One can definitely get inexpensive and high-quality rubidiums for dirt cheap on the second-hand market.   I've specifically ignored those when discussing price as options as one can never be sure about their accuracy or long-term reliability, and I try to filter my suggestions on NANOG based on what I'd want to put in my production network (putting on my ISP network administrator hat for a second).  Many of those rubidiums have lived their lives in the harsh environment of a cellular tower's equipment enclosure and were pulled out after many years of service.

A brand-new rubidium oscillator is typically just under $2K  (A common brand is $1695 as we speak).   This is just the oscillator.  To this you have to add various support items such as power supplies and signal conditioning and heatsinks and enclosure to end up with something which will connect to some sort of NTP server.   Note that this doesn't provide any way to actually phase-align that rubidium oscillator to UTC.  For that, you'd have to add more hardware.

Professionally, I keep looking for a supplier of lower cost rubidum oscillators but have not found anything much cheaper than the SRS PRS10's at the $1695 I mentioned above.   

There are lots of ways to improve a GPS-based NTP server.  Better antenna positioning.  Better GPS chipset.  Paying attention to antenna patterns.   Adding notch filters to the GPS feed.  And so on.   Once you get the 1PPS out of the GPS receiver you then can utilize the 1PPS signal to discipline some sort of clock.  Maybe a TCXO for a day or so of NTP holdover.   Maybe a rubidium for a year or more of holdover depending on the accuracy you need.   You can add software to filter the GPS signal to limit the likelihood of time injection attacks.  And on and on.    It really comes down to how much money you want to put into the "appliance" you're building.   But, in the end, there is nothing better than adding a second GPS source at a diverse location as far as improving reliability, provided that's an option based on timing needs.

I can also attest that there is at least one overlap between time-nuts and NANOG.... occupational hazard here.  From my desk I can reach out and touch two rubidium oscillators, one is GPS synchronized and one is freerunning.   There are a couple others unpowered in the box of spares in the other room.   There are also at least 3-4 TCXO-based GPSDOs floating around (including one I use for my reference source), and don't get me started on the T&M equipment I have collected for comparing various time sources.   So far I've avoided spending the mid-5 figures to get a decent cesium oscillator.