Speechify, ElevenLabs and NaturalReader make the voice on their servers. Every word you listen to costs them computing time, so they charge monthly and cap how much you can listen to.
Natural voices no longer need a data centre. Open models such as Kokoro run on an ordinary graphics card or an Apple silicon chip, and when the voice runs on your computer, there's nothing to meter. That's why the pay-once and free options are, almost always, the ones that work offline.