Fiber optic networks provide the connectivity ubiquitous to modern society, enabling financial transactions, business critical traffic, national defense, entertainment, and personal interactions across the globe. Optical fiber extends to homes, businesses, and 5G radios in metro networks, connects cities within countries, and countries within continents. Subsea fiber cables spanning seas and oceans extend that connectivity between continents.
The massive data-carrying capacity of optical fiber has led to wide deployment of subsea fiber cables. Today, over 570 in-service subsea cables carry more than 99% of all international traffic and connect data centers around the globe to deliver cloud-based and emerging AI services.
A massive global data center build-out is currently underway to satisfy the compute needs of AI training. These build-outs are being led by hyperscale, neo-cloud, and AI companies, all of whom are seeking to develop market-leading AI models, agents, and applications that seek to transform work and society. These data centers will house ever-larger clusters of inter-connected GPUs, which serve as the engines for machine learning and AI training.
Modern AI training and large language models (LLMs) require a “scale-up” design, where thousands of GPUs function as a single logical unit, with the bandwidth needed to interconnect these GPU clusters soon exceeding 1 Pettabit per second; multiple orders or magnitude greater than the server-based internet traffic of traditional data centers.
While the interconnection bandwidths needed within a data center explode for both “scale up” and “scale across” requirements to increase GPU training clusters, what impact will this will have on inter-DC bandwidth, and across trans-oceanic subsea cables?
Some of these increasing bandwidth requirements will bleed into metro data center interconnections (DCI), as limits on maximum DC size and improvements in training protocols require, and enable, connectivity for AI training to be distributed across multiple DCs using “scale-across” architectures.
Beyond the metro, the roll-out of AI applications delivered from data centers to global end-users will drive even more global subsea capacity and connectivity. Nokia Bell Labs is forecasting that global AI wide area network (WAN) traffic will add one Zettabyte/month of capacity by 2033. In terms of annual bandwidth growth, this represents a continuation of the recent industry trend of approximately 30% per year growth in bandwidth transmitted over subsea cables.
More fundamentally, however, the continued growth in cloud-based services, augmented and accelerated by new AI services, is leading to fundamental changes in both the players involved and the network architectures optimized for global end-to-end subsea terrestrial networks. This article will describe these dynamics, and what changes are occurring in subsea networks.
Taking a step back in time to 1866, when the first practical trans-oceanic copper cable was put into operation connecting Ireland to Newfoundland, Canada, one sees that the cable landing station (CLS) performed a key role, with an operator detecting the trans-Atlantic telegraph signal, which was then handed off to another telegraph operator to re-transmit over terrestrial copper cables to New York City or London.
Fast-forward 150 years to the mid-2010’s, and trans-oceanic cables evolved from copper to fiber. The dominant operators of subsea networks were international communications service providers (CSPs) and bandwidth wholesalers, who led the deployment of new cables, or participated in cable consortium for cost-sharing of these expensive infrastructure projects.
While data rates had increased tremendously since 1866, from approximately 8 words per minute to 12Tb/s per fiber pair, little had changed in the CLS architecture. Subsea traffic continued to be terminated at the CLS with transponders, to then be connected via back-to-back client interfaces to another set of transponders, which re-transmitted the subsea traffic to its ultimate destination over a terrestrial backhaul network.
But today’s cloud and AI- based data traffic is different. Firstly, for the past decade or more, the majority of new subsea fiber cable build-outs have been led by hyperscalers, to support the large traffic demands between data centers spread across different continents. With that, traffic across subsea cable s does not originate or end at a CLS; it flows end-to-end between data centers that may be located meaningful distances away from a CLS. One impact of this for subsea networks is that DC operators are increasingly optimizing their subsea-terrestrial networks end-to-end, encompassing everything between the data centers at each end.
This leads to a change in the network architecture, where back-to-back transponders at the CLS are removed, and subsea traffic is transparently passed through in the optical domain to directly to the terrestrial network fiber, all the way to the data centers (see figure 1).
Optimizing subsea-terrestrial networking by connecting wavelengths end-to-end between data centers using optical pass-through at cable landing sites. Credit: Nokia
This optical pass-through of subsea wavelengths is enabled using the latest generation of multi-degree reconfigurable optical add/drop multiplexers (ROADMs) supporting subsea-optimized features, such as amplified spontaneous emission (ASE) spectrum power insertion into the subsea fiber, spectrum sharing, and optical channel monitoring, while also supporting the features needed for terrestrial networks.
Optical pass-through at the CLS sites eliminates two-thirds of the transponders needed for end-to-end connections, and reduces power and space usage at the CLS, where these are often at a premium. While the extra backhaul distance may marginally reduce the maximum data rate of each connection, this is partially offset by the performance of the latest generation of 140 GBaud coherent optics.
Thus, a CLS-to-CLS connection at 800Gb/s per wavelength across the Atlantic will be able to achieve an end-to-end speed of 700Gb/s when connecting to the data centers at each end of the subsea-terrestrial link. This end-to-end optimization provides important benefits to hyperscalers and AI companies that control both the subsea and terrestrial fiber routes that connecting their data centers.
The growth in cloud-based and AI services also implies global coverage and reach, balancing the fact that foundational AI models and training occurring in one region will be delivered via inferencing to other regions or continents. Overlay with that with national data sovereignty requirements and regulations related to network ownership, and oftentimes, a hyperscaler or AI service provider may need to partner with local network operators to provide terrestrial backhaul connectivity from a subsea cable to data centers in other countries.
This terrestrial backhaul connectivity from a CLS to a data center is often provided by a Managed Optical Fiber Network (MOFN) provided by a local CSP, which can sometimes come at high cost if extending subsea wavelengths across the terrestrial network requires as many fibers for the backhaul as exist in the subsea cable.
An alternative end-to-end optimization leverages the use of additional bands in the terrestrial backhaul fiber spectrum, such as C+L or Super C+L, to enable more wavelength division multiplexing (WDM) channels (see figure 2). While a subsea fiber typically provides 4.5THz of WDM spectrum, C+L supports 9.8THz and Super C+L supports 11.6THz, or 2.5x more than the subsea cable.
Combined with coherent wavelengths on the backhaul fiber operating at higher data rates than on the subsea fiber due to shorter distances, leveraging C+L or Super C+L WDM can reduce the number of backhaul fibers needed by 50-75% compared to the number of subsea fibers, providing significant overall savings for leased backhaul capacity.
Reducing the cost of subsea traffic backhaul by using more WDM spectrum in the terrestrial network using C+L. Credit: Nokia
Operators can complement the end-to-end network co-optimizations described above by using a common transport solution across their subsea-terrestrial networks. This can, in turn, deliver more than lower costs and the operational simplicity of common spares, training, and management; it also unlocks the benefits of end-to-end automation to plan, optimize, turn-up, and monitor the network.
Network automation can help reduce operational costs and enable faster turn-up of end-to-end subsea bandwidth, leveraging zero-touch commissioning tools to automate subsea channel optimization and service configuration, reducing the manual time and effort needed to turn-up traffic, while also ensuring optimal performance and resource usage.
The well-used expression “it’s a small world” takes on a new meaning in today, where AI applications will permeate the globe. Ensuring cost-efficient and ubiquitous global connectivity and providing equitable access to new AI-based applications will help bring people, cultures, and societies together across continents; more than ever enabled by the fabric of end-to-end subsea terrestrial networks.
