Metadiscourse and lexical bundles are two closely related concepts and both operate as overlapping functional units in texts. Metadiscourse analysis always takes a top-down approach, in which discourse analysts begin from pre-determined metadiscourse items down to the analysed texts. Lexical bundle analysis usually uses a bottom-up approach, in which the analysis begins with bundles, extracted automatically from texts, up to generate metadiscourse items to reach an understanding of the discourse. The bundle-driven bottom-up approach is likely to lead to the discovery of longer metadiscourse units and create new categories, while at the same time allowing for the verification of existing researchergenerated metadiscourse lists. While many researchers have focused on examining metadiscourse in academic writing, few studies have used a bottom up approach beginning with lexical bundles in this way to explore the use of metadiscourse. Moreover, research on sentence initial bundles is rare. The present study explores the metadiscourse functions of generated four-word sentence initial bundles from the corpora of Chinese L2 and New Zealand L1 masters and PhD theses, and compares bundle distributions between L1 and L2 thesis writing. Except for a few propositional bundles, all the other bundles were identified as metadiscourse bundles and two new categories (introduction bundles and condition bundles) were created in to supplement those in Hyland's (2005a, 2005c) metadiscourse model. In contrast to New Zealand thesis writing, both the Chinese masters and PhD corpora were characterised by the heavy use of code gloss bundles (e.g. In other words, the), condition bundles (e.g. In the case of) and booster bundles (e.g. It is obvious that), and a relatively low use of endophoric bundles (e.g. The use of the), introduction bundles (e.g. There are a number), attitude bundles (e.g. It is interesting to), hedge bundles (e.g. It is possible that) and self-mention bundles (e.g. In this chapter, I). These findings indicate how productive bundle-driven metadiscourse analysis is in expanding the scope of current metadiscourse studies. It also suggests that L2 students could benefit from having attention drawn to lexical bundles as metadiscourse devices to support their academic writing.