If the delay down the copper wire is less than the encoding delay though, surely bits will be being received too quickly at recipient end, faster than they can be decoded into a digital stream which could be buffered?
I think the key point you're missing is that it's entirely possible for more than one bit to be "in flight" on the wire at any given time.
For example, if the wire is 100 m long, and the velocity is 192 x 106 m/s, and the bit rate is 100 Mb/s, then 52 bits of data will actually be "on the wire" at any given time. The receiver, however will only be aware of the 1 bit that is actually arriving at the receiver at that instant.
If the transmitter is sending bits at 100 Mb/s, then the receiver must receive and decode these bits at 100 Mb/s. The length of the wire changes the latency time between these two events, but it has nothing to do with the rate at which the receiver must deal with the incoming data.
Usually the receiver doesn't deal with the incoming bits one at a time, doing calculations at 100,000,000 operations per second. Instead it simply queues the bits up into something like a shift register, and then operates on them at a much lower rate, maybe 12.5 million operations per second, but operating on full bytes with each operation (or even at slower rates, but operating on larger data words).
Firstly you will need to check your end devices are 802.3af mode b compliant, otherwise you can't inject the power onto a discrete pair, but instead if they are mode a, you need to put the power onto the data pair which is usually a function of the switch.
With regards to UTP. It is possible you might be able to use non twisted cables if the distance is short, your devices will put up with a lot of dropped packets, sub 10Mbit speeds are ok and you don't have too much issue with external EMF sources.
But I would be much more inclined to simply run two shielded UTP cables.
Best Answer
No, you don't need 2.5 GHz bandwidth to transmit 2.5 Gbit/s. You can do it of course, but that would be very inefficient.
The data is encoded into symbols.
100Base-T Ethernet uses one pair in one direction and uses MLT-3 encoding to convert 4B5B encoded 125 Mbit/s bit stream into a signal that has 31.25 MHz bandwidth, so it uses 3.2 bits per hertz per channel.
2.5Base-T Ethernet uses all four pairs and complex encoding methods to achieve 6.25 bits per hertz per channel to end up being 100 MHz bandwidth for which the Cat 5e cable is rated for.
The encoding method includes the use of PAM-16 modulation, which means that each symbol on each wire pair is one of 16 voltage levels and thus carries four bits per symbol. It also uses Tomlinson-Harashima precoding (THP) precoding and DSQ128 encoding to achieve the goal of getting 2.5 Gbit/s sent over four pairs with 100 MHz bandwidth per pair, or 400 MHz total bandwidth for the whole link.