I would disagree that the majority of images on the web are photos. Yes, there are many photos on the web, but I'd wager the vast majority of images (or at least the vast majority of image requests) are tiny bits of connective tissue of web layout - little rounded corners, gradient backgrounds, icons, etc.
Rounded corners and gradients can both be handled without images these days, but even if you use images for things like that, that bandwidth (certainly) and number (probably) of photos dwarfs them.
Vector formats can replace simple images, sure. They can't replace photos. And vector art complicated enough to be interesting can be a huge CPU load to draw... and can often be a LARGER file than a bitmap made at a resolution appropriate for the screen. It's not a more efficient representation for complex images until you start getting to 300+ dpi images that fill a whole printed page.
Source: I'm an artist who works primarily in Illustrator and checks to see how horrible the svg performance on her more complicated images is every few years.
Every game company ever disagrees that vector graphics are inherently slow. They just need to be designed to work fast.
SVG just happens to be a very slow implementation of vector graphics.
And my point isn't that they're better than a 300 dpi image, it's that they're better than an infinite list of high-DPI images for every twisted combination of DPI and screen-size.
Rendering my art takes noticeable amounts of time whether it's a PDF renderer, an SVG renderer, or Illustrator's internal renderers. Even after a recent Illustrator update's main bullet point was "we rebuilt a significant chunk of AI's internals to make it tons faster". It may be theoretically possible to render vector images much faster than this, but ain't nobody stepping up to do this.
Moreover: a lot of the images people would like to serve in multiple sizes are presumably photos. Photos and vectors don't go together well; vectors are a lot better at simple images.
And finally. [Here](http://egypt.urnash.com/rita/chapter/01/) is the graphic novel I'm drawing entirely in Illustrator. The 72dpi bitmaps range from 15 to 212 kb. Printable 300dpi bitmaps range from 1.75 to 2.3 mb. They're CMYK rather than RGB, let's guesstimate they'd be 3/4 of the size -- 1.3 to 1.7 mb. Illustrator files? 948-34 MB. And that's with 'create PDF compatible file' turned off, which generally halves the size of an .AI file.
I'm going to pause here to let that sink in. Even distributing a CD full of 300DPI images of a nearly 200 page graphic novel is less data than the source Illustrator files. And converting them to SVG/SVGZ/PDF makes little in the way of size gain, either. A modest list of high-DPI images sounds like a hell of a lot less stuff to upload to your server, never mind that you'd only be sending the single bitmap that best fits the display at hand.
3D games have a lot of tooling to make things Render Fast. There are a ton of ways people have figured out to fake more complex images - LOD, bump maps, baked luminosity, lots more stuff I don't know offhand because I'm not a 3D person. Nobody has put anywhere near the same kind of effort into getting 2D vector images to render fast.
I would love to see someone build tools for doing the 2D equivalent of things like "turning a super hi res mesh into a simpler one plus bump maps" and rendering them quickly. Maybe I could finally stop rendering my stuff into bitmaps for the web. But I don't see anyone doing that any time soon.
anyway.
TL;DR: 2D vector rendering is sluggish and I don't see that changing any time soon, and the file sizes for interestingly complex images is a couple orders of magnitude larger than a 300dpi bitmap. Only the most trivial images are more efficiently served as vector files than as multiple bitmaps.
Serioulsy, fast vector graphics were a solved problem back in the late '90s. How is this still a problem today on the Web?