He has no engineering degree. Jensen Huang knows his name anyway.
His parents ran a 24-room motel in rural Georgia. His mother managed the desk; his father had worked a Subway in Los Angeles before pooling family savings - and family loans - to buy a foreclosed property sight unseen. Dylan and his brothers grew up rotating through reception, laundry, and the night shift.
The night shift had a computer. The computer had the internet. And somewhere between the night audit spreadsheets and the quiet hours after 2am, Dylan Patel discovered chip-geek forums. He toggled between RuneScape and datasheets on semiconductor process nodes, because that is what some kids do when the world is asleep and there is nothing stopping them.
He applied to MIT. Rejected. Applied to Stanford. Rejected. Enrolled at the University of Georgia Terry College of Business instead, studying Data Analysis, Risk Management, and Legal Studies - none of which are semiconductor engineering, and all of which, it turns out, are exactly what you need to build the most influential chip research firm on the planet.
He graduated in 2017, spent a short stretch at a financial firm, grew irritated with how bonuses were distributed, and decided to do something else. On May 22, 2020 - his 24th birthday - he posted his first blog on SemiAnalysis. He had no investors. He had no team. He had a Substack account and an obsessive command of how silicon actually moves from Austrian chemical suppliers to your data center.
Semiconductors are infinitely complex. There's millions of people working on things and each person is in such a niche.
- Dylan Patel, SemiAnalysisSix years later, Jensen Huang stood at the GTC 2026 keynote and said his name. Out loud. In front of the entire semiconductor industry. Huang cited SemiAnalysis's InferenceX benchmark report and spent five minutes walking through its findings. Dylan Patel was one of exactly two individuals Huang named by name.
This is not a story about overnight success. It is a story about someone who understood something more deeply than the people who were supposed to understand it - and who would not stop writing about it until the entire industry had to take notice.
SemiAnalysis is not a newsletter. It started as one, but calling it that now is like calling NVIDIA a graphics card company. The research firm has 85+ employees across 11 countries, 260,000+ subscribers ranked #2 in Technology on Substack, and is projected to surpass $100 million in annual revenue in 2026. The Information ran a major profile in April 2026 with the headline: "How Dylan Patel and SemiAnalysis Grabbed Sway in Silicon Valley."
The business model is unusual. The majority of revenue does not come from subscriptions. It comes from selling deep quarterly research reports and proprietary data products to the companies that move the most capital in technology: venture firms, hedge funds, hyperscalers, chip manufacturers, and government agencies. When you want to know how many H100s NVIDIA is actually shipping per quarter, or where the HBM bottleneck will crack, or whether Intel's foundry ambitions have any basis in reality - you pay SemiAnalysis.
The flagship data product is InferenceX, a benchmarking and rating system for AI companies and their datacenter infrastructure. At GTC 2026, Jensen Huang dedicated five minutes to walking the audience through InferenceX findings. That is not an advertisement. That is the world's most powerful chip executive using your research as a third-party validator of his company's performance claims.
People don't understand how fragmented the semiconductor supply chain really is and how many monopolies there are. Austria has two companies with super high market share and very specific technologies - and there is no alternative.
- Dylan PatelWhat makes SemiAnalysis different from every other tech research shop is the analytical method. Dylan Patel tracks semiconductor supply chains from the chemical suppliers in Austria all the way to the GPU sitting in your data center, accounting for every monopoly and bottleneck in between. He uses satellite imagery to track data center construction - watching billions in capital flow from orbit. He builds TCO (Total Cost of Ownership) models for AI cloud infrastructure that force executives to confront math they would rather not do.
When he published criticism of AMD's MI300X chip, AMD CEO Lisa Su personally arranged a 90-minute one-on-one meeting within 24 hours. That is not the typical response to a newsletter. That is a trillion-dollar company recognizing that one analyst with a blog had more credibility with their potential customers than their own marketing department.
Dylan Patel's three big bottlenecks to scaling AI compute: logic (TSMC capacity), memory (HBM), and power. Everyone in the industry focuses on GPUs. He focuses on what makes GPUs possible - and what will stop them.
His most contrarian call: by 2028 or 2029, the binding constraint on AI compute will not be TSMC, not NVIDIA, not HBM. It will be ASML - the Dutch company that manufactures the EUV lithography machines required to make every advanced chip. ASML can produce roughly 70 EUV tools per year now and may reach 100 by 2030. That is the ceiling for the entire global AI infrastructure buildout.
An H100 is worth more today than it was three years ago.
There's actually a humongous increase in GPU capacity. There's more GPU FLOPS shipping this year than Nvidia shipped their entire history.
NVIDIA has captured the compute and network layer, and is actively moving into the storage, software, and infrastructure operations layers over time.
By 2028 or 2029, the bottleneck falls to the lowest rung on the supply chain, which is ASML.
Memory vendors are going to double or triple price again as HBM demand escalates.
Amazon and Google are about to spend $200 billion per year on AI infrastructure. Google might have no profits in 2027.
There is a reliable test for whether someone actually matters in an industry: do the people running that industry change their behavior in response to them? By this measure, Dylan Patel matters more than almost anyone who writes about semiconductors.
After he published a critical analysis of AMD's MI300X GPU, AMD CEO Lisa Su personally arranged a 90-minute one-on-one meeting within 24 hours. This is a billion-dollar CEO clearing her calendar for a blogger from Georgia. The meeting happened because AMD's customers - the people AMD needs to convince - had read the analysis and were asking questions.
At NVIDIA's GTC 2026 conference, Jensen Huang did something unusual. In a keynote speech that could have cited analysts from Goldman Sachs, Morgan Stanley, or any established research institution, Huang singled out exactly two individuals by name. Dylan Patel was one of them. Huang spent five minutes explaining InferenceX findings to the audience - SemiAnalysis's own benchmarking product - as validation for NVIDIA's Blackwell performance claims.
Sam Altman has referred to him as "that SemiAnalysis guy." This is the way powerful people reference someone they are paying close attention to but would prefer not to amplify too directly.
Jim Cramer has cited his work on national television. The Information has profiled him. His "GPU-rich vs GPU-poor" framework from the "Gemini Eats the World" report is now standard terminology across the AI industry. When language enters a field through a single document, the author of that document has done something notable.
GPU-rich vs GPU-poor is now the vocabulary the entire AI industry uses to categorize itself. Dylan Patel wrote those words first.
- Industry observation, widely reportedThe model he has built is unusual enough to generate genuine controversy. SemiAnalysis blurs the line between media, research, and investment - a fact The Information flagged explicitly in its April 2026 profile. In early 2026, the firm was in early talks to raise hundreds of millions of dollars for a formal venture capital fund, which would make Patel not just the person who analyzes where chips go, but one of the people deciding where semiconductor investment goes next.
He has already personally invested in approximately 20 startups, including Thinking Machines Lab and Enfabrica. He personally led a $50 million Special Purpose Vehicle for Fluidstack's $700M fundraising round. The research that informs those decisions is the same research his clients pay for. This is the territory that generates lawsuits - and also the territory that generates outsized returns.
The criticism that follows SemiAnalysis is not about accuracy. It is about architecture. When your research firm also invests in the companies you write about, and when your personal SPV is embedded in deals where clients might be seeking your independent analysis, the question of influence becomes uncomfortable.
In January 2026, SemiAnalysis terminated employee Wei Zhou. The firm filed a lawsuit against him for breach of contract and trade secret misappropriation. Zhou counter-sued in April 2026, alleging wrongful termination. His complaint alleged he was fired for refusing to incorporate Material Non-Public Information from Fluidstack - the same company in which Patel had personally led a $50 million SPV - into paid research reports sold to clients.
SemiAnalysis denied these allegations directly. In a public statement on Twitter, Patel said Zhou was terminated for severe workplace misconduct - including coming to work drunk, making sexually explicit comments about a coworker, and repeated disruption of colleagues. The firm called the lawsuit "meritless."
The legal outcome remains pending. The underlying tension does not. As SemiAnalysis moves toward a formal venture capital fund, the question of whether its research can remain independent of its investment interests becomes more pointed, not less. The Information's framing was precise: SemiAnalysis has "blurred the lines between media, research, and investment."
This is not a unique problem. Every sell-side analyst on Wall Street operates in roughly this structure. The question is whether the audience understands the architecture - and whether the analysis holds up regardless. By most external accounts, it does. Jensen Huang citing your benchmarking platform is not the endorsement a captured analyst receives.
Dylan Patel does not have a filter. His Twitter/X account (@dylan522p) is a direct feed of what he actually thinks about NVIDIA's supply chain allocations, what China's EUV progress actually means, and why your assumptions about AI inference economics are probably wrong. He has 172,000+ followers there, not because he is diplomatic, but because he is specific.
The specificity is the point. He will tell you that Austria has two companies with near-monopoly market share in semiconductor chemicals, and that no chip at seven nanometers or below gets made without touching one of them, and that there is no alternative. This is not a fact that shows up in mainstream semiconductor coverage. It is the kind of fact that only someone who has traced the entire supply chain from raw material to finished GPU would know - and Patel has done that tracing compulsively, for years.
He uses satellite imagery to track data center construction. He builds models to project when Amazon and Google's AI capital expenditure will reach $200 billion per year. He wrote a report called "Claude Code is the Inflection Point" in February 2026 and noted that 4% of all public GitHub commits were being authored by Claude Code - a data point that had not appeared anywhere else before he published it.
His personality on podcasts is the same as his personality on Twitter. When Dwarkesh Patel (no relation) interviewed him in March 2026, Dylan walked through the three bottlenecks to AI compute scaling for over an hour without notes, without hedging, and without stopping. The confidence is not performance. It is what happens when someone has read the primary sources, built the models, and spoken to the engineers at every level of the supply chain.
He grew up working a motel desk in rural Georgia. His parents were immigrants who bet their family savings on a foreclosed property in a town most people have never heard of. He got rejected by the two universities that would have handed him credentials. He built the most important semiconductor research firm of the AI era on a Substack account he started on his birthday, for free.
There is a very specific kind of competence that comes from not having been given the door. You build the door. You analyze the door's supply chain. You write a 10,000-word report on the Austrian companies that supply the hinges. And then Jensen Huang names you at his keynote.