It’s both the data uploaders, and the network. Increasingly more the data uploaders, until the network pays nothing. That is the mechanism to get the unminted money out into circulation.
The exact rate at which data uploaders take the real cost of storage, and network ceases to mint new coins (to the benefit of farmers), is an implementation detail.
I have implemented a solution (presumably) that gives a decline up until ~130 billion nodes, while seeing fairly OK costs also at the very first moments of the network. That is equivalent to a world wide adoption with some IoT. (There are caveats, but the details of these might very well be outdated and made irrelevant in any of the upcoming testnets.)
It is possible to imagine farming systems where this is not decided by the network (i.e. a handful of developers at design time…) - such as PAC - but that is beyond the scope of this task at this point of time.