Earlier quoted context omitted.
And hardware vendors are and will be very eager to add hardware specific quirks, tons of it. So there is no point to "abstract graphics API for any OS/CPU instruction set" then, when most of the work - and we are talking about graphics acceleration, where performance is everything - is and increasingly will be handling all of that quirks.
That's not how it works. Most games and apps are just fine with leaving some amount of graphics performance on the table. That's why engines like Unity are so popular: sure, you could always go a bit faster writing your own engine, but in most cases it isn't worth the trouble. Likewise, with graphics APIs, you have to weigh the significant benefits of greater compatibility against the benefits of using vendor-specifi…
Beyond that, even where performance is the priority, a generic performance win is usually more interesting unless the IHV is buying the developer time.