This paper reports the results of a self-paced reading experiment in Japanese in which the materials consisted of four versions of successively more nested syntactic structures. It was found that (1) people read the more nested materials slower than the less nested materials; and (2) the locus of the relative slowdown occurred early in the nested structures. There was no corresponding slowdown when processing the verbs at the end of each clause. The results are therefore not predicted by retrieval-based integration accounts of syntactic complexity. Rather, the results support expectation-based accounts of syntactic complexity for these materials.