The problem comes from the number of inlets. You can:
a) a common inlet to all of your channels, or b) independent inlet for each channel, or for several channels
a) might seem intuitive, the problem is limitations on channel width, the inlet would have to feed a single channel which would then split (they would split to the inlet channel from the paper), structural limitations of PDMS and manufacturing have maximum widths and heights in millimetres at best. Which would not be enough to achieve the throughput required. The original 'master' inlet would probably have to be at least in the 10-20 cm range to achieve flow rates that could make this a household filter.
Even with another, stronger material, and perhaps manufacturing techniques I'm not aware of you still have limitations of pressure, their channels are very small in the paper (for a reason), pressure becomes a limiting factor to prevent their failure.
The problem with b is sheer complexity, you're talking about 10 000 tubes and connectors and holes punched into a chip if you even do a 1 - 60 split. The chip with just the channels would be about the size of a tissue box, this wouldn't be able to accomodate the channels so now you're talking about something tens of metres x tens of metres. This is hugely cost prohibitive and the channels from connectors would have to be really long, really long channels need more pressure to drive the liquid requiring more energy which reduces the efficiency significantly.
The reason parallelizaton works in pharmaceutical / DNA testing applications is you're going from microlitres of DNA/samples to nano, or picolitres in the channels, a single inlet can sufficiently provide that throughput (you're talking 10 uL / hour perfusion rates)
Edit: TL;DR version: Construction constraints would make this have worse efficiency than reverse osmosis and cost a lot more to manufacture as well.
The advantage being reverse osmosis is more efficient when the salinity difference is smaller. Still I don't think the added complexity is going to be worth slightly lower energy costs.
What about parallelization? Are there any obvious downsides to simply having lots of the active units, aside from the (what seems to me largely manageable) increase in micro fluidic chip complexity?