplexing (SDM) technology, enabling individual cables to carry dozens of fiber pairs — dramatically increasing total throughput per system. Modern cables now routinely target capacities exceeding 20 terabits per second per fiber pair, a far cry from systems laid even five years ago.

The challenge isn’t just capacity — it’s time. Submarine cable projects typically require three to five years from planning to activation, encompassing marine surveys, international permitting, manufacturing, and careful deep-sea deployment. With AI infrastructure demand accelerating on a quarter-by-quarter basis, that timeline feels increasingly incompatible with market reality. Some carriers are exploring modular upgrades to existing cable landing stations and wet plant repeater upgrades to extract additional capacity from existing routes while new systems come online.

Low Earth Orbit: Bridging the Gaps AI Can’t Afford to Ignore

While submarine cables handle the transoceanic heavy lifting, low Earth orbit (LEO) satellite constellations are emerging as a critical — and surprisingly complementary — layer of the AI-era network. Constellations like SpaceX’s Starlink, Amazon’s Project Kuiper, and OneWeb (now Eutelsat OneWeb) are no longer simply rural broadband stopgaps. They are increasingly being evaluated as legitimate backhaul and redundancy solutions for enterprise AI workloads in underserved geographies.

LEO satellites orbit at altitudes between approximately 340 and 1,200 kilometers, delivering round-trip latencies in the 20–40 millisecond range — a dramatic improvement over legacy geostationary satellites that sit at 35,786 kilometers and impose latencies exceeding 600 milliseconds. For certain AI inference applications, edge computing deployments, and IoT data aggregation use cases, LEO connectivity is becoming genuinely viable in the network architecture conversation.

Telcos with satellite subsidiaries or partnership agreements are moving aggressively to integrate LEO capacity into their multi-layer network offerings. The integration challenge, however, remains significant — seamless handoffs between LEO, 5G terrestrial networks, and fiber backhaul require sophisticated software-defined networking (SDN) and network function virtualization (NFV) capabilities that many operators are still actively developing.

The Middle Mile and Metro Fiber: The Overlooked Chokepoint

The narrative often focuses on transoceanic cables and space-based networks, but industry veterans are quick to point out that middle-mile and metro fiber infrastructure represents an equally pressing bottleneck. As AI workloads concentrate in tier-one and tier-two data center markets, the regional fiber networks connecting those facilities to peering points, edge nodes, and enterprise campuses are experiencing unprecedented congestion.

Carriers are accelerating dark fiber deployments, wavelength service expansions using DWDM (Dense Wavelength Division Multiplexing) technology, and metro ring upgrades across major markets. Some are deploying coherent optical transceivers capable of 400G and 800G wavelengths to dramatically increase per-fiber capacity without requiring new conduit runs — a critical capability given that permitting and civil construction remain the longest lead-time items in any fiber build.

Capital Intensity Is Back — With a Vengeance

The financial dimension of this buildout cannot be understated. After years of investor pressure on telcos to moderate capital expenditure and prioritize free cash flow, the AI infrastructure imperative is forcing a recalibration. Several major carriers have signaled elevated capex guidance in recent earnings cycles, citing AI-related network demand as a primary driver.

The risk, of course, is timing. Carriers that over-build ahead of demand destruction or consolidation could face return on investment challenges. Those that under-build risk losing strategic positioning in an AI-driven economy where connectivity quality becomes a genuine competitive differentiator for enterprise customers.

Industry Outlook: Build or Be Left Behind

The emerging consensus among network strategists is stark: the telcos that succeed in the AI era will be those that move decisively on physical infrastructure now, even at the cost of near-term financial pain. The traffic cycle being driven by AI is unlike previous demand waves — it is more geographically concentrated, more bandwidth-intensive at the node level, and more latency-sensitive across the end-to-end path.

From the crushing depths of the Pacific Ocean to the orbital mechanics of LEO satellite constellations, telecommunications companies are being asked to build faster, smarter, and at greater scale than at any point in industry history. The technology exists. The demand is real. The question is whether the industry’s capital structures, regulatory environments, and engineering talent pipelines can keep pace with an AI economy that refuses to wait.