The same reason Facebook loads multiple MBs of JavaScript code for something that could have been not more than 1 MB. One can't help but wonder about nefarious purposes: either collecting user data or abusing the computing resources. For companies that tend to hire the best and the brightest, the 'software bloat' theory is not compelling.
Curiously, major Linux distributions have also gotten significantly slower compared to early 2000s versions.
I wonder, for a thought experiment, what if companies stopped development on software when it reaches certain stage of maturity, say Windows 2000, providing only necessary security updates or optional visual changes?
I was very reluctant switching from win xp to win 7. Unless the 64bit era would have forced me I rather would have stayed put.
New software is really not adding much to the table after some point of completion since the software companies seems to mostly add pet feutures and user hostile fads be it star menu, complete gui changes, 'enterprise' admin lookout or ads and tracking.
E.g. Facebooks 1000s of developers seem to add a net of antifeutures to their site. On Netflix you can't even disable autoplay.
It is the same really for software moving to remote mainframes. The companies rather hide and burry the old desktop versions deep.
The same trend can be seen with linux distributions. A lot is surely down to composited window management. There's also things that we didn't run as services back in Windows 98 days - network capabilities, bluetooth, low power states, file hashing for quick search, assistants and other ways of rapidly finding and launching programs, auto updates, telemetry, automatic mounting of connected devices.
EDIT: and in the case of Windows 10, adverts/trialware and their associated animations.
That's an interesting point, but there was certainly an increase in CPU usage with the switch from gnome 2 to unity and gnome 3, even though they were GPU accelerated. I guess CPU is still needed.
It doesn't require an SSD in the strictest sense of the word, but it is developed with the strong assumption that you have at least something as fast as an SSD, so the experience of running it on a spinning rust disk is an exercise in patience.
The anti-malware measures in modern Windows are at least part of the issue.
There was an interesting conference video exploring this, about eighteen months ago. It was part of a project to make git faster on windows. I can't find it, maybe someone else will remember.
I wish a kernel engineer could give a good answer to that question.
One side of the answer could be that the software that is bigger, but honestly that doesn't explain everything.
I wish somebody could confirm Wirth's law is real and that there are valid example of it.