One of the factors that the other, excellent answers don't address is signal to noise, although I agree that latency is often the main factor.

Wireless often has a dreadful signal to noise ratio. This weighs on the spectral efficiency $\eta$: how many bits one can send per symbol. For additive Gaussian noise, the number of bits per symbol that can be sent is, through the Shannon-Hartley form of the Noisy channel coding theorem (see also here):

$$\eta = \frac{1}{2}\log_2(1 + SNR)$$

(see also my answer here where I calculate in detail the capacity of fibre optic links). If a copper pair achieves a signal to noise of $10^4$ whilst a wireless network (often realistically) an SNR of 10, then this is roughly a factor of four gain in speed of the copper relative to wireless. The Shannon theorem assumes our coding is "perfect", which means that there is enough redundancy (checksums of parity bits and so forth) that the probability of error is vanishing. Real codes are not as good as this, and wireless networks often corrupt data; when this happens there needs to be a retransmission. If retransmissions are often, this devastates the spectral efficiency as what should take one transmission often takes two or three tries.

Lastly, copper pairs are not hugely slower in transmission than freespace. Although the drift velocity of electrons through wires is tiny, the wires only guide the signal energy, which mainly propagates in the freespace (or insulator) around the wires. So the signal velocity is an appreciable fraction of $c$. TEM modes in minimally insulated pairs propagate at just under $c$, the only deviation is from:

1. The skin effect, i.e. that the EM field penetrates the conductor slightly; and
2. the permitivity of the small amount of insulation you will need. Two bare wires running parallel in space support modes that propagate a fraction of a percent below $c$. A good working figure would be more likely to be $0.5 c$ or something like that. The old co-axial cables with the braided outer conductor and polymer in between the conductors get about $0.7 c$ if you measure the delay of a square pulse through one of these on an oscilloscope. It's an easy experiment to do if you can get hold of a roll of 10 metres or so of cable and even a modest, 100MHz scope. Try it; you'll be surprised! Indeed it would be good if you reported the answer for Ethernet cable back to this question as your own answer!

What you find is that most networks have a speed $\times$ bandwidth product and the underlying physics for this is a signal to noise ratio that worsens with length of transmission. To understand this, you need to so roughly the same calculations as I have done in my estimate of the capacity of fibre optic links: the principle is the same for copper and for freespace, assuming that retransmissions are not needed.

One of the factors that the other, excellent answers don't address is signal to noise, although I agree that latency is often the main factor.

Wireless often has a dreadful signal to noise ratio. This weighs on the spectral efficiency $\eta$: how many bits one can send per symbol. For additive Gaussian noise, the number of bits per symbol that can be sent is, through the Shannon-Hartley form of the Noisy channel coding theorem (see also here):

$$\eta = \frac{1}{2}\log_2(1 + SNR)$$

(see also my answer here where I calculate in detail the capacity of fibre optic links). If a copper pair achieves a signal to noise of $10^4$ whilst a wireless network (often realistically) an SNR of 10, then this is roughly a factor of four gain in speed of the copper relative to wireless. The Shannon theorem assumes our coding is "perfect", which means that there is enough redundancy (checksums of parity bits and so forth) that the probability of error is vanishing. Real codes are not as good as this, and wireless networks often corrupt data; when this happens there needs to be a retransmission. If retransmissions are often, this devastates the spectral efficiency as what should take one transmission often takes two or three tries.

Lastly, copper pairs are not hugely slower in transmission than freespace. Although the drift velocity of electrons through wires is tiny, the wires only guide the signal energy, which mainly propagates in the freespace (or insulator) around the wires. So the signal velocity is an appreciable fraction of $c$. TEM modes in minimally insulated pairs propagate at just under $c$, the only deviation is from:

1. The skin effect, i.e. that the EM field penetrates the conductor slightly; and
2. the permitivity of the small amount of insulation you will need. Two bare wires running parallel in space support modes that propagate a fraction of a percent below $c$. A good working figure would be more likely to be $0.5 c$ or something like that. The old co-axial cables with the braided outer conductor and polymer in between the conductors get about $0.7 c$ if you measure the delay of a square pulse through one of these on an oscilloscope. It's an easy experiment to do if you can get hold of a roll of 10 metres or so of cable and even a modest, 100MHz scope. Try it; you'll be surprised! Indeed it would be good if you reported the answer for Ethernet cable back to this question as your own answer!

What you find is that most networks have a speed $\times$ bandwidth product and the underlying physics for this is a signal to noise ratio that worsens with length of transmission. To understand this, you need to so roughly the same calculations as I have done in my estimate of the capacity of fibre optic links: the principle is the same for copper and for freespace, assuming that retransmissions are not needed.

One of the factors that the other, excellent answers don't address is signal to noise, although I agree that latency is often the main factor.

Wireless often has a dreadful signal to noise ratio. This weighs on the spectral efficiency $\eta$: how many bits one can send per symbol. For additive Gaussian noise, the number of bits per symbol that can be sent is, through the Shannon-Hartley form of the Noisy channel coding theorem (see also here):

$$\eta = \frac{1}{2}\log_2(1 + SNR)$$

(see also my answer here where I calculate in detail the capacity of fibre optic links). If a copper pair achieves a signal to noise of $10^4$ whilst a wireless network (often realistically) an SNR of 10, then this is roughly a factor of four gain in speed of the copper relative to wireless. The Shannon theorem assumes our coding is "perfect", which means that there is enough redundancy (checksums of parity bits and so forth) that the probability of error is vanishing. Real codes are not as good as this, and wireless networks often corrupt data; when this happens there needs to be a retransmission. If retransmissions are often, this devastates the spectral efficiency as what should take one transmission often takes two or three tries.

Lastly, copper pairs are not hugely slower in transmission than freespace. Although the drift velocity of electrons through wires is tiny, the wires only guide the signal energy, which mainly propagates in the freespace (or insulator) around the wires. So the signal velocity is an appreciable fraction of $c$. TEM modes in minimally insulated pairs propagate at just under $c$, the only deviation is from:

1. The skin effect, i.e. that the EM field penetrates the conductor slightly; and
2. the permitivity of the small amount of insulation you will need. Two bare wires running parallel in space support modes that propagate a fraction of a percent below $c$. A good working figure would be more likely to be $0.5 c$ or something like that. The old co-axial cables with the braided outer conductor and polymer in between the conductors get about $0.7 c$ if you measure the delay of a square pulse through one of these on an oscilloscope. It's an easy experiment to do if you can get hold of a roll of 10 metres or so of cable and even a modest, 100MHz scope. Try it; you'll be surprised! Indeed it would be good if you reported the answer for Ethernet cable back to this question as your own answer!

One of the factors that the other, excellent answers don't address is signal to noise, although I agree that latency is often the main factor.

Wireless often has a dreadful signal to noise ratio. This weighs on the spectral efficiency $\eta$: how many bits one can send per symbol. For additive Gaussian noise, the number of bits per symbol that can be sent is, through the Shannon-Hartley form of the Noisy channel coding theorem (see also here):

$$\eta = \frac{1}{2}\log_2(1 + SNR)$$

(see also my answer here where I calculate in detail the capacity of fibre optic links). If a copper pair achieves a signal to noise of $10^4$ whilst a wireless network (often realistically) an SNR of 10, then this is roughly a factor of four gain in speed of the copper relative to wireless. The Shannon theorem assumes our coding is "perfect", which means that there is enough redundancy (checksums of parity bits and so forth) that the probability of error is vanishing. Real codes are not as good as this, and wireless networks often corrupt data; when this happens there needs to be a retransmission. If retransmissions are often, this devastates the spectral efficiency as what should take one transmission often takes two or three tries.

Lastly, copper pairs are not hugely slower in transmission than freespace. Although the drift velocity of electrons through wires is tiny, the wires only guide the signal energy, which mainly propagates in the freespace (or insulator) around the wires. So the signal velocity is an appreciable fraction of $c$. TEM modes in minimally insulated pairs propagate at just under $c$, the only deviation is from:

1. The skin effect, i.e. that the EM field penetrates the conductor slightly; and
2. the permitivity of the small amount of insulation you will need. Two bare wires running parallel in space support modes that propagate a fraction of a percent below $c$. A good working figure would be more likely to be $0.5 c$ or something like that. The old co-axial cables with the braided outer conductor and polymer in between the conductors get about $0.7 c$ if you measure the delay of a square pulse through one of these on an oscilloscope. It's an easy experiment to do if you can get hold of a roll of 10 metres or so of cable and even a modest, 100MHz scope. Try it; you'll be surprised! Indeed it would be good if you reported the answer for Ethernet cable back to this question as your own answer!

What you find is that most networks have a speed $\times$ bandwidth product and the underlying physics for this is a signal to noise ratio that worsens with length of transmission. To understand this, you need to so roughly the same calculations as I have done in my estimate of the capacity of fibre optic links: the principle is the same for copper and for freespace, assuming that retransmissions are not needed.