I think there are two concurrent forces at play pulling in a opposite directions.
On one hand, the higher end of the embedded world can be replaced by a powerful CPU that can run Linux, and there is not much differences between embedded vs non embedded there. This is the "easy" but not cheap solution.
On the other hand, as always more sensors are needed, and as objects are more connected than ever, the density of micro controller running on a given device is exploding (from automotive to various medical devices).
So all in all, I am not sure that embedded is decreasing.
Don't forget per-unit savings at scale. That's the main reason 8-16 bitters (a) still exist at all and (b) sell for over a billion dollars a year. The 32-bit RISC chips on modern nodes are starting to eat into them. Yet, there will always be companies willing to use a tinier, weaker, or just weirder device to increase per-unit profits.
Heck, there's even still a market for 4-bit chips used by the likes of Gillette:
Yeah. If you can save 5 cents on the part cost, and you're going to sell 100 million of them, it's worth it (for the company) to burn a few person-years of engineering time dealing with the limitations of the chip (provided it's at least barely adequate for the task).
Yeah, exactly. And automotive itself is kind of just the tip of the iceberg (although admittedly huge by volume).
My current company is working with heavy machinery manufacturers, and the industry is years behind consumer automotive. There's a similar kind of explosion in the amount of small intelligent modules waiting to happen.
