Poisoning models during pretraining is a particularly concerning threat because training data is sourced from the public web, which adversaries can easily manipulate (Carlini et al., 2023). Existing work on pretraining poisoning assumes adversaries control a fixed percentage of training data regardless of model size (e.g. 0.1% in the work of Zhang et al. (2024)). However, since the optimal amount of training data scales with model size (Hoffmann et al., 2022), even small poisoning percentages translate to unrealistically large volumes of poisoned content for large models, implying the practical risk of these attacks reduces with scale. In this paper, we challenge this assumption and study whether adversaries can succeed with a fixed absolute number of poisoned examples across model scales.