If you've fallen down the AI rabbit hole, you've probably come across SemiAnalysis, founded by Dylan Patel. He’s technology’s trusted source on semiconductors, GPU’s, and the infrastructure powering AI. So naturally, I asked him the internet's burning questions & he gets into: AI taking our jobs, data centers, AI doom, the "permanent underclass," and why he's actually optimistic about where all of this is headed. Q: Why shouldn’t we be scared of AI? A: So I think everyone's scared of AI because they're afraid that AI is going to take their jobs & cause this massive wealth divide. But really, AI is doing the opposite. Everyone who's a blue-collar worker was really screwed over in the last four or five decades in America. And now AI is making a lot more demand for electricians, plumbers, people in construction. The blue-collar worker's actually winning. It's people who have white-collar jobs who are scared of losing out because they have a "fake email job" much like myself. But I think they shouldn't be afraid of that. They should be excited that fellow Americans are going to get to make more money because they have actual things to build. Everyone's jobs also seem to be a lot easier with AI now. So I don't think people should be scared. I think people should be excited that everyone's life is easier, & they get paid more. Q: What's one piece of advice you have for somebody who wants to do what you do? A: I think the best way is just to follow your passions, right? Very cliche answer, but for a long time I was considered a weirdo. Still am, but I think for a long time, no one really understood why I was so obsessed with the stuff I'm obsessed with. And I think still people question it. But you should just do what you're obsessed with. Q: What is your most controversial take about all of this? A: Going back to your first question, why is everyone so pessimistic? I feel like everyone is very pessimistic about, oh, AI is going to take our jobs, or pollute the water, the air & all this sort of stuff because of people building data centers. But actually, I'm quite optimistic in that next year AI adoption is going to be huge & the economy's not going to lose jobs. We're actually going to be gaining jobs. Nominal GDP growth will be ~5%. And unemployment will be lower than it is today. And so despite all the fears, I actually think AI will lead to people being more fulfilled in their lives because they have jobs, the economy's growing, there's more abundance for everyone to split. All these people in San Francisco think if they don't work really hard & don't have fun, they'll escape the "permanent underclass", which is a really stupid concept. Everyone will be fine, things will be great & humanity gets better every generation. 📌 Follow along for more. 📨 Where tech, money, & culture meet: jaiyagill.com #AI #Technology #ArtificialIntelligence #Semiconductors #FutureOfWork
The third pressure point is Dylan’s singular authority versus the actual collective machinery. Patel recently described SemiAnalysis as having roughly 90 people, including engineers across the supply chain and former hedge-fund personnel. Yet the public identity still centers heavily on Dylan as founder, CEO, chief analyst, and authoritative interpreter of nearly every layer from process chemistry through model economics. The reasonable audit question is which conclusions are Dylan’s work, which come from named specialists, which are produced by proprietary models, and which depend on unattributed sources. That does not show that the research is weak. It tests whether the heroic-founder presentation accurately represents the epistemic production system behind it.
The second pressure point is independence versus embeddedness. SemiAnalysis calls itself an independent research company, but it also sells retained advisory engagements, bespoke projects, proprietary industry models, and hourly consulting. Its stated clients include hyperscalers, semiconductor companies, and public- and private-market investors. That access flywheel is a legitimate moat, but it is also a dependency: the companies being analyzed may simultaneously be clients, sources, collaborators, benchmark participants, or beneficiaries of the resulting narrative. SemiAnalysis has installed trading restrictions, MNPI procedures, and supervisory review, which indicates that it recognizes the institutional problem. The sharper question is not whether it has compliance paperwork. It is whether an outsider can distinguish reproducible analysis from privileged proximity and ecosystem consensus.
What he appears to be running from is therefore category collapse. He owns the forum-moderator and shitposter history when it functions as an origin myth. What threatens the finished identity is the suggestion that SemiAnalysis never fully escaped that category, but merely monetized it through access, staffing, presentation, and favorable timing. Patel himself acknowledges that everything “blew up all together” at the right moment, while also presenting the company as the product of superior foresight. The sensitive question is how much of the outcome represents uniquely calibrated analysis and how much represents being an obsessive, talented commentator positioned directly in front of the largest capital-expenditure cycle in modern technology.
That is the seam underneath the preciousness. SemiAnalysis wants the cultural permission of a disruptive shitposter and the evidentiary deference of a regulated research institution. The control question that forces those identities to reconcile is: What observable result would make SemiAnalysis conclude that the AI infrastructure cycle had been materially overbuilt, and where did the firm publish that falsification threshold before the result arrived?
This also explains why ordinary mockery may produce less leverage than expected. Patel openly describes himself as a former forum moderator, says he enjoys argumentative environments, and invokes the idea that wrestling with him is pointless because he enjoys the fight. Calling him arrogant or wrong simply returns him to familiar terrain, where speed, confidence, and combativeness are advantages. The harder move is a calm institutional audit that refuses the forum contest altogether: dated predictions, confidence intervals, revisions, misses, methodology changes, source boundaries, client overlap, and author-level attribution. I found extensive product descriptions and a compliance page, but I did not find a single public, consolidated calibration ledger covering the firm’s major forecasts.
A short and sweet interview. Gets straight to the point..thanks!
Such a great interview!!!
It's interesting how many of the worries about AI parallel historical fears around automation, yet the shift in demand for blue-collar skills is a genuinely new wrinkle. The "fake email job" observation feels pretty accurate for a lot of us who spend our days in tools. It really makes you think about what true value AI unlocks for those who build things.