As someone who has worked with GPUs and software renderers for over a decade: I am pretty sure that their tech depends on a few types of repeatable data, which

As someone who has worked with GPUs and software renderers for over a decade: I am pretty sure that their tech depends on a few types of repeatable data, which they are able to cache effectively based on rotation -- in other words, they have come up with an efficient way of querying the front-facing voxels in a large set of data based on the resulting view matrix. Where this falls flat is if the data is not procedural, or it is not diverse - as you can see in the video, there is a bunch of the same data copied over and over. However you compress it, such detail is not free and I am guessing there is a lot of data that depends on a good deal of memory/storage to work properly. I am not so worried about animation or dynamic lights or textures as everyone else is. If they can render it to a buffer and get the normals/depth/UV coordinates, the rest of the rendering can be done in screen-space, including SSAO, deferred lighting, and similar rasterization tricks. Animation can also be rendered on top of the scene, and intersected with the former depth buffer. The only thing I am worried about is the size of the data set and ability to create more diverse landscapes.