Escaping the Matrix: AI and Submarine Cables

We at the moment are transferring from sci-fi to actuality with the launch of many AI providers which are poised to rework industries and reshape our every day lives.

Whereas knowledge facilities are increasing quickly to help AI necessities, there’s much less readability on how AI would possibly impression long-haul community infrastructure, particularly submarine cables.

That is precisely what I delved into throughout my keynote at Submarine Networks World 2024.

On this presentation, I checked out three huge questions:

1. How huge are AI coaching knowledge units and fashions?

To know AI’s potential impression on submarine cables, we should think about the sheer dimension of coaching knowledge and fashions.

I used to be curious to grasp simply how huge the coaching knowledge and fashions are when it comes to gigabytes, as a substitute of the extra obscure metrics used within the AI world like parameters or tokens. In spite of everything, this knowledge is the uncooked materials that could possibly be carried across the globe on submarine cables. 

2. How will AI impression demand for submarine cable bandwidth?

Over the previous couple of years, I have been asking individuals how AI would possibly form the demand of submarine cable bandwidth. I introduced some outcomes of this non-scientific, pretend survey within the chart beneath.

Even with no definitive reply to this query, it’s value digging into what components might play a job.

Two phrases: location issues.

A mannequin might require pulling coaching knowledge from sources situated everywhere in the world, or it could possibly be relying purely on a knowledge set already saved domestically. Mannequin coaching and inference (utilizing an AI mannequin) might happen throughout the similar metropolis, or the mannequin could also be distributed globally throughout a number of knowledge facilities.

And that is simply the tip of the iceberg.

Components similar to knowledge dimension, mannequin coaching frequency, energy availability, knowledge sovereignty legal guidelines, inference latency, and on-device inference can even affect bandwidth necessities.

3. Is there an AI-fueled bubble in new submarine cables?

Many veterans of the telecom business will recall the submarine cable funding bubble (and subsequent meltdown) within the late Nineties/early 2000s. Is the present degree of funding much like what occurred 1 / 4 of a century in the past?

Whereas a whole lot of billions of {dollars} are flooding into GPUs, knowledge facilities, and energy technology, the funding into new cables is extra modest.

Shifting ahead

I closed with a take a look at some components to control because the AI revolution progresses. Particularly, the situation of future AI factories is vital. Will these be in areas which are unconnected or underserved by submarine cables?

Extra Information This Means 👇

You could find my full 2024 Submarine Networks World presentation over right here.

Obtain the slides. Free your thoughts.