D3DUSAGE_DYNAMIC is only useful when the buffer is locked frequently; data that remains constant should be placed in a static vertex or index buffer.

http://msdn.microsoft.com/en-us/library/windows/desktop/bb147263(v=vs.85).aspx

From a GPU point of view, it's a pretty bad idea. For most GPU's reading from a dynamic vertex buffer will be slower compared to a static vertex buffer, since dynamic vertex buffers need to be placed in CPU-accessible memory. From a CPU point of view it might let you reduce draw calls quite a bit which can remove a lot of CPU overhead, but of course you need to spend some CPU time copying the vertex data into the dynamic buffer as well (you may also need to transform the vertex position/normal/tangent data depending on how you set things up). In general instancing will be a better way of reducing draw calls, if you can make use of it.

Well he's correct that rendering a large number of vertices per-draw call is better than rendering a small number, but you'll get more mileage by ordering your data at load time (or - even better - in your content creation pipeline) so that you can achieve this with static buffers. And MJP is absolutely correct that you can burn more CPU than you save by doing too much work to order stuff at runtime; generally you should be aiming to get your runtime work as low as possible, preferably just setting states and issuing draw calls (which is not always achievable - e.g some particle system approaches - but should be possible for ~90% of what you see on screen).

An alternative mode of operation is to create a library of dynamic textures and create texture atlases so that all the vertices that belong to one model have access to all their different textures on one atlas.

I'm also curious about what the tradeoffs are. I think its cool he got it to work, how does he handle other operations that break a frame's worth of rendering into multiple draw calls? Stuff like different vertex declarations between two models, different render states, different shader constant registers, etc? Unless you have one uber vertex declaration, one uber vertex/pixel shader and do instancing on everything so that you don't need to change a world matrix using SetVertexShaderConstantF between each new model draw, I don't see how he can even win more than once or twice per frame.