Yes, exactly, I simplified it a bit for the limited space here. Basically C ~ B * SINR where B is the amount of spectrum and C is capacity. The reason mm-wave yields such whopping throughputs is 1. space diversity, that is, each user gets a beam 2. the dedicated beam means that one site is not interfering with it's neighbor (co-channel interference) . So if you can control the RF into a tight beam, the SINR is greatly improved. So it boosts both B because not you are not sharing B you get all the B available to the carrier in your beam and SINR is greatly improved because your beam is not being interfered with by other cell sites.
You can not, unfortunately, modulate an infinite number of bits in practical implementations. The more QAM point require higher SINR to resolve. You can't have infinite SINR or Infinite Spectrum. Say you have 100Mhz of spectrum, in low band you are sharing it with others but also getting interfered by neighboring sites particularly at the cell edge.
My understanding of the Shannon limit is that it depends not only on the bandwidth but also on the signal-to-noise ratio. If you had no noise at all (including infinite control over the modulation and demodulation process), you could modulate an unlimited number of bits per symbol within an arbitrary finite bandwidth. Like a QAM constellation with 2⁵⁰⁰⁰⁰⁰⁰⁰ points or something, all of which could in principle be accurately distinguished from one another in the absence of any noise.
If the signal-to-noise ratio isn't a fixed thing for RF communications in different times and places, in what way is cellular communication newly reaching the limit now? I mean I understand that some digital coding methods are considered to be fairly near optimal for a given RF environment; is the idea that they previously weren't and now they are, so there's no longer much benefit to be had from engineering improvements in that area?