The Niquist rate says that the maximum rate of independent symbols (fp) is twice the frequency bandwidth of the channel (B).
fp ≤ 2B
Hartly's law introduces the idea that the amount of information that can be encoded in each indepenent symbol depends on the number of different levels that the receiver can distinguish and hence the information rate depends on the frequency bandwidth, the amplitude of the signal (A) and how accurately the receiver can distinguish signal levels (ΔV).
R ≤ 2Blog2(1 + A/ΔV)
The logarithm in there is notable, as it shows us that while increasing the number of levels the receiver can distinguish increases the data rate there are diminishing returns.
The Shannon-Hartly theorem builds on these ideas and the statistical properties of noise and relates the maximum theoretically achievable information rate to the Bandwidth and signal to noise ratio for a channel subject to additive white Gaussian noise and with an encoding scheme that has unlimited context to use for error correction.
C = Blog2(1 + S/N)
Theoretical maximums and real-world systems are not the same thing. 10BASE-T Ethernet was designed more for simplicity than for bandwidth efficiency and only gets 1 bit per second per Hz for a 10MHz bandwidth. 100BASE-TX uses a muli-level encoding scheme which achieves just over 3 bits per second per Hz for a 31.25 MHz bandwidth. 1000BASE-T uses a more complex encoding scheme which pushes up to 4 bits per second per Hz and it uses all four pairs for transmission in both directions so a 1000 MBps data rate is achieved with only a 62.5 MHz bandwidth.
In other words, thanks to more advanced interface designs, while the data rate increased by a factor of 100, the bandwidth only increased by a factor of 6.25. Unfortunately after 1000BASE-T diminishing returns started to kick in.
10GBASE-T pushed up to 6.25 bits per second per Hz, but with four data lanes that still results in a 400MHz bandwidth. Too much for many existing cable plants. This resulted in the creation of 2.5G and 5G standards based on the modulation scheme used by 10GBASE-T but with lower data rates.
There are also 25GBASE-T and 40GBASE-T standards which take the modulation scheme used by 10GBASE-T and scale it up to higher bandwidths, but I can't seem to find any evidence of anyone actually selling them.
The headline frequency figures for a cable type are only part of the story. A cable doesn't suddenly stop carrying signals at all at a given frequency, the performance just gets gradually worse (more attenuation, more crosstalk, more dispersion). Eventually the signal degradation will reach a point that signals can no longer be reliably received.
Each speed of twisted pair Ethernet was designed for a particular cable type, but there are some sublties.
- 10BASE-T was designed for Cat 3 or better.
- 100BASE-TX was designed for Cat 5 or better.
- 1000BASE-T was designed to run on Cat 5, however it was discovered that the Cat 5 specification did not sufficiently constrain crosstalk, so while it would usually work there was a possibility of problems, Cat 5e was introduced to rectify these deficiencies.
- 10GBASE-T requires Cat6A for the full 100 meter distance, it can work over real world CAT 6 at shorter distances, but similar to the situation with 1000BASE-T over CAT 5 this can't be gauranteed on a "worst case" cable.
- 2.5GBASE-T is designed for Cat5e
- 5GBASE-T is designed for Cat6 and shorter runs of Cat 5e.
- 25GBASE-T and 40GBASE-T are supposed to require Cat 8, but it's not clear to me whether those standards will ever be a thing, or if speeds that high will stick to fiber (and direct attach cables for short runs).
Ethernet did not traditionally take cable parameters into account when determining what speed to link up at, though some manufacturers have added their own extensions to the negotiation process which do. For 2.5 GBASE-T and 5GBASE-T the NbaseT alliance came up with a "downshift" feature to take account of cable quality, but it's not obvious if it's an IEEE standard or not.
Even where negotiation does take account of cable parameters, it's still limited to a fixed set of standard speeds. It's not like Wi-Fi or modern DSL variants which are highly adaptive to channel conditions.
So what to do in practice?
If you think 1G will be enough to last the life of the cabling, then just get at least Cat5e.
If you think you may want 10G now or in the future then you may consider Cat 6 or 6A, the trouble is though that actually achieving the Cat 6 and 6A standards can be challenging, there is lots of crap out there that doesn't meet the standards it claims and field terminations require a lot of care to meet the 6A specification. Cat6a also seems to be substantially more expensive than Cat5e.
I don't think there is any point installing Cat 7 or Cat 8 at this stage, if you think you may need more than 10G at some point in the future, either install single mode fiber now or install ducts/conduits so that you can install the cable you need when you need it.