Dylan Patel is the founder, CEO, and chief analyst of SemiAnalysis, the research firm that turned semiconductor supply-chain minutiae into one of the most-read intelligence products in the AI economy. What began on his 24th birthday — May 22, 2020 — as a real-name blog about chips now sells data and analysis to hyperscalers, AI labs, chipmakers, and the investors financing them, with The Information reporting roughly 60 employees and more than $100 million in projected 2026 revenue.
Identity and formation
Patel's origin story is unusually well-documented because he tells it himself. He grew up inside his family's small businesses in rural Georgia — living in the motel his parents ran, later working around their gas station. The hardware obsession started with a broken Xbox 360: fixing the "red ring of death" by shorting a temperature sensor opened, in his phrase, Pandora's box. By about age twelve he was moderating online forums and Reddit communities covering Android, Apple, Google, Intel, NVIDIA, AMD, and PC building — years of public arguing about silicon before he owned much of the hardware he argued about.
He studied at the University of Georgia's Terry College of Business from 2014 to 2017, taking degrees in Risk Management & Insurance and Management Information Systems plus a legal studies certificate — no semiconductor-engineering degree. He then spent about two years as a quantitative analyst at a small risk firm, leaving after a dispute over credit and bonus. Through it all he kept posting, anonymously, until an internet argument got him doxxed; after a brief scare he concluded the obvious and started writing under his own name. That blog became SemiAnalysis, and the posting habit never left — he still answers critics online personally, against the advice of his own marketing team.
Building SemiAnalysis
The firm's model is to be more granular than anyone else willing to publish. SemiAnalysis reports treat the AI industry as a physical object: HBM stacks and CoWoS packaging capacity, accelerator bill-of-materials, cluster network topologies, chiller counts, substation timelines, and turbine fleets — each converted into dollar and gigawatt figures that executives and traders can price. The Datacenter Industry Model sells building-by-building tracking of more than 5,000 datacenters assembled from property records, permits, power usage, FOIA requests, and computer-vision reads on satellite imagery.
He began hiring in 2022 — the first employees were people he knew from Discord — and by his own account the team passed thirty in 2025. The newsletter itself — free analysis up front, paid depth behind it — is among the largest technology publications on Substack, past 200,000 subscribers on reported figures. The business expanded from a solo newsletter into institutional products — the accelerator, datacenter, energy, memory, and cloud-TCO models, plus the open InferenceX inference benchmark — and, in 2026, into adjacent businesses: a reported $400 million venture fund filing and the September acquisition of Citrini Research. Patel's own career now spans analyst, publisher, and investor: reporting describes stakes in roughly twenty AI-infrastructure startups.
Signature reports
Three 2023 pieces made the firm's name. In May, SemiAnalysis published the leaked internal Google memo "We Have No Moat, And Neither Does OpenAI" — a Google researcher's argument that open-source AI would eat both companies — after verifying its authenticity. In June came "AMD MI300 – Taming The Hype," the deep dive on the first credible NVIDIA challenger. In July, with Gerald Wong, he published detailed alleged specifics of GPT-4 — about 1.8 trillion parameters, a 16-expert mixture-of-experts design, roughly 13 trillion training tokens — a report whose numbers were never officially confirmed but shaped public understanding of the model. A December Gemini analysis coined the vocabulary the industry still uses: "GPU-Rich" and "GPU-Poor."
The second signature era is physical infrastructure. The March 2024 "AI Datacenter Energy Dilemma" forecast US AI datacenter demand above 28 GW by 2026 and argued power, not chips, was the binding constraint — a call the firm considers vindicated. The January 2025 "DeepSeek Debates" dismantled the viral "$6 million model" figure during the market panic. And its xAI coverage counted chillers, turbines, and Megapacks across the Memphis Colossus sites while the world's largest training cluster went up.
Influence and friction
The reach is real: Odd Lots, Dwarkesh Patel, BG2, and Sequoia's Training Data have all hosted him, and at GTC 2026 Jensen Huang gave InferenceX a keynote callout — conceding on stage that Patel's "sandbagging" accusation about benchmark claims was "not wrong." Ben Thompson and Jon Y endorse the publication on its own site.
The friction is real too. In spring 2026 SemiAnalysis sued former employee Wei Zhou for breach of contract and trade-secret misappropriation; Zhou countersued, alleging he was fired for refusing to put what he described as material non-public information into client models — allegations the firm denies and that surfaced questions about conflicts between Patel's personal investments and the research his firm sells. Both disputes were pending as of this index.
What the record does not settle
The soft spots are biographical precision and contested claims. The birthdate rests on his own birthday-post account; the headcount and revenue figures are snapshots from different months; the GPT-4 numbers were sourced estimates OpenAI never confirmed; and the Zhou litigation — including the conflict-of-interest questions it raised about his investing alongside his analysis — is unresolved. The index preserves those seams rather than smoothing them over.
This index was compiled from public sources and does not imply the subject's endorsement. Citations live in the packet's source catalog.