Basically, nothing is really a hard limit. Seismologists push the data to its limits, so depending on the purpose you _really_ do need a broadband station. For other purposes, though, you could definitely get away with off-the-shelf accelerometers. We'll take whatever we can get. The more precision the better, but the more sensors you have, the less the precision matters.

As far as whether or not you really need 20 bits of resolution over the full frequency spectrum, you certainly don't for every application. For some applications, though, you need it over most of the frequency spectrum. For others (e.g. strong ground motion) you don't at all. (For a quick overview, see here: http://www.passcal.nmt.edu/content/instrumentation/sensors In most cases, the high frequency component isn't the problem on the sensor side, it's maintaining accuracy on the low-frequency component. High frequency sampling generates more data and results in more power consumption, though, so there's a different set of challenges there.)

The power limits depend on where the station is installed. Typically, they're solar powered with lead-acid batteries (or maybe something else these days) to store energy. Solar panels are expensive and fragile. The less power it needs, the easier it is to reliably deploy.

Overall, if the data is available, someone will push it to its limits. You have to make compromises, and which ones to make depends on exactly what you're trying to do. A ton of cheap, easily deployable instruments changes things. There's a lot of talk about things like this, but it has been hard to do in practice so far.

Take anything I say with a grain of salt, though. I'm an exploration seismologist who crosses paths with earthquake seismologists. I don't really know what I'm talking about.

At any rate, this is all largely a non-answer. I think a lot of this is more within-reach than I realize.