He says in the video the high voltage driver is incorporated...

He says in the video the high voltage driver is incorporated into the chip itself (as a tiny ASIC chip which is a technical marvel in its own right) into one package. The other circuitry (FPGA driver) drives the chip by providing on/off commands to the individual pixels.

Efficient optimised code has to be created by software developers to run on a mobile processor - it is a vast amount of data that needs to be generated in real-time to produce sound given a digital input, to convert it using AKP's algorithms into the commands for the thousands of individual pixels. Presumably they have decided it can be done, and there is little reason to doubt this - previous demos have shown the pre-computation taking (say) 3 times longer than the time it takes to play the sound itself (i.e. to play 10 seconds of sound, it takes 30 seconds to generate the pixel command data) and for all we know, this may have actually been the time required to transfer the data over a network to the on-board memory on the doughnut, not the pre-computation. In any case, assuming the worst, only a 3x improvement is required. In the world of software where changing to efficient algorithms and then optimising an efficient implementation can easily result in 100x or more speedups, this does not seem difficult. Understandably, this is not something you would do until the very end - because it will result in less flexibility. Given the recent software developments in reducing acoustic noise, it just shows that this task was rightly deferred until late in development.