Network delay is often said to be composed of propagation, serialization, and queuing delay.
Propagation delay is the time it takes the physical signal to traverse the path. This delay is usually fairly constant if there are no route changes (it's more constant if the end-points are static connected via optical fiber and has more variation if one of the end-points is a moving airplane or satellite).
Serialization delay is the time it takes to actually transmit the packet. Depending on bit-rate it may be a significant portion of overall delay (for a modem user connected at 33600 baud upstream it would take about 450ms to transmit a 1500 bytes packet [each byte actually takes 10 bits to transmit counting start and stop bits; on the other hand modems have hardware compression that reduces serialization delay somewhat]), or may be negligible (it only takes 73us to transmit a 9180-byte packet on a 1Gb/s connection). In ARPANET days serialization delay was a significant delay component, since there were typically several low bit-rate links between remote hosts. In modern world where high bit-rate connections are becoming the norm serialization delay is becoming more and more irrelevant. In any case, for given packet size and path, serialization delay is constant (this may not be completely true due to hardware compression).
Queuing delay is the time a packet spends in router queues. This time depends naturally on queue lengths: for unloaded network it would be negligible; for network that is heavily congested it is usually the main delay component. It is the most variable delay component in a typical modern network.
For high-performance networks, such as those used by Internet2, delay is mostly the propagation delay, which is determined by fiber length and speed of light in fiber, about 201000km/s. In this case it evidently cannot be significantly lessened due to fundamental physical reasons if physical network configuration is such that packets take almost the straight line as their path.
Assuming steady network state, TCP throughput is inversely proportional to round-trip delay for a given loss rate and maximum packet size. Notice that this by itself does not limit throughput for high-delay links (such as satellite connections), if there's no random loss due to bit corruption or equipment inadequacy.
Delay is a significant parameter for interactive networking applications. Comfortable human-to-human audio is only possible for round-trip delays not greater than 100ms. Requirements of other applications may be even stricter.
For some applications, such as tele-immersion, delay on the order of 100ms per se may be acceptable, but any significant jitter is not.
Different QoS techniques may seek to minimize queuing delay in conjuction with loss, or make different promises.