Why Enough Bandwidth Still Stutters: Queueing and Round-Trip Time
First, separate two things: bandwidth is "how much data can be moved per second," and latency is "how long one round trip takes." The former is like how many…

First, separate two things: bandwidth is "how much data can be moved per second," and latency is "how long one round trip takes." The former is like how many lanes a road has; the latter is like how long it takes to drive from home to the office. Widening the road lets more cars pass at the same time, but it does not shorten the journey. A 1 Gbps link is not "faster" to a given destination than a 100 Mbps link; it simply fits more bits into the same interval.
Propagation time is fixed by physics. Light in fiber travels at about two-thirds of its speed in vacuum, roughly 200 kilometers per millisecond. Across the Pacific, about 10,000 kilometers, one way takes around 50 milliseconds, and a round trip (RTT) is about 100 milliseconds. That distance cannot be made up for with bandwidth. Conversely, sending the same 1,500-byte packet over a 1 Gbps link takes only 12 microseconds to serialize; over 100 Mbps it takes 120 microseconds. Compared with 50 milliseconds of propagation, that difference is almost negligible. So latency on a wide-area network is mainly not "not enough bandwidth" but "the distance is too great."
What is truly easy to overlook is queueing. A router's egress is a shared single-lane bridge; when packets arrive, they have to wait in line. Queue length does not grow linearly with average utilization: at 50% load there is almost no queue; at 80% it starts to stall occasionally; above 90%, queues suddenly lengthen, and latency jumps from a few milliseconds to tens or even hundreds of milliseconds. Averages are misleading because traffic is bursty—a link averaging 500 Mbps may be momentarily saturated within a given 10 milliseconds.
There is another commonly misunderstood concept: the bandwidth-delay product. The amount of data a link can hold "in flight" equals bandwidth times RTT. On a 1 Gbps path with 50 ms RTT, about 50 megabits—more than six megabytes—can be inside the pipe at the same time. The sender must have that much data outstanding before receiving acknowledgment in order to fill the pipe; if the window is too small, bandwidth sits idle. This is why simply upgrading the line does not change transfer speed for older protocols—the bottleneck is the window, not the pipe.
Applications also pay a "handshake tax." The TCP three-way handshake costs one RTT; TLS 1.3 adds another; an HTTP request to first byte costs another half to one RTT. At an RTT of 200 milliseconds, a brand-new connection has already spent around 500 milliseconds before the first byte goes out. This is also the point of CDNs and connection reuse: not to make signals travel faster, but to bring the starting point closer to you and, incidentally, to save repeated handshakes. HTTP/3 uses QUIC to merge the transport and cryptographic handshakes, typically saving one RTT, and even less when resuming a session.
Queueing also drags in other flows. A large download fills the egress buffer, and the adjacent voice packet that could have completed in 20 milliseconds has to wait behind the big packets—this is bufferbloat and head-of-line blocking. Link utilization may look like only 60%, yet the call breaks up.
How can you tell them apart concretely? First measure the path RTT; ping and mtr can reveal distance and hop count; then use iperf3 to drive traffic and observe throughput, but high throughput does not mean low latency. The real clue is this: if RTT is stable when idle, spikes as soon as load is applied, and recovers slowly, that is queueing rather than propagation. You can also run a "latency under load" test to see exactly how much RTT rises at full load.
The limitations should also be made clear. A single ping is subject to ICMP rate limiting, asymmetric paths, and NAT, so it cannot be fully trusted; retransmissions and channel contention on home Wi-Fi make RTT jittery, but that is a link-layer problem, not bandwidth; clocks and CPU preemption in virtualization and containers can also fake "latency." Measurement itself is noisy, so you need to look at multiple points, multiple times, and both directions.
So the conclusion is not complicated: bandwidth buys capacity, while latency buys distance and queueing. If the line is wide enough but still slow, first ask three questions—how far the signal has to travel, how much queueing is on the path, and how many times handshakes are repeated. Moving content closer to users, applying fair scheduling to buffers, and reusing connections are often more effective than pulling yet another gigabit line.