Agree. And just to add further, micro compute... I.e. smaller...

Agree.
And just to add further, micro compute...
I.e. smaller and more specific workloads rather than all things to all people...

some of the other technologies being used to achieve compute on AI right now are so large and power hungry, it's not actually efficient to run small/repetitive workloads through them anymore.

i think BRN giving an example of security cameras is a good example here....

current state : farm all images off to the cloud and then come back if I detect a person in the image with AI. Imagine how many cloud servers are dedicated to doing this for a company like Arlo for all their cameras across the globe? Doesn't sound efficient to me if you can do it on the device itself.

future state - train the camera to know how to recognise a human and only send large workloads to the cloud where necessary (in my case, storage of the footage/images)

takes a while to get this if you don't have experience with this sort of infrastructure but it's worth getting educated if you have a view on investing in microchips imo.