I have read the entire report.
Crazy. Just crazy.
What a terrible AI world in the background.


And how urgently POET's products can AND WILL provide a remedy here (as we know).
The conclusion of the report's end :

The second major change is computing density. As we’ve discussed in depth for many years, networking is a key piece to increase cluster sizes. This is a significant cost item in terms of CapEx, but even more importantly, a poorly designed network can significantly reduce the utilization rate of your expensive GPUs.

Generally, in a large cluster, we want to use as much copper as possible, both for the scale-up network and the scale-out network. Using copper electrical cables for communications avoids the use of fiber optic transceivers, which are costly, consume power and introduce latency. But copper’s reach is typically limited to just a couple meters at most when transmitting at very high speed – therefore, GPUs must be kept as close together as possible in order to run communication over copper.