Two things you mentioned that are incorrect though
The first is the presumption that you need a GPU to do machine learning - the reality is that the average business doesn't need nor has the data to benefit from such large models
The second is that PHP is slow - but when you do comparisons with Python, PHP is actually faster
https://benchmarksgame-team.pages.debian.net/benchmarksgame/...
The reason why Python, for example, is faster in some machine learning applications is because it's probably getting help from a library such as numPY or Tensorflow
Although these tools are missing currently in the PHP ecosystem, I'm confident that we'll have them eventually
For example, here is an attempt to bring numPY-like functionality as a PHP extension
https://github.com/phpsci/phpsci-ext/tree/1.0-dev
We've also had some progress being made accessing Tensorflow from PHP using an FFI (foreign function interface)
https://github.com/dstogov/php-tensorflow
So this is where the convo should begin in my opinion