They run on machines that are designed to have unused computer capacity. Desktops, laptops, tablets, phones. Responsiveness is the key.

Ffmpeg runs on a load of server farms for all your favourite streaming services and the bajillion more you’ve never heard of. Saving compute there saves machines and power.

Your point is well taken but there is a distinction that matters some.

Does it?

I've had some instability using PyTorch on my desktop computer, that only appeared if I was using the computer. I just a few minutes ago discovered what the problem was, because while running a calculation I opened FreeTube, a YouTube interface that runs in an Electron instance, and my computer immediately restarted. Apparently 8 cores, 96 MB of cache, 32 GB of system RAM, and 16 GB of VRAM doesn't leave enough unused capacity for one more web browser instance.

I've helped multiple people figure out that their games were getting latency spikes because of the resources used by chat applications running on Electron or similar, which no amount of RAM or processing power seems to fix.

Besides the opertunity cost of the resources that extremely inneficient applications use making the computer much less useful for other tasks, power consumption itself can matter a lot. Sure, using more power on a desktop computer may only cost a fraction of a cent per day, but on laptops, tablets, and phones that don't have user replacable batteries, which is pretty much every one in production, every percent power using increase not only wares the battery the same percentage faster, which proportionally reduces the useful life of the device, but it reduces how long the device is usable between charges, requring topping off and reducing usability and reliability.

Also, running ICQ on a 90's computer with a mechanical hard drive was far more responsive than running Slack or Discord on the fastest computer you can buy today today. I can guaruntee you that switching from an Electron/HTML/CSS/JavaScript/WhateverJavaScriptFramework stack to a C/GTK or similar stack will not only reduce resource consumption by an order of magnitude or two, increase security, and make the codebase simpler and easier to mantain, it will also be much, much more responsive.

Saving compute in consumer facing devices saves machines and power, too.