Algorithms for answering XPath queries on Xml streams have been studied intensively in the last decade. Nevertheless, there still exists no solution with high efficiency and large coverage. In this paper, we introduce early nested word automata in order to approximate earliest query answering algorithms for nested word automata. Our early query answering algorithm is based on stack-and-state sharing for running early nested word automata on all answer candidates with on-the-fly determinization. We prove tight upper complexity bounds on time and space consumption. We have implemented our algorithm in the QuiXPath system and show that it outperforms all previous tools in coverage on the XPathMark benchmark, while obtaining very high time and space efficiency and scaling to huge Xml streams. Furthermore, it turns out that our early query answering algorithm is earliest in practice on most queries from the XPathMark benchmark.