Around 2018, YouTube became heavily criticized for its radicalizing function by allowing far-right actors to produce hateful videos that were in turn amplified through algorithmic recommendations. Against this ‘algorithmic radicalization’ hypothesis, Munger and Phillips (2019, 2020) argued that far-right radical content on YouTube fed into audience demand, suggesting researchers adopt a ‘supply and demand’ framework. Navigating this debate, our article deploys novel methods for examining radicalization in the language of far-right pundits and their audiences within YouTube’s so-called ‘Alternative Influence Network’ (Lewis, 2018). To that end, we operationalize the concept ‘extreme speech’—developed to account for ‘the inherent ambiguity of speech contexts’ online (Pohjonen and Udupa, 2017)—to an analysis of a right-wing ‘Bloodsports’ debate subculture that thrived on the platform at the time. Highlighting the topic of ‘race realism’, we develop a novel mixed-methods approach: repurposing the far-right website Metapedia as a corpus to detect unique terms related to the issue. We use this corpus to analyze the transcripts and comments from an archive of 950 right-wing channels, collected from 2008 until 2018. In line with Munger and Phillips’ framework, our empirical study identifies a market for extreme speech on the platform, which came into public view in 2017.