How MoE Capacity Factor Silently Drops Tokens and Degrades AI Model Output

In Mixture-of-Experts (MoE) models, each expert is allocated a fixed-size token buffer determined by a capacity factor formula; tokens that exceed this buffer are silently discarded rather than processed. A dropped token bypasses the expert's feed-forward block and passes through on the residual stream alone, effectively turning that transformer layer into a no-op for that token. Because MoE routers tend to self-reinforce load imbalance, token dropping is a default risk rather than an edge case, and the problem surfaces under real production traffic rather than single-sequence evaluations. The failure is difficult to detect because no errors are thrown and output logits can appear normal, making per-layer, per-expert drop-rate logging essential for diagnosis. Potential fixes include raising the capacity factor, switching to dropless block-sparse kernels, or adopting auxiliary-loss-free balancing strategies such as DeepSeek-V3's per-expert routing bias.