There were a few things (GDI handles and suchlike) that had very small limits. Once you exhausted them, the system was basically unusable. There was a little program you could run that would show the number allocated vs allowed. By the time you'd launched one program, they were normally 60-90% gone.

These two are related. OS/2 needed 16MB of RAM to be useable back when I had a 386 that couldn't take more than 5MB (1MB soldered onto the board, 4x1MB matched SIMMs). Windows NT had the same problem - NT4 needed 32MB as an absolute minimum when Windows 95 could happily run in 16 and unhappily run in 8 (and allegedly run in 4MB, but I tried that once and it really wasn't a good idea). The advantage that Windows NT had was that it used pretty much the same APIs as Windows 95 (except DirectX, until later), so the kinds of users who were willing to pay the extra costs could still run the same programs as the ones that weren't.

I never ran 3.0 on a 386 to try that. On Windows 3.1 it wouldn't work, because the OS required either (286) protected mode or (386) enhanced mode. Running 3.0 on a 386, the DOS prompt would use VM86 mode (yes, x86 has had virtualisation support for a long time, but only for 16-bit programs). Windows 3.0 could run in real mode, so would work inside VM86 mode. In real mode, it didn't have access to VM86 mode (no nested virtualisation), so probably couldn't start again.

I'm not sure if it ever became a real product, but the Novell guys did have Netware running as a paravirtualised Xen guest. Apparently it was quite an interesting challenge, as Netware was one of the few operating systems to use rings 1 and 2 on i386.

It's like rain on a wedding day.

What you need depends on the application, what you have depends on you meeting your design criteria.

32MB memory chips cost $5 in quantities of 1. If you have an IoT application that needs a full Android OS on it for some reason then the memory won't break the bank.

You answered your own question in the example. You want networking you either need to implement a network stack or offload that stack to another component in which case you're just cheating by moving code around and ultimately still using the same amount of resources.

Though I admit that 32MB sounds high for what we're talking about, I'll happily callout anyone who compares it to programming a simple PIC micro and thinking they are badass.

You jest but I'll bite. Google most definitely care about your privacy. In the number of companies willing to share your information with 3rd parties Google is right at the bottom of the list for one simple reason, control. There are a lot of companies out there that share the data they collect. Google keeps the data and provides an API to let someone target your anonymised data set.

Google is the number 1 offender for collecting my private data.
Google is also the number 1 company I trust not to simply sell it or share it.

Not quite the same. When you consider who to connect one of these microcontrollers an an IoT device you need all sorts of additional parts. We're basically offloading the software from the OS to specific hardware / firmware combinations that handle it for us. There's a lot more than 64bytes of code running on your project.

IoT is nothing of the sort. A few questionable corporations are creating that product that way but just because it says IoT doesn't mean that you need to give anyone else data.

Yeah it seems like only yesterday where whole OSes fitted on 10k.

Oh Wait...

And Windows 3.1 lost real mode support. You could run Windows 3.0 on an 8086 with an EGA screen and 640KB of RAM (I did - the machine originally shipped with GEM). I think 3.1 still have 286 protected mode support, but didn't work very well unless you ran it in 386 enhanced mode. It was a bit sad that the version of Windows that required an MMU didn't use it to implement memory protection...

Sort of. The desire not to cannibalise sales was a key factor in the design of the PC, but these were also features that IBM didn't think would be missed.

IBM knew what multitasking was for: it was to allow multiple users to use the same computer with administrator-controled priorities. Protected memory was for the same things. Why would you need these on a computer that was intended for a single user to use? A single user can obviously only run one program at a time (they only have one set of eyes and hands) and you can save a lot in hardware (and software) if you remove the ability to do more. And, of course, then no one will start buying the cheap PCs and hooking them up to a load of terminals rather than buying a minicomputer or mainframe.

My father's company got their first Windows 3.0 install because they bought a diagram tool (Meta Design, I think), that came with a free copy. The company that made it had decided that bundling a copy of Windows 3.0 was cheaper than writing (or licensing) a graphical toolkit for DOS and an associated set of printer drivers. I don't know if they were the only company to do this, but after a year or so they stopped bundling Windows and just expected their customers to either have a copy already or go and buy one.

Historically it was always the other way around - Russia had to defend themselves from France (Napoleon campaign), UK (Crimean war, intervention in the civil war 1918), us Germans (WW1 and 2) and so on.