Good demo. The benefits are indeed attractive, but you still have not sold me on why this is a better approach than ... well, everything that's been suggested in this thread. This stuff is possible without what you're proposing, as you yourself admit. What I see amp giving is a framework in which it is impossible to "mess up", which can be beneficial but I agree with the general sentiment here: New tech has a cost to it. Learning it has a cost. Implementing it in browsers has a massive cost (because I'm sure this will eventually find its way in Chrome, which means Firefox will follow suit, which means a larger potential for technical debt and a higher bar to developing a web browser from scratch).

Like someone else said: If Google wants performance, Google can start aggressively penalizing article-type websites which make unnecessary requests etc. I know you guys already penalize slow sites, so what's wrong with that approach? Especially since such an approach has a much higher chance of actually working.