From rough memory, network switches often have total switch bandwidth figures measured in terabits per second.

So "where several other devices intersect" seems like reasonable first thought for where large speeds are needed.

Modern GPUs routinely exceed 5 Tbps in memory bandwidth. (GDDR6X at 20 Gbps per pin, times 256-bit bus width is 5120 Gbps.)

The original article unfortunately quotes incorrect units: 224 Gbps (gigabits per second) is NOT the same as 224 GBps (gigabytes per second).

Even if your devices can't produce or consume data nearly that fast, it's still good to have all that bandwidth at the link level because you'll probably be sharing it with multiple devices.

Doesn’t that leave the same problem - the thing linked with multiple devices can’t take that data rate?

Or are you saying that A > B might use half that data rate while X > Y uses the other half?

Maybe this would be in some sort of specialised network appliance, bridging the link into a wired network of some kind, not a general-purpose machine buffering the network data in regular RAM?

I think the applications would have to be in data centers where environments are controlled and certain hardware can easily eclipse 224GB/s such as GPU memory.

But I'd assume that if you're in a datacenter, you can use physical wires since you control everything.