It's true that most floating point computations involve rounding, but there's no contradiction in that. If you think floats are inexact you can get into a lot of confusion about things; you might miss, for example, that `(a + b) / 2` is exact (except on overflow to infinity, or languages that don't enforce floating point correctness) or that Javascript's integers are actually just floats. Considering floats as inexact might lead you to believe they represent ranges of numbers, eg. x ± ½eps, which is confusing. I've seen more damage from this belief than most other myths about floats, at least among experienced programmers.
The whole point of floating point arithmetic is to be approximate, maintaining only a particular precision because otherwise the number of bits in the answer doubles every time we multiply. If someone cares about exact arithmetic, floats are the wrong tool for the job.
As such, to do pretty much anything with floats you have to keep some idea in mind of the relative accuracy of various values, and the condition number of every operation/algorithm. None of this is explicitly included in the number, you just have to track it manually.
Instead of thinking of a float coming out of some algorithm as an exact value, you should generally think of it as a representative of some implicit range of possible values, a range typically somewhat bigger than ±½eps.
Rather than thinking of a float as some range of values, you should think of it as a randomized sample inside some (probably unknown) range.
I think it’s misleading / counterproductive to talk about IEEE floats as subsets of “real numbers”, because e.g. Math.PI is a clearly different thing than the “real number” π (whatever that might be).
Calling floats “exact” is also IMO problematic, since in practice every computation using floats involves rounding.
(I also think it’s problematic for Gustafson to call the numbers in his system “real numbers”, even if you could in principle build new hardware with the idea to perform arithmetic on any particular (decided before hardware design) small set of real numbers you want.)