If one were to make a markov chain with the same amount of input data, would the result be the same? Markov chain chatbots have been a thing for years, just on a much more limited set of data.
For an example, if your training set consists of the words rain and thunder used interchangeably a lot, but the word "today" is only used once in the sentence "there is no rain today", then a Markov chain based on the data would never output "there is no thunder today", but a transformer might.
In other words, information compression (eg. equating rain with thunder) isn't just for practicability, it's a necessary requirement for (the current generation of) good language models.