Dylan Patel, SemiAnalysis
Dylan Patel argues data center risk is now about permits, not labor or equipment, that AI products matter more than models, and that an objective research firm must stay independent of its customers and investments.
Watch on YouTube ↗Charlie's summary
- ▶ 3:35
Data center delays now come from permits, not labor or equipment, and Texas could stop projects through air and water permits
Dylan says SemiAnalysis tracks every data center in the world and has pushed multiple gigawatts out of 2027 into 2028, with more slipping within 2028. The cause is “permits are just not available.” Texas's governor has levers on water, wastewater, storm water and air permits, and can block even grid-independent sites. The union labor question is the other sketchy point.
- ▶ 6:10
Most people don't know the data centers near them exist, so the backlash is a vague anti-AI feeling
Dylan says a poll of counties with a 100 megawatt-plus data center found 49% of residents did not know one was in their county, and only 40% knew one was coming. He thinks people hold a nebulous bad view of data centers, boosted by astroturfed content and inflated claims, such as a bestseller overstating water use by a thousandx. Data centers are an easy target for dislike of AI. Dylan thinks companies could defuse it, for example by subsidizing power for everyone within 30 miles.
- ▶ 9:12
Water fears are overblown when data centers are built properly, but a weak local return fuels resentment
Dylan says all of Elon Musk's data centers use less water than three In-N-Outs, counting embedded water in the supply chain. Agriculture wastes far more water, but nobody protests it. The real grievance is when a data center gets a property tax break, may raise grid prices and employs almost no one after construction, giving a boom then bust. Done properly, though, the benefits can be large.
- ▶ 14:20
Products, not model quality, are deciding who wins, and OpenAI and Anthropic are bad at product
Dylan thinks Meta's Muse and a rival app are better products than OpenAI's despite weaker models, partly because OpenAI's safety guardrails overreach. He cites OpenAI's cluttered chat app with its many model toggles. SemiAnalysis went from zero to 45% of its $15 million annual AI spend on persistent agents, mostly through Perplexity Computer rather than the labs' own tools. Dylan expects consumer products to be sticky, coding tools less so, and enterprise contracts stickiest, such as discounts for 70% of spend.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.