Huygens is pretty cool, and we're still digesting it, but I'm not sure yet it's a game-changer. NTP is enough to synchronize clocks to sub-millisecond levels in most cases (both in our experience and in the comparisons in the Huygens paper), and that's enough for strict serializability. The problem is that what matters for CockroachDB is not the typical clock offset, but the worst case. Multi-millisecond spikes can still happen in all but the most carefully-tended environments (especially when virtualization is involved), and as far as I can tell Huygens doesn't do anything to avoid those. (TrueTime's innovation was in making the margin of error explicit, not just in tighter GPS-based clock synchronization)
Where Huygens is potentially a big deal for us is that in contrast to NTP, it is non-hierarchical: it is designed for nodes in a cluster to synchronize as a group instead of with an external authority (NTP is not exclusively hierarchical but that's how it's typically deployed). This could form the basis for a zero-configuration synchronization network within a cockroach cluster, more or less independent from the system clock.
-Ben