The first article I found about this was https://timkadlec.com/remembers/2018-03-19-how-fast-is-amp-r..., which claims that AMP is actually faster in practice, probably mostly because it encourages good habits. Where are you seeing that it's not faster? This article is from '18 so perhaps there has been a change in how well AMP performs versus the baseline.

Also worth noting that we should probably judge AMP as a whole product. What if it didn't make things faster at all, but did make them predictable and amenable to prerendering? The product is AMP+Cache+Prerendering. Prerendering unquestionably gives huge speed advantages in many scenarios. If you don't have AMP, you don't have prerendering, which means AMP-as-a-product gives significant speed advantages, since it's on the critical path to prerendering. (All of this is of course predicated on the idea that you can't practically prerender normal HTML pages, which perhaps you can, I don't know. I'm just saying if that justification for AMP is correct, it shouldn't necessarily matter that HTML+AMP markup is not faster than HTML markup.)