You're arguing this from the standpoint of someone who assumes all scientific papers are centered around MathML. You're ignoring humanities, for example, and behavioral sciences where images and graphs are a lot more common than equations. Anything that can be a vector should be a vector, in my mind, and PDF works very well for this. It's not new technology, it's mature, and it works on essentially any device and platform.
What is more, PDF works for papers already published. I need to access stuff from JStor, from pre-digital archives of Nature and Science, and so forth. If you want to go and retroactively convert millions of articles into XHTML, you go ahead. The rest of us will scan them into a PDF with OCR.
What you really NEED for scientific papers is a format that works on any device, including paper. We need to have the information available to anyone, anywhere, not just on 1st world computing devices. PDFs do this quite well.