On September 22, SemiAnalysis released an in-depth technical report that deconstructs LLM inference services into four working stages—Prefill, Midfill, Decode Attention, and Decode Experts—forming a “token factory.” It analyzes the differing requirements for compute, memory bandwidth, and network across stages, uses Nvidia Blackwell GPUs as a case study, and argues that mixing these stages together loses the structural advantages of MoE architectures.
SemiAnalysis has acquired thematic investment research firm Citrini Research. Founders Dylan Patel and James van Geelen confirmed the deal to Bloomberg on September 11. Van Geelen will remain CEO of Citrini for now, and Citrini will maintain operational independence. The transaction price was not disclosed.