Misconception 1 — "MoE is always cheaper"
Cheaper to compute per token, yes. Cheaper to host, no. The total parameter pool has to live in memory. For workloads where memory is the bottleneck (single-GPU local inference, long batched generation, extreme context), MoE often makes total cost worse than dense.
Misconception 2 — "Each expert specializes in a topic"
Tempting but false. Research on Mixtral and Switch Transformer shows expert specialization is mostly at the token-level / syntactic scale — punctuation, certain morphemes, language-pair patterns — not topical ("the science expert", "the literature expert"). The headline interpretability picture is more mundane than people hope.
Misconception 3 — "MoE pricing should be based on active parameters"
Providers price based on total cost of serving (memory + utilization + load-balancing overhead), not just active FLOPs. A 671B-A37B model is not priced like a 37B model; it is priced like a "large MoE" tier. Read provider pricing pages literally.
Misconception 4 — "MoE is the future, dense is the past"
Both are still actively developed in 2026. Dense is still the right answer below ~30B and for fine-tunable, locally-deployable, and debuggable workloads. MoE is still the right answer for frontier capacity at affordable serving cost. Neither replaces the other.
Misconception 5 — "Top-K higher means better"
Higher K means more experts firing per token, which means more expressive routing decisions and more compute. It is a tradeoff knob, not a quality knob. Switch Transformer's top-1 worked. DeepSeek's top-8 works. The right K depends on N, model scale, training regime, and serving target.
The reading rule
When you see an MoE claim, ask: "Is this on the compute axis, the memory axis, or both?" Most marketing collapses them. Most architectural reality keeps them separate.