SemiAnalysis founder Dylan Patel explains why he raised a $500 million solo venture fund, bought the AI-bearish Citrini Research, and thinks Meta’s Muse beats OpenAI on product despite a weaker model. In a conversation with Molly O’Shea on Sourcery, recorded November 11, 2026, Patel also covers data center permits pushing gigawatts of capacity into 2028, where his firm spends $15 million a year on AI, and why he calls Gartner “a racket.”
TLDW
Patel says permits, not labor or equipment, are pushing data centers back, and that a SemiAnalysis poll found 49% of people in counties with a 100 megawatt data center did not know it existed. He argues applications now matter as much as models: Meta’s Muse and Instinct are better products than OpenAI’s Dots, and 45% of SemiAnalysis’s $15 million AI budget now goes to persistent agents, mostly through Perplexity Computer. He explains the $500 million seed and Series A fund focused on hardware and infrastructure, the Citrini acquisition, and why SemiAnalysis never sells opinions, even when it means criticizing its biggest customers. He closes by predicting an explosion of hardware startups.
Thoughts
Patel’s whole business rests on one asset, and this interview is mostly about how he plans to protect it while adding new conflicts. SemiAnalysis sells data and consulting to hyperscalers, chip companies and AI labs, and he says those clients only pay because the firm is “an objective truth-seeking player.” He calls Gartner’s model “a racket” and says SemiAnalysis has never taken ads or paid opinions. His evidence is specific: the firm has called some of Nvidia’s treatment of neoclouds anti-competitive, called IREN’s cloud technology unreliable while praising its powered land, and publicly said Fluidstack’s data centers are delayed even though he has a large investment in it. Now he runs a research firm, a consulting arm, a newsletter, a separate bearish research shop and a $500 million venture fund at the same time. His answer is firewalls and compliance spending. Readers should hold him to the Fluidstack standard: the test is whether SemiAnalysis keeps criticizing companies the fund owns.
Buying Citrini Research is the most interesting strategic move he describes. Patel is “super bullish AI in the long term,” while Citrini turned bearish on AI months ago, and its February paper shook the markets. He says Citrini staff asked whether they were still allowed to be bearish after the deal, and he told them to “do whatever you want.” Owning a bear is a smart hedge for a research brand built on credibility. If AI sentiment turns, the group has a well-known voice on the other side, and the deal makes his objectivity claim more believable than any statement could. The two firms also do research differently. SemiAnalysis counts cables, servers and fab tools, while Citrini sent someone to the Strait of Hormuz to see whether oil would stop shipping and built an AI-designed scorpion antivenom to understand AI drug design.
His product critique matters because it comes with spending data. Over the last few months, SemiAnalysis went from no persistent agents to 45% of a $15 million annual AI budget, and most of that runs through Perplexity Computer, which uses Claude and OpenAI models underneath. He says it is simply a better harness, and that OpenAI’s Guardian safety layer makes Dots and Codex refuse things “overzealously.” Instinct, he says, mostly uses open models and still beats OpenAI on product. The implication for the labs is uncomfortable: the best model does not guarantee the customer relationship. He expects coding tools to be easy to switch, but enterprise deals to stick, citing Anthropic contracts that offer discounts if 70% or more of a company’s AI spend goes to Anthropic.
The data center section offers a grounded view of the backlash. SemiAnalysis has pushed multiple gigawatts of capacity from 2027 into 2028, and more out of 2028, not because of labor or equipment shortages but because permits are not available. Patel notes that most jurisdictions banning data centers have none, that data centers were not a top issue in election exit polls, and that his firm’s poll with David Shor found most people near big projects did not know about them. He thinks the opposition is about AI more than buildings, and that companies can buy goodwill cheaply, for example by subsidizing power for everyone within 30 miles. His point that all of Elon Musk’s data centers use less water than three In-N-Out restaurants, counting the cattle supply chain, will get quoted. His more useful point is that the real local harms are noise and tax breaks for facilities that employ almost no one once they are built.
His closing prediction ties the fund to the research. Patel expects hardware to “accelerate like crazy,” with the number of companies building hardware, and building it well, set to skyrocket within a year. That is why the fund is aimed at robotics, silicon, optics, networking and manufacturing rather than apps. He says he knows nothing about scaling a consumer app to a billion users, but has staff from ASML, Applied Materials, TSMC, Intel, Nvidia and the hyperscalers to judge hardware deals. His new GDPX benchmark, which tests models on cyber, chip design, materials science, bio and quant work, is the bridge: if models get good at those tasks, the hardware startups he is funding get much faster engineering.
Key Takeaways
- SemiAnalysis and Citrini together employ about 115 people in 15 countries, with offices in New York, Chicago, San Francisco, Oregon, Singapore and Korea.
- Patel expects some gigawatt-scale data center capacity in space this decade, perhaps around 2030, but nothing meaningful in the next three or four years.
- SemiAnalysis has pushed multiple gigawatts of data center capacity from 2027 into 2028 because permits are not available.
- In a poll with David Shor, 49% of people in counties with a 100 megawatt data center did not know there was one.
- Patel says Meta’s Muse and Instinct are better products than OpenAI’s Dots despite weaker models, partly because OpenAI’s Guardian is overzealous.
- 45% of SemiAnalysis’s $15 million annual AI spend now goes to persistent agents, mostly through Perplexity Computer.
- He has seen Anthropic offer enterprise discounts if 70% or more of a customer’s AI spend goes to Anthropic.
- He raised a $500 million solo GP fund for seed and Series A deals in robotics, silicon, optics, networking, manufacturing and infrastructure software.
- He avoids app investing because he does not know how to scale a consumer app, and says hardware needs both a great founder and a great idea.
- Citrini remains a separate company free to stay bearish on AI, with its director of research controlling what gets published.
- SemiAnalysis never takes ads or paid opinions, and Patel says its biggest customers are often the companies it criticizes most.
- He calls Gartner “a racket,” pointing to a ranking that put IBM’s coding agent above Cursor or Claude Code.
- SemiAnalysis is testing inference chips from Huawei, Amazon and startups, and launched GDPX, a benchmark spanning cyber, chip design, materials, bio, quant and recursive self-improvement.
Chapters
02:17 Data Centers in Space and Texas Permits
Patel takes the over on Elon Musk’s 2028 plans for data centers in space and expects a gigawatt in orbit sometime this decade. In Texas, he says the governor has signaled she will not stop data centers if they avoid raising power prices and use water responsibly, but a union labor requirement would be unpopular, and water, wastewater and air permits give the state easy ways to block projects.
06:42 Who’s Really Against Data Centers
Most jurisdictions that ban data centers have none, and SemiAnalysis’s poll found half of residents near large ones did not know about them. Patel blames online narratives, some astroturfed, and a bestselling book that overstated data center water use by about 1,000x. He says Elon Musk’s data centers use less water than three In-N-Outs, and that the real issues are noise and tax breaks for facilities that employ few people.
12:04 Anthropic vs OpenAI, and Why Meta’s Muse Wins on Product
Patel says Anthropic crushed it in Q2, OpenAI in Q3, and Anthropic is back on top with Opus 5.5 and Fable 5.2, while markets smooth out the swings. He calls Muse and Instinct far better products than OpenAI’s Dots, mocks the ChatGPT app’s confusing model options, and says Meta’s Alexandr Wang let him beta test Muse.
15:55 Where SemiAnalysis Spends $15 Million on AI
Most of the firm’s AI spend went to Claude for most of the year, but persistent agents now take 45%, largely through Perplexity Computer rather than the labs’ own agent products. He expects consumer products and enterprise contracts to be sticky, and coding tools to be easy to switch.
20:38 Is There Real ROI in AI Startups?
Most startups will fail, Patel says, but companies like Instinct, Sierra and Harvey have staying power and fast-growing revenue. He questions whether some inference providers deserve their valuations and notes that some application companies are pure pass-throughs with volatile margins.
22:30 Inside the $500 Million Fund
Patel lists his goals as having fun, making money, knowing everything, influencing what matters and spending time with family. The $500 million fund invests at seed and Series A in hardware and infrastructure, kept separate from the research business. He will not disclose check sizes, avoids apps, and says most pitches are “boring and wrong.”
30:30 Buying Citrini, the AI Bears
Patel says Citrini has been right many times and does research very differently, from sending an analyst to the Strait of Hormuz to building an AI-designed scorpion antivenom. He wants to help it run better as a company without changing its voice, and says it remains separate and free to be bearish.
34:03 Why Gartner Is a Racket, and Staying Objective
Patel says information is more valuable than ever while firms like Gartner and IDC are useless, and that legacy media brands have slipped. SemiAnalysis never sells opinions, criticizes its biggest customers, and has had clients threaten to pull contracts only for their teams to insist on keeping the data. He says the fund will never use the research platform to promote portfolio companies.
41:07 The Hardest Lessons of Running SemiAnalysis
Running a company means lost sleep, missed moments and constant tough calls, from customer negotiations to firing people and being sued. Patel says many investors have never made decisions like that. He also jokes about knowing just enough Chinese to make people laugh.
46:32 Why Hardware Is About to Explode
Patel is looking forward to SemiAnalysis’s inference tests of Huawei, Amazon and startup chips, and to models improving on GDPX. He predicts the number of hardware companies, and of good ones, will skyrocket within a year, along with advances in robotics, packaging, networking and memory.
Notable Quotes
“49% of people in counties that they have a 100 megawatt data center did not even know there is a data center in their county.”
Dylan Patel, on SemiAnalysis’s poll with David Shor
“All of Elon Musk’s data centers use less water than three in-n-outs.”
Dylan Patel, on data center water use
“Instinct and Muse despite being worse models than OpenAI’s models are way better products than OpenAI dots.”
Dylan Patel, on why product beats model quality
“I’m still super bullish AI in the long term.”
Dylan Patel, on owning the AI-bearish Citrini Research
“If you look at like Gartner, the way they monetize it is they are like a racket.”
Dylan Patel, on paid analyst rankings
“We criticize the things that we love the most.”
Dylan Patel, on staying objective about companies he invests in
“I think hardware is going to just accelerate like crazy.”
Dylan Patel, on what he expects in the next 12 months
Watch the full conversation here.
Related Reading
- SemiAnalysis Patel’s research firm covering chips, data centers and AI infrastructure.
- Citrini Research the AI-bearish research shop SemiAnalysis acquired.
- Magic Quadrant (Wikipedia) background on the Gartner rankings Patel criticizes.
- Data center (Wikipedia) background on power, cooling and water use in the facilities behind AI.