The idea of a metaverse holds considerable promise for language learning. The phenomenon of an integrated digital and physical world, accessed in a 3D immersive environment, offers powerful opportunities for incidental language learning through informal chats, simulations, and games, as well as for dedicated language acquisition through personalized, adaptive tutoring. For that vision to be fulfilled, multimodal AI will be needed, moving beyond text-only language models, enabling input and output in any combination of media. Integrating multimodal AI into virtual reality (VR) will allow immersive experiences to be wide-ranging and free form, supplanting scripted interactions that offer limited learner agency. At the same time, memory capabilities built into AI systems will enable creation of learner profiles used to individualize interactions according to learner goals/interests and proficiency level. The copresence of peer learners, as well as the growing ability of AI systems to mimic human communicative practices, will enable a version of the metaverse, with its co-habitation of humans and AI agents, to become a collaborative space for social learning. While such a system will provide an engaging learning space, concerns of privacy and ethics will need to be addressed. For learners, it will be important to supplement AI-based virtual interactions with authentic human-to-human communication, for example, through engaging in virtual exchange. To understand the complex, intertwined interactions between humans and AI, it will be helpful to look beyond tool functionality and consider AI (and VR) from a holistic, ecological perspective, using frameworks such as sociomaterialism and complexity theory, along with insights from non-Western ontologies.