Dylan Patel
The independent analyst who traces the semiconductor supply chain down to its raw chemicals
Dylan Patel is the most influential independent semiconductor analyst of the AI-compute era and the Founder, CEO, and Chief Analyst of the research firm SemiAnalysis. He holds no engineering degree: rejected by MIT and Stanford, he graduated from the University of Georgia's Terry College of Business in data analysis, risk management, and legal studies. As a child working night shifts at his immigrant parents' rural Georgia motel, he taught himself semiconductor process technology and supply chains on chip-geek forums. On May 22, 2020 — his 24th birthday — he launched SemiAnalysis as a solo Substack post, growing it from a one-person venture into a multi-country research firm serving hedge funds, venture capital, hyperscalers, and chip manufacturers. His core method is tracing the supply chain from Austrian chemical suppliers all the way to the GPU in a datacenter, using satellite imagery to track datacenter construction and TCO (total cost of ownership) models to force executives to confront math they would rather avoid. In May 2023 he published the leaked internal Google memo 'We Have No Moat' and introduced the now-standard 'GPU-rich vs GPU-poor' framework. He appears frequently on Lex Fridman, BG2 (Bill Gurley & Brad Gerstner), Dwarkesh, Latent Space, and Stratechery. His business model centers on deep reports and data products sold to institutional clients, though the blurred line between research and investing has drawn controversy.
Methodologies
- End-to-End Supply-Chain Tracing - Trace from the most upstream chemical supplier all the way to the GPU in a datacenter, marking every monopoly and chokepoint along the way.
- TCO and Bottleneck Cost Math - Use total-cost-of-ownership models to cut through nominal compute, then walk the chain to find the bottleneck that truly caps expansion.
Key decisions and timeline
- 2011 Motel Night Shifts and Self-Teaching on Chip Forums - Real expertise can be born outside institutions, given enough curiosity and first-hand digging.
- 2014 Rejected by MIT and Stanford, Enrolls at the University of Georgia - Rejection does not mean the path is blocked; an unconventional skill mix can be rarer and more valuable.
- 2020-05-22 Founds SemiAnalysis on His 24th Birthday - Public, free, hyper-specific content is the fastest path to building authority from zero.
Beliefs and mental models
- Belief 1 - Having no engineering degree and being rejected by elite schools is no barrier to becoming the most authoritative voice in a field. Real authority comes from understanding something more deeply than the people who are supposed to — reading primary sources, building models by hand, talking to engineers at every layer of the supply chain, and writing about it relentlessly until the whole industry has to take notice.
- Belief 2 - In the AI race, outcomes are decided less by small algorithmic differences than by who has more and more-efficient compute. His 'GPU-rich vs GPU-poor' lens splits the entire industry: the acquisition, deployment, and unit economics of compute predict an organization's fate more reliably than the model itself.
- Belief 3 - Everyone watches the GPU, but what makes GPUs possible — and what will stop them — sits upstream: TSMC's advanced logic capacity, HBM high-bandwidth memory, CoWoS packaging, grid power, and at the very bottom ASML's EUV lithography machines. To understand the limits of AI-compute scaling you must walk the supply chain to find the single most-constrained 'lowest rung.'
- Model 1
- Model 2
- Model 3