Mixture of Experts (MoE) ☆

An MoE model holds many expert sub-networks but routes each token through only a few — DeepSeek-V3 stores 671B parameters yet computes with 37B per token. That decouples capacity from cost and is how open labs reached frontier scale on constrained hardware. The trade: all those parameters still occupy memory, which pushes HBM demand even higher — an architecture choice you can read directly in the memory market.