They still cascade normally like CSS classes... see this example:
<div style="color: red">
<p>red text</p>
<p>red text</p>
<p style="color: blue">blue text</p>
</div>
Ironically, it's actually complex CSS classes and long selectors that produce high specificity, which then requires even more specificity to override and eventually the dreaded !importantImagine styling a table row. That is one style attribute. Now imagine styling a 100 tables rows. That is style attribute duplicated a 100 times. Think of the amount of unnecessary bandwidth consumed to load that HTML file with 100 duplicate style attributes. Not only does it slow down page load times, it also slows down page render times.
Now if you add selectors to style attributes then it solves the above issue altogether. But, then you are just re-inventing the wheel. Instead of CSS in a separate file you now have the same functionality but all inlined in style attributes.
So in this situation you would have a <my-row /> component, which would contain the tailwind classes. This reduces load time as duplication is eliminated, but I'm not sure it would reduce the render time.
Page load times are interesting, but again, unless you're loading an insane amount of data from the server, it will never be a real issue.
Some sort of component architecture solves the problem you bring up w/ duplication. You should never be writing the same thing 100x over. JS frameworks all render 100's of the same component with the same attributes and performance is fine for 99% of real use cases.
The downsides for Tailwind all revolve around personal preference and how your team organizes a project. There is no performance hit (unless you ship the entire un-purged CSS file to production, even then, it's minimal).
What? Do you ever leave the city? It seems like that’s an unsupported use case any many developers minds.