Here's my uneducated take..
First off, AKIDA is now freely available, so why are they comparing their latest iteration, to a GPU?
48 million neurons, well and truly trumps AKIDAs 1.2 million, but it is the number of connections between neurons, synapses, that makes something "smart".
AKIDA claims 10 billion (but I understand the ratio of neurons and synapses are on a sliding scale?).
They are using 16 bit (vs AKIDA 1 to 4) activations, to remember the previous step (something we won't achieve, until LSTM?).
So that's the only thing I can see, that they are doing better, but they would be using a lot more power, than when we have next gen.
As others have said.
No learning (how can something even be considered "intelligent", if it can't learn?)
Not available as IP = huge disadvantage in the OEM marketplace.
It's competition, but not good competition, in my opinion.