For each eye, there is an imaging device (720p/1080p resolution for example) that displays an image. In front of the imaging devices there are high speed mask device (LCD?) that blocks off portions of the image that are not at the current focal depth. The focal depth cycles quickly between 12 different values and easily driven by the z-buffer values from the renders. The light not blocked by the mask is partially collimated and guided towards the eye via optics and then a high resolution high speed digital zone plate is used to refocus the light at different depth levels using diffraction. The mask and zone plate update together at 12X (360Hz-720Hz) the rate of the image device (30-60Hz). The result is that pixels that should be at different depths are actually focused near that depth and the eye blends the 12 partial frames together to form one image that has a range of depth in it.
I expect they are using off the shelf parts for the imaging device (and maybe mask LCD) but manufacturing their own digital zone plates since they need super high resolution and high speed for those (at least 360hz) but only on-off values not color or grayscale. This is probably where a bit of their money is going.
If it works, you could get high resolution images per eye, and also see 12 buckets of depth so your eye would refocus to see things close up versus far away. This could potentially create a pretty good impression of something existing close to you, compared to existing AR where all objects are focused at a single fixed depth and there is a conflict between accommodation and stereopis cues.
Of course this is one of the patents filed, it could be just an idea they decided to file on, not what they are actually building.