We may not be able to shrink fabrication processes much more than they are, but we can probably shrink them enough to get an order of magnitude improvement. Then, we can get one or two orders of magnitude improvement by building hardware designed for the task. Then, we may be able to get yet more improvements by going with creative solution. For example: chips that use analog electronics to implement multipliers with just one to five transistors. Noisy and not very accurate, but still usable for neural networks.
Besides this, Facebook has the compute power, right now, to run facial recognition on every picture being uploaded. Running facial recognition on videos isn't that far-fetched. Even if you think that's too computationally expensive, you could just run the facial recognition at a lower frame rate. There are also tricks that can be used to save computing power. If the system is able to track specific people, it can avoid having to analyze all the data coming in all the time.