从论坛发帖人到 AI 基础设施分析师：Semi Analysis 的起源故事

From Forum Poster to AI Infrastructure Analyst: The Story of Semi Analysis

迪伦·帕特尔 Dylan Patel · WisdomTree欧洲 · 2026-07-09 · 约分钟 · 原视频 ↗

本期速览 · Overview

Dylan Patel 分享了 Semi Analysis 从青少年时期论坛发帖到如今 90 人研究公司的历程，专注于 AI 和半导体供应链。

Dylan Patel shares the journey of Semi Analysis from his teenage forum posts to a 90-person research firm covering AI and semiconductor supply chains.

要点 · TL;DR

AI 模型成本每年下降 60 倍，重塑基础设施需求。
AI model cost drops 60x yearly for same quality, reshaping infrastructure demand.
推理模型和智能体导致内存多年短缺。
Memory faces multi-year shortage due to reasoning models and agents.
强化学习和智能体推高 CPU 需求，但稳态 GPU 比更低。
CPU demand spikes from RL and agents, but steady-state ratio to GPUs is lower.

核心观点 · Key points

AI模型在给定质量水平下每年成本降低约60倍。
AI models improve at about 60x per year in cost for a given quality level.
由于推理模型和智能体工作流，内存需求激增，导致多年短缺。
Memory demand is soaring due to reasoning models and agentic workflows, causing a multi-year shortage.
由于强化学习和智能体工作流需要更多CPU交互，CPU需求正在激增。
CPU demand is inflecting because of reinforcement learning and agentic workflows requiring more CPU interaction.
由于制造挑战，共封装光学（CPO）要到2028年底或2029年才能大规模部署。
Co-packaged optics (CPO) will not ramp until late 2028 or 2029 due to manufacturing challenges.
数据中心的自发自用发电正在激增，燃气发动机和太阳能+电池成为关键。
Behind-the-meter power generation for data centers is soaring, with gas engines and solar+battery becoming key.

反共识 · Contrarian takes

由于词元效率，使用最智能的模型通常更便宜，而不是更贵。
Using the smartest model is often cheaper due to token efficiency, not more expensive.
内存价格将飙升到智能手机和笔记本电脑被挤出市场的程度。
Memory prices will soar so high that smartphones and laptops will be priced out of the market.
由于CPO延迟和铜缆创新，铜缆在中短期内将优于CPO。
Copper will outperform CPO in the medium term due to CPO delays and copper innovation.
CPU需求激增是追赶效应；与GPU的稳态比率远小于当前炒作。
CPU demand spike is a catch-up effect; steady-state ratio to GPUs is much smaller than current hype.
将柴油卡车发动机改装为燃气发动机是数据中心发电的可行方案。
Diesel truck engines converted to gas are a viable solution for data center power generation.

本期章节 · Chapters（共 21）

引言与半分析起源 Introduction and Origin of Semi Analysis
构建SemiAnalysis Building SemiAnalysis
GTC时刻 The GTC Moment
推理基准测试与Blackwell性能 Inference Benchmarking and Blackwell Performance
开源vs闭源与AI投资回报率 Open Source vs Closed Source and AI ROI
Semi-Analysis的AI支出轨迹 AI spending trajectory at Semi-Analysis
成本优化与模型选择 Cost optimization and model selection
通过Token效率实现成本优化 Cost Optimization Through Token Efficiency
内存需求与定价动态 Memory Demand and Pricing Dynamics
RL与智能体引发的CPU需求拐点 CPU demand inflection due to RL and agents
市场结构与CPU竞争 Market structure and CPU competition
面向智能体的CPU优化与GPU-CPU比 CPU optimization for agents and GPU-to-CPU ratio
智能体工作流中的CPU权衡 CPU trade-offs in agentic workflows
CPU需求比与市场修正 CPU demand ratio and market correction
CPU需求动态 CPU demand dynamics
网络与光学的演进 Networking and optics evolution
数据中心能源挑战 Data Center Energy Challenges
数据中心与能源供应链 Data centers and energy supply chain
电源转换供应链 Power conversion supply chain
总结与供应链复杂性 Closing remarks and supply chain complexity
结束语与免责声明 Closing Remarks and Disclaimer

阅读全文双语转录 →