The Impact of AI on Submarine Cable Networks

Submarine networks enable uninterrupted connectivity among geographically dispersed physical data centers, effectively forming a global data center without walls.

Substantial investments driving the swift growth of AI infrastructure in both existing and new data centers are set to significantly influence data center interconnect (DCI) networks, which are essential for the success of AI applications and use cases.

Research Director Alan Mauldin recently joined Ciena’s Brian Lavallée for a live webinar exploring these technical and strategic drivers transforming the AI landscape.

To kick things off, Alan explained the impact of training data size, dataset scale, and model complexity on AI capabilities. He also shared insight on how AI could shape the future of long-haul data demand and submarine cable networks.

Some key takeaways from the discussion include:

• AI Training Data: AI models require vast and growing datasets for training, with examples like Common Crawl (405 TB) and Meta's exabyte-scale data.  Factors influencing data size include model complexity, goals, and data quality. 
  
  
• AI Model Size: Model size depends on parameters, architecture, training data, and compression techniques. Higher parameter counts don’t always equate to larger models.
  
  
• AI Infrastructure: AI operations rely heavily on cloud-based systems, including training clusters, inference clusters, and end-user devices. Training often occurs in locations with high power availability, while inference prioritizes user proximity for performance.
  
  
• Demand for Submarine Cables: AI-driven factors shaping long-haul bandwidth demand include distributed training, federated learning, spatial-temporal load shifting, and legal issues like export controls and data sovereignty. 
  
  
• New Bandwidth Consumers: Hyperscalers and emerging AI-focused companies (e.g., CoreWeave) may drive demand for subsea bandwidth. Some carriers are offering GPU-as-a-Service (GPUaaS). 
  
  
• Infrastructure Expansion: AI-focused data centers ("AI factories") may require new submarine cables to connect underserved or unconnected locations.

If you missed the live event, we've got you covered.

👉 Grab a copy of Alan’s slides (and queue up the on-demand webinar recording) to get caught up.

A few of Alan's presentation slides: How does AI work? How big are AI training data sets? How big are AI models? Existing and planned cloud data centers. Used international bandwidth by network type.