HTML5 video is completely fucking useless, because:
1. You can't stream video. (No, not a file, I mean live video.)
You can stream live video just fine, if the format and browser supports it. I've heard reports of people getting this to work just fine with Ogg. This isn't the big use-case anyway, though.
2. You can't full screen HTML 5 video. (The spec forbids this as a security flaw.)
3. There is no standard format, leaving you to encode an unknown number of versions. Hell, even if you stick with just H.264, you still need to encode to multiple profiles if you want to support everything.
HTML5 video is not yet usable as the only video format anyway, since you have to support IE
4. You can't seek in videos in anything remotely near a reliable manner. You know how you can link to a certain time in a Youtube video? Not possible in HTML5.
5. You can't switch to lower/higher-bandwidth versions while the video is playing.
The HTML5 spec as is stands today is useless. The features it does offer above HTML4 already exist and are handled better via existing specs or plugins. Pretty much anything that isn't canvas or video isn't implemented anywhere, making the features entirely useless instead of done better elsewhere.
HTML5 is a vast spec that includes a huge number of improvements and refinements. The HTML5 parsing algorithm, for instance, will make a lot of weird websites work exactly the same across browsers when formerly they all behaved a bit differently – it's enabled by default in Firefox 4, and experimentally available for WebKit. This and many other clarifications are giving each new browser release more opportunity to be consistent with other browsers.
Canvas, video, and audio are not yet as reliable, widely available, or full-featured as their plugin-based equivalents, but they're suitable at least as fallbacks where the plugins aren't available (iOS, sometimes Linux). SVG and MathML embedded in HTML (not XHTML) documents are usable on some cutting-edge browsers, with painless ways to fall back to rasterized versions (requiring barely more work than just using the rasterized version). Microdata can be used today and some search engines will recognize it.
HTML5 includes a lot of nice little things that can easily be used to augment your existing pages without too much work, while keeping your legacy code for legacy browsers. Examples of this are the autofocus attribute, the placeholder attribute, and getElementsByClassName(), all of which are supported in recent enough Firefox and WebKit (for instance).
And it also makes a lot of convenient things valid that browsers already supported anyway. This includes <meta charset="">, omitting quote marks and closing tags in more cases than HTML 4 permitted, standardizing innerHTML, standardizing contenteditable (although not precisely enough for real interop yet), standardizing autocomplete, and permitting custom attributes if they start with "data-".
I encourage you to read the Editor's Draft of HTML5 differences from HTML4 for a more comprehensive list of what HTML5 adds. Many of these features are supported in enough browsers to be useful today, which is why some HTML5 features are used by practically all major sites and all up-to-date JS libraries. There is no reason not to use the things that are useful today, which is a fairly decent amount – although nothing compared to what we'll be seeing five years from now when big features like canvas and video/audio are better-developed and more reliably available.
And, of course, all of this only talking about the HTML5 spec proper. If you include all the other things that are often labeled "HTML5", there's tons more to talk about: CSS3, Web Storage, Web Databases, WebSockets, and a host of other features. This is a time where the web is moving forward at a very rapid pace, in sharp contrast to the stagnation we saw five to ten years ago – and yes, quite a lot of features are already usable to some degree.