There's one beautiful, simple equation that has it all, called the Shannon–Hartley theorem:
$$C = B \cdot log_{2}\Big(1+ \frac{S}{N}\Big)$$
It says that over a channel with a given quality, the capacity (bit rate) C is proportional to the channel's bandwith B. The quality of the channel (signal vs. noise) hides within \$log_{2}(1+S/N)\$, and the bit rate includes redundancy (error checksums and the like).
The best data rate can be achieved with a low-noise installation of a channel that offers a high bandwith.
Of the wiring systems in question, a simple two-wire phone line will have the lowest bandwith and the worst noise properties (crosstalk and interference from neighboring lines, ...), twisted pair wires increase the bandwith and are more immune to external noise with an increasing "CAT number" (6 being better than 5e, being better than 5, ...) and systems with optical fibers are even better.
A telephone connection has an audio bandwith limited to a few kHz. Old systems had filters and the wires were often not capable of much more than the filter-defined bandwith. Digital subscriber lines (DSL) take advantage of the fact that many phone lines, when not being filtered, can take more than the mediocre audio bandwith of telephones. Beyond approx. 200 bps, it depends on the installation of the last mile and in your house (and your provider's willingness to use it in the best way). Typically, fiber can handle a greater bandwith than copper, but good quality can be achieved with copper, too.
Note: Someone selling you "Fiber to the Home" over copper wires is just doing (un-? clever?) marketing. Claude Shannon was way cooler, he didn't even care about the type of channel (copper, fiber, radio waves, whatever), he just looked at the bandwidth and the quality (signal-to-noise ratio). You can join Shannon and, like him, enjoy the theory and also don't care about the material of your wires. When I took my communication theory class at college, my professor was actually very right when he pointed out the beauty of shannon's work and said that the equation mentioned above was the E = mc2 of the information technology age.
Well I've made 1 million units for a product and they all had silkscreen and we fought over the cost of a resistor so it's not that cost prohibitive. Yeah I guess there is a cost associated with that but it's not that much. Also when you need to do rework, or when at the end of the line they are repairing boards that didn't pass testing, you want to be able to say "yeah replace U1 and change R17 to 33 Ohms" without having to haul out the schematic and the layout. Sure some factories will have computers with your drawings out there, and some have dirt floors ;)
For 402 components or vias just move your silkscreen, I mean I have 201 components that are labeled properly it's a matter of taking the time to do it.
So in short I agree with you I always prefer silkscreen, the only time I don't do it is when I'm making something for a hobby for myself and I'm being really cheap. Even then I usually try to label the parts in copper. Not saying you should do that for a real board though.
Best Answer
For items such as you use as examples, there is usually a NAND flash chip (or more than one) hanging off the main processor.
For smaller items (say microwaves), the code will either be in internal flash on the microcontroller, or in very cost-sensitive high-volume applications will be in ROM, again usually on the same die as the micro.