1) There's definite overhead on the API itself, e.g. the slow shader compilation times due to the verification, translation, etc. That'll get better over time, but it'll never match the desktop; I personally think a caching approach will be a big win here, but no one has done it yet.

2) You don't have geometry shaders, hardware tessellation, hardware instancing, and a number of other features you get on the desktop with modern hardware. You also don't get the ability to render to multiple targets at once (critical for performant deferred shading).

3) There's significant 'extension lag' as it isn't as simple as just implementing an extension on one chip and coding for that; it has to hit a certain level of penetration before it'll be ratified for inclusion in WebGL, then it has to be made safe and implemented.

Those are the core things holding WebGL performance back (IMO), excluding other browser bits that may get in the way, which aren't relevant to this discussion directly. A lot of this will go away over time, but I honestly have no idea how any of that will happen.

(Disclaimer: I work on graphics stuff for Mozilla, but don't work much on the WebGL side of things; mainly write demo code there)