In theory, almost any base in the human genome may mutate, so you have billions of variants to go. Ok, this is not so useful.
A related and potentially useful question is: given a human, what is the fraction of his/her variants seen in the the previously sequenced samples? For a Wright-Fisher population, there is an analytical answer: x% of variants on a haplotype has frequency below x%. Replace "x" with "1": 1% of variants on a haplotype have frequency below 1%. The 1000 genomes project concluded that 99% of SNPs at frequency 1% or higher have been called by the project. If we all came from a Wright-Fisher population, 99% of of SNPs have been called in 1000g. Given that the project also calls SNPs at lower frequency, nearly all SNPs in a newly sequenced sample are present in 1000g.
The above is a theoretical analysis. In practice, there are two major complications. First, we are not a Wright-Fisher population. Due to recent population expansion, each haplotype harbors more new mutations than the Wright-Fisher model would predict. Given a known population history, it is actually possible to numerically compute $f(x)$, the fraction of variants with frequency below x (for Wright-Fisher, $f(x)=x$). I don't have this result, though. Second, 1000g and many other genome projects didn't call SNPs in repetitive regions and missed a significant portion of high-frequency indels and structural variations. With Illumina, you can only call ~30% of long deletions callable with PacBio data, but we only have a dozen of PacBio human genomes in public. There are still a lot we don't know.