The earliest mention of the 30x paradigm I could find is in the original Illumina whole-genome sequencing paper: Bentley, 2008. Specifically, in Figure 5, they show that most SNPs have been found, and that there are few uncovered/uncalled bases by the time you reach 30x:
These days, 30x is still a common standard, but large-scale germline sequencing projects are often pushing down closer to 25x and finding it adequate. Every group doing this seriously has done power calculations based on specifics of their machines and prep (things like error rates and read lengths matter!).
Cancer genomics is going in the other direction. When you have to contend with purity, ploidy, and subclonal populations, much more coverage than 30x is needed. Our group showed in this 2015 paper that even 300x whole-genome coverage of a tumor was likely missing real rare variants in a tumor.
On the whole, the sequence coverage you need really depends on what questions you're asking, and I'd recommend that anyone designing a sequencing experiment consult with both a sequencing expert and a statistician beforehand (and it's even better if those are the same person!)