A very big portion of this is just Wirth's Law in motion. Software is getting worse rapidly, and AI is accelerating the decline in software quality, where performance is an aspect of quality.
But the other part of this is the latency vs throughput tradeoff. More and more of our systems are optimized for throughput, or worse are optimized for throughput in ways which causes significant experience degradation if you are unable to meet the throughput demands. There is a natural tension between throughput and latency, and we are seeing this play out across every aspect of both hardware and software system design. As a simple example, nothing you do to USB4 will ever match the latency of PS/2 connection for an input device, but a PS/2 port delivers around 7-12 kiloBITS/second of throughput vs USB4 delivering up to 120Gbps and a /minimum/ of 20 gigabits/second of throughput. PS/2 uses blocking direct hardware interrupts for input devices, vs asynchronous polling on USB. This architectural difference will never allow for USB to match PS/2 latency, regardless how "fast" USB gets.
This tradeoff is leaning more and more towards throughput everywhere you look. Just today, the new M5 Ultra is announced w/ 1.2TBs/ of memory bandwidth. The M5 Max had 14ns of memory latency while high-performance DDR4/DDR5 typically achieves 8-9ns of latency. That's 36% more latency (conservatively), but nearly 10x the throughput. When you stack up small changes like this at every layer of the system, including in our overreliance on micro services and networked data I/O vs local data in applications, and it adds up. There are so many places in modern software where there is some network connection required for something that could have been achieved without that connection, and waiting on that connection consumes a huge amount of wait time relative to the total time for the operation, and we all feel it.