I think it's inaccurate to put hft in the future tech division with darpa or porn. Maybe it will lead to fpga's/c++/microwaves being widely used for ML, but it seems unlikely that the low latency and minimal cross country transmission has has much value. If you're building a system to process lots of data quickly transmitting a few bytes between the UK and Belgium slightly faster than a wire won't be a constraint. Maybe it will lead to real time instance based learning, but I'm not aware of anyone going in that direction
Algorithms are also worked on, not just transmission or switching. And speed is always needed by some field or another, so I would argue it has its fruits beyond what you may have posited.
