Dylan Patel on the AI Market Shakeout & Building SemiAnalysis
- 01Data center permitting, not labor or equipment, is the binding constraint on AI buildout
- 02The anti-data-center backlash is largely a perception problem, not a local-experience problem
- 03Space data centers are real but too late to matter for the current cycle
- 04Product quality and harness, not raw model quality, is becoming the competitive moat
- 05Enterprise contract design is where stickiness is actually being engineered
- 06Valuations move more slowly than the news cycle suggests
1. Key Themes
Data center permitting, not labor or equipment, is the binding constraint on AI buildout
Dylan Patel, who tracks every data center globally, says he has already pushed out capacity timelines because of permits. The risk is regulatory, not supply-chain. "We pushed back, you know, multiple gigawatts out of 2027 into 2028. And in 2028, there's quite a few gigawatts that are getting pushed out, not because they can't build it, not because the labor is not there, not because of some like shortage of equipment, but because, you know, permits are just not available." 00:05:46 He explains why a governor has leverage even without a ban: "There's water permits, wastewater permits... And there's like air permits. And those three things she can control and stop very easily... There's data centers that aren't even connected to the grid that use very little water. But she can still stop them via the air permit." 00:04:54
The anti-data-center backlash is largely a perception problem, not a local-experience problem
SemiAnalysis polled counties with large data centers and found ignorance, not opposition, is the norm. "49% of people in counties that they have 100 megawatt data center did not even know there is a data center in their county." 00:00:01 On the cause: "People hate, dislike AI more than they dislike your data centers, but then data centers are easy way to point at it." 00:11:57 He adds that misinformation compounds it: "One of the books in the New York times bestsellers overestimated the water usage of data centers by like a thousand X. And then they like revised it, but like no one pays attention." 00:11:42 His proposed fix is cheap: "You could just subsidize the power of everyone in a 30 mile radius and it costs like nothing." 00:08:34
Space data centers are real but too late to matter for the current cycle
Patel puts a gigawatt in orbit at roughly the end of the decade and says it solves nothing near-term. "I think space is like a solution still two years away... I know Elon's trying to do a shitload in 2028 but I'll take the over on that. But probably like at some point in this decade it could be like there's some, you know, like a gigawatt in space." 00:03:00 And: "So that doesn't solve anything in the next like three or four years. Or it's not a meaningful amount of the capacity." 00:03:30
Product quality and harness, not raw model quality, is becoming the competitive moat
The frontier-lab leaderboard shifts quarter to quarter, but Patel argues applications and workflow stickiness will decide durable winners. "Instinct doesn't have the best model. Um, they're primarily using open models. Um, and yet the, and, and yet OpenAI is terrible versus them. Right. So I think product is very important." 00:16:24 His own firm's usage proves the point: it went "from, you know, zero persistent agents to 45% of our $15 million a year AI spend being persistent agents... through, um, perplexity computer, which on the backend is still using Claude and OpenAI is just a better harness." 00:15:55
Enterprise contract design is where stickiness is actually being engineered
While coding tools are easy to swap, labs are locking in enterprises through discount thresholds. "Developers can switch easy. Um, but enterprise contracts will be sticky and there's like contracts floating around where Anthropics like, oh, if 70% of your AI spend is with us, we'll give you this discount. And if it's not, you don't get that discount." 00:16:52
Valuations move more slowly than the news cycle suggests
Patel pushes back on the "who's top dog this quarter" framing: "The market doesn't actually shift the valuations that much... it kind of, uh, smooths in the curve." 00:12:53 His read of the quarter-by-quarter swings: "We're going to watch this and over index on any individual data point. Um, that is just like human nature to over index on data points." 00:13:19
Most AI startups will fail, but application companies with real revenue have ROI
Patel is measured rather than bearish on the startup layer: "Most startups will fail. That's just, like, a fact of life. But some of them will be successful as well." 00:20:54 He singles out those with fast growth: "Companies like Sierra and Harvey, like, they're growing revenue really fast... I think, like, those companies have ROI." 00:20:54 He flags margin risk for pure pass-through plays: "Some of them are just, like, pure pass-through, though, right? Like, I don't know if you saw Harvey's, like, blogs. They, like, showed their margins of just seesawed all over the place." 00:21:44
Objectivity is the product: refusing paid opinion is how SemiAnalysis preserves its value
Patel explains that the data and consulting business only works because the opinions can't be bought. "We've never done ads. We've never done, you know, you pay us for an opinion... And so our biggest customers are the ones we hate on the most, right?" 00:35:52 On why it works commercially: "We've had people try and pull their contract because we, you know, they were mad at what we said. But then their team was like, well, no, actually, we still want their data." 00:36:50
A radical-honesty investor model: invest where your firm has unfair expertise
His $500M SoloGP seed/Series A fund stays inside infrastructure because that is where SemiAnalysis's bench is deep. "I don't know shit about applications. So why would I invest in applications, right? Like, I don't know how to scale a fucking app to a billion active users." 00:25:46 Instead: "I have people who have worked at ASML and applied materials... people who did R&D at TSMC and Intel... people who have built power plants. It's, like, I have people who have done all of these things in my company." 00:26:16
Models getting good at hard technical domains will accelerate hardware innovation
Patel's 12-month view centers on the new GDPX benchmark and what capable models unlock. "We launched a new benchmark called GDPX. It tests across many domains, cyber, chip design, hardware design, material science, bio, quant, RSI." 00:46:42 The downstream thesis: "I think hardware is going to just accelerate like crazy... the number of companies making things in the hardware realm is going to skyrocket. And the number of companies that can do it well is going to skyrocket." 00:47:08
2. Contrarian Perspectives
Data center backlash is mostly narrative, not community experience
Conventional wisdom says local communities are rising against data centers. Patel's polling shows most people near large ones don't know they exist. "49% of people in counties that they have 100 megawatt data center did not even know there is a data center in their county... only 40% of people knew a data center was even coming where there was a hundred megawatt data center coming in the next three years." 00:00:01 He also notes data centers didn't register at the ballot box: "If you look at all the exit polls, like data centers was not like any of the top issues... it was like inflation and war in Iran." 00:07:13
Data centers use less water than the burger chain next door
Against the "AI is draining water" meme, Patel claims the whole Elon Musk fleet is a rounding error on an agricultural supply chain. "All of Elon Musk's data centers use less water than three in and outs." 00:09:15 He defends the methodology, counting embedded water on both sides: "Even if you do the embedded water usage of the data center versus the embedded water usage of in and out, it's three in and outs is equal to all of the data center capacity Elon Musk has." 00:09:50
Frontier labs are bad at product, and that opens room for outsiders
Most commentary treats OpenAI and Anthropic as default winners. Patel says the leading labs' products are weak and that incumbents like Meta execute better. "Have you seen the OpenAI chat app? It's so dumb... This is a billion user fucking app." 00:17:22 He credits Meta's new product: "Muse is actually intuitive... it was like clear and good UI. And it's like, this is not the case for open AI and Anthropics." 00:18:18 His explanation: "It's the cycle of like higher meta VP that wasn't like actually that important at meta, you know, create a bunch of garbage, hire new meta VP." 00:18:44
Coding agents are the least sticky AI category, while "safe" products may be too safe
The consensus treats Claude Code and Codex as entrenched. Patel disagrees: "I think like Claude code and codecs are not going to be that sticky because developers can switch easy." 00:16:52 He also argues that over-caution hurts: "Good products need to be good products before, you know, and safety is like important. But like... if it like texts someone incorrectly or like books an appointment incorrectly, like fine, whatever." 00:15:09
Investors who can't operate shouldn't pretend to know everything
Patel calls out a cultural flaw in venture: "One of the most toxic traits about whether it be San Francisco or New York investors is, like, they think they know everything about something that, like... they don't know anything about." 00:27:44 He adds that most have never lived the hard decisions: "A lot of investors have never made tough decisions... Have you fired people? That fucking sucks. Have you been sued by someone you fired?" 00:42:33
3. Companies Identified
SemiAnalysis
Patel's research, data, and consulting firm tracking every data center, chip, and supply chain input globally. Mentioned as his own platform and the foundation for the fund. "We've got people in 15 countries, there's about 115 people, offices in New York, Chicago, SF, Oregon, Singapore, Korea. You know, we track every data center in the world and we as a firm know more about data centers than anyone else." 00:00:01 Its customers: "My biggest customers are hyperscalers, semiconductor companies, and AI labs." 00:00:01
Citrini
Macro and thematic research shop acquired by Patel and kept as a separate company; known for its bearish-AI and "Citrini paper" market-moving work. "They have captured the market's imagination. They've been right a shitload of times. And their way of doing research is very different than the way of semi-analysis's research is done." 00:30:40 On its depth: "They wanted to know what's going on with AI drug design. So they went and developed a scorpion antivenom with AI... or they wanted to know if oil was going to stop being shipped. So they sent someone to the straight of Hormuz." 00:32:59
Instinct
AI consumer product that Patel says is gaining real users with open models. "Seems like Instinct has some staying power. They've gotten a lot of users. Now, I don't know if it's worth, like, the valuation they've got, but they're fucking kicking ass, right?" 00:20:54 And: "Instinct doesn't have the best model. Um, they're primarily using open models. Um, and yet the, and, and yet OpenAI is terrible versus them." 00:16:24
Meta (Muse)
Meta's new agent product (referenced as Muse), which Patel beta tested and praised for UI. "I was able to beta test Muse. Um, Alex like texted me... So I tried it out and I was like, fuck, this is pretty good. And it was like clear and good UI." 00:18:18 He expects stickiness: "Muse is probably just going to be a lot more sticky." 00:16:24
Perplexity (Perplexity Computer)
Now the largest slice of SemiAnalysis's agent spend. "45% of our $15 million a year AI spend being persistent agents. And that 45%, the vast majority of it is not through, you know, Anthropics, Claude tags or OpenAI's... It's through, um, perplexity computer, which on the backend is still using Claude and OpenAI is just a better harness." 00:15:55 Patel is unsure of its margins: "I don't know what perplexity's margin is on that. Yeah. I don't think it's that high." 00:22:09
Anthropic
Frontier lab Patel says has crushed it in Q2 and again recently, with sticky enterprise deals. "Obviously Anthropic crushed it in Q2... Now they're back to crushing it because of Opus 5.5 and Fable 5.2." 00:12:53 Also noted for its tiered-discount enterprise contracts: "If 70% of your AI spend is with us, we'll give you this discount." 00:16:52
OpenAI
Frontier lab that "crushed it in Q3" but is criticized on product design and delayed releases. "OpenAI keeps delaying releases of models." 00:13:19 And: "OpenAI's product sucks in many cases." 00:37:47
Harvey
Legal AI company, cited as a positive example of an application with ROI and fast revenue growth. "You look at, like, companies like Sierra and Harvey, like, they're growing revenue really fast." 00:20:54 Patel also notes its margin volatility from public blogs.
Sierra
Customer-service AI agent company, grouped with Harvey as a fast-growing application. "Maybe their customer acquisition cost is really high, but, like, you know, they're solid and they're in there, right?" 00:21:21
IREN (Iris Energy)
Neocloud/data center operator. Patel makes the distinction between weak cloud technology and valuable powered land. "We've said Iris Energy, IREN, has terrible technology on the cloud services side. You know, their clusters can be unreliable... At the same time, we've said, oh, but they have powered land, and that's worth a lot." 00:38:46
Fluidstack (transcribed "Foodstack")
Neocloud in which Patel holds a large personal investment while publicly criticizing its data center delays. "I have an investment, large investment in Foodstack. And yet, we've still said their data centers are delayed publicly." 00:38:46
NVIDIA
Dominant AI chip vendor; SemiAnalysis has publicly raised concerns about its conduct. "We've said things like, oh, we think some of NVIDIA's practices are anti-competitive. Right? Like, you know, with regards to how they treat certain neoclouds." 00:37:47
Tesla
Used by Molly as a product-discipline analogy. "Tesla at least will take things off the production line. Why don't the models do that?" 00:18:11
Cerebras
Mentioned among the year's AI IPOs. "We had Cerebrus earlier this year." 00:12:04
SpaceX / xAI
Referenced as having IPO'd and become an AI company; Elon Musk's data centers and space ambitions also discussed. "SpaceX had a big one earlier this year and they are now an AI company." 00:12:04
Noted as returning to prominence in the rotation of leading AI companies. "Google got back in." 00:13:19
Huawei, Amazon (custom chips), and AI startup chip makers
Patel plans to test their inference performance. "There's testing of various chips, including Huawei chips, Amazon's chips, various AI startups chips on inference performance." 00:46:42
Gartner and IDC
Cited as negative examples of legacy research. "Everyone recognizes, like, the Gartners and IDCs and those kinds of companies just suck and are useless. And yet information is more valuable than it's ever been." 00:34:13 On pay-to-play dynamics: "They put IBM, like, whatever the fuck IBM's coding agent is above, like, I think they put it above Cursor or Cloud Code." 00:35:41
IBM
Used as an example of Gartner's questionable rankings. "How the fuck is IBM's coding agent better?" 00:35:52
Cursor and Claude Code
Cited as the benchmark coding agents in the Gartner critique and as SemiAnalysis's primary tool for much of the year. "The majority being Claude code, um, for most of the year." 00:15:32
ASML, Applied Materials, TSMC, Intel
Named as employers whose alumni staff SemiAnalysis. "I have people who have worked at ASML and applied materials land research... I have people who did R&D at TSMC and Intel." 00:26:16
Google, Microsoft, Amazon
Named as employers of SemiAnalysis's infrastructure experts. "People who have done infrastructure at Google, Microsoft, Amazon." 00:26:46
Altimeter
Fund of Apoorv Agrawal's guest-episode reference. "I recently had a pour of Agrawal on the pod... He has, he has, yeah. Altimeter." 00:12:04
Brex, Harvey, Zone, Mercor, VCX by Fundrise, Public, Deel
Podcast sponsors named by Molly. Zone is described as "developing next-generation data center campuses" 00:20:17; Brex as "the intelligent finance platform that combines cards, expenses, and banking" 00:19:20.
Wall Street Journal, Bloomberg, New York Times, Substack, YouTube, X, Instagram
Cited in Patel's discussion of media and attention. "There's still awesome reporters at, like, places like Wall Street Journal or Bloomberg or New York Times. But those brands have definitely taken a huge hit." 00:34:43
4. People Identified
Dylan Patel
Founder of SemiAnalysis and Citrini's owner, and a solo GP raising a $500M early-stage fund. Subject of the episode. "I raised $500 million, SoloGP. You know, we're focusing on Seed Series A." 00:23:56
Molly O'Shea
Host of Sorcery and a former VC. "I come from the VC world." 00:01:04
Elon Musk
Referenced for space data centers and data center water efficiency. "I know Elon's trying to do a shitload in 2028 but I'll take the over on that." 00:03:00
Governor Greg Abbott (transcribed as "Hinojosa")
The Texas governor discussed as holding permit leverage over data centers; Patel says she "has been pretty positive and constructive." 00:03:56 Her key levers: "Water permits, wastewater permits, stormwater... and there's like air permits." 00:04:54
James (Citrini)
Citrini's leader, praised by Patel. "James is awesome. He's, you know, he's awesome. And the whole Citrini team is awesome." 00:30:40
Kyle (Citrini)
Citrini's director of research, who controls publishing decisions. "Kyle, who's their director of research, is, like, got the most sort of voice on what gets published and not." 00:31:34
Alex (Meta)
Reached out to Patel to beta test Muse. "Alex like texted me. He's like, Hey, try this new product out." 00:18:18
David Shor
Pollster who collaborated with SemiAnalysis on the data center poll. "We did this poll with David Shore and he collaborated with our team." 00:06:43
Apoorv Agrawal
Altimeter investor with the "four seasons of AI" framework (Google, Anthropic, OpenAI, SpaceX). "He has this like whole thing on the four seasons of AI." 00:12:34
Dina Powell
Referenced for a story on stage at the All-In Summit about a county that funds its education from data center revenue. "A city council leader and they now can fund all the education in their area because of the benefits and the revenue." 00:08:47
5. Operating Insights
Separate the revenue engine from the opinion engine with real compliance walls
Patel runs consulting and data sales alongside public research and a venture fund, but keeps them firewalled. "That stuff is sort of firewalled away from the research." 00:23:56 And: "The fund is a completely separate entity and company than SemiAnalysis and Citrini. So, they will never say anything to pump a company that is, like, on the venture side." 00:39:58 On compliance: "You spend a lot of money on compliance. You can't, like, fucking, like, just, like, willy-nilly do stuff." 00:40:23
Buy a complementary franchise and leave its voice intact
When he acquired Citrini, he explicitly preserved independence and editorial control. "They were, like, oh, are we allowed to be bearish? I'm, like, dude, do whatever you want. Like, I'm not here to change your voice. I'm here to help you run this as a better company." 00:31:34 The value-add was operational: "Let's operationalize it. Let's help them be better." 00:31:09
Keep your own buying decisions flexible and measure AI spend by workflow, not vendor
Patel's firm seesawed vendors all year and was willing to move 45% of spend to a better harness. "Throughout the year, our spend is like seesawed all over the place in terms of what we're using." 00:15:32 "We're a small nimble organization that moves fast." 00:16:24
Stay opaque on fund details to preserve optionality
Asked for check size, he declined. "Why would I tell people what I'm doing? That removes optionality for me to do something different." 00:28:52 On sourcing: "There's 115 people work for me. If they see something interesting, they'll refer it." 00:29:38
Use your platform to help portfolio companies without pumping them
"There's many ways to help them, right? I have 115 people... If this company needs help, you know, maybe we sell them research. Maybe we do consulting. If they're super early-stage, we can invest and give them some help pro bono." 00:24:25
6. Overlooked Insights
Voluntary tax-credit mechanics may be quietly turning data centers into local government windfalls
In a brief aside, Patel describes a Louisiana county where property tax rules forced the revenue to flow to government employees. "There's a county in Louisiana who got massive, uh, they paid out all their teachers like 20 hay... There's a rule that you, your property tax has to be, you know, all the tax you collect, like they just kind of have to credit it back to, uh, government employees." 00:09:08 This points to a replicable playbook: structured community-benefit mechanisms can convert data center revenue directly into constituent goodwill, defusing the backlash he describes elsewhere.
Patel treats frontier model "harnesses" as a separate, investable layer from the models themselves
Mentioned in passing, but important: his firm routes 45% of a $15M AI budget through a company that "on the backend is still using Claude and OpenAI is just a better harness." 00:15:55 If a harness layer can capture that much enterprise spend while renting the underlying model, it implies a middle layer between labs and end users with real pricing power, yet Patel also admits he doesn't know its margin ("I don't think it's that high" 00:22:09), which is exactly the open question for investors in that layer.