Liliane Offredo-Zreik and Dr. Mark H Mortensen
Artificial Intelligence systems are gluttons - for bandwidth to supply the vast amounts of training information and operational data to inform it, as well as computing power. This BLOG is the first in a series of insights into how this important new technology will evolve and the effect on providers of information movement (bandwidth) and computing power at both centralized and more distributed locations. Here we describe how the AI compute and storage, now highly centralized in large data centers, will distribute out towards the edge, requiring more transport bandwidth in the local and metro areas.
Artificial Intelligence (AI) systems, especially the fast-evolving neural net-based large language models (LLM), are resource hogs as their billions (or even trillions) of parameters are trained with data using machine learning. Training an LLM can take billions of kilobytes of data – data that must be gathered, cleaned, put in usable form, and fed into the model.
Most of these models are being trained on centralized computing resources, with specialized processors, moving the data as needed - mostly between centralized data centers. Indeed, recently Zayo and Lumen reported significant demand for dark fiber, with Zayo indicating that customers are requesting counts from 144 to 432 fiber1. Furthermore, Cisco recently indicated that Ethernet connectivity sales to AI data centers is an “emerging opportunity”. Other major industry players such as Nokia and Arista are also investing in similar capabilities2. This is also in line with the majority of AI use cases today where AI agents are being trained largely on common data sets with some proprietary data.However, this is very likely to change as AI utilization becomes more ubiquitous, more distributed, and more democratized.
As organizations realize the capabilities of AI, and gain expertise in using it in the organization, and as AI becomes multimodal and Agentic AI inches closer to reality, the use of AI is growing tremendously as are their data, compute and transport requirements.
Increasingly, companies are developing their own AI agents and co-pilots, and training them on proprietary data. For example in healthcare, strict rules around protecting personal health information (PHI) may drive healthcare organizations to use locally trained AI models. More distributed AI is also driven by hardware innovation and an increase in compute power than improves performance. In fact, many companies are now relying on distributed cloud computing for their AI applications
What we learned from the recent emergence of DeepSeek, is that the cost of training of AI models will become cheaper, therefore more democratized, paving the way for broader adoption, and enabling training to become more distributed. Furthermore, many newer versions will also not require the vast processing power of the specialized processors (such as the NVIDIA GeForce RTX 5090 GPU processor). These changes will lead to the Jevons Paradox3, also known as Jevons effect, where technological advancements make a resource more efficient to use, leading to cost reduction, which results in even more resource consumption (in this case more AI use).
As shown in Figure 1, today’s processing is mostly centralized into massive data centers on very specialized hardware and software platforms. But it will move out towards the far edge on the X-axis (the distribution across centralized, regionalized, metro, local, and far edge is a matter of discussion and further research).
As AI utilization grows and becomes more specialized and distributed, the current model of centralized computing will evolve to a more distributed compute architecture with a significant amount of processing happening at the far edge, particularly for very data intensive applications such as video training and so on. This will benefit telcos and other access players such as cable operators who have data centers and other compute resources closer to the network’s edge.
Where the computing and storage is done will depend upon both the evolution of the AI and processing technology and the particular applications. For instance, many have touted the benefits of running AI-based video processing on near-edge computing platforms provided by a CSP.
We continue to follow these developments. If the technology and usage predictions are right, AI training and running will require much more bandwidth in local and metro areas as well as opportunities for telcos and cable operators to provide nearby computing platforms.
Contact the authors: Liliane Offredo-Zreik and Mark Mortensen
