The youtube client I use is able skip (some) mid roll ads.

It's not Twitch, video is encoded in advance in various formats for different devices.

Embedding the ads into the stream is pretty difficult.

It's not uncharted territory. It's just more expensive comput-wise, and I expect google to make that calculation to conclude it's not worth the cost yet.

But compute costs could drop, and change the equation.

It's not just compute. It's storage, and network costs too. Right now they can cache the ads and video next to the user. But Google likes to target ads at particular users, in fact they formalised this in FLoC. In FLoC (since renamed to Topics API and now shipping in Chrome) groupings of users are called cohorts. There are lots of cohorts, eg: male / female, baby boomer / gen-Y / gen-Z / millennial, parents / single, religious / secular, conservative / liberal, income range. An ad might be targeted at a specific combination, such as conservative male parent. But there are a lot of possible combinations. Not 10's but 100's. This means you have to produce version of the video that targets each possible combination, and that in turn means your storage, network and computer costs balloon by 100's.

It's a big number. They avoid it by having the browser doing the mixing. I can't see that changing.

They don't want to show the same ads to everybody: they maximize their advertising income when they show different ads to different people depending on what people are interested in and buy.