Teahose.
SIGN IN
NEW HERE — WHAT TEAHOSE DOES
We read the entire AI & tech firehose — so you don't have to.
PODPodcastsAll-In, No Priors, Acquired…
NEWNewslettersStratechery, Newcomer…
PAPPapersPhysical AI research
PHProduct Huntdaily launches
VCInvestor ScoutSequoia, a16z, Benchmark…
CLAUDE DISTILLS →
7 reads, 30 sec each — free, 6 AM ET.
+ a live graph of the companies, people & themes underneath.
HOME/SOURCERY/Dylan Patel on NVIDIA's Future,…
POD
// EPISODE
SOURCERY

Dylan Patel on NVIDIA's Future, SemiAnalysis & His $500M VC Fund

DATE November 16, 2026SOURCE SOURCERYPARTICIPANTS DYLAN PATEL, MOLLY O'SHEA
// KEY TAKEAWAYS6 ITEMS
  1. 01Jensen is a single point of failure for NVIDIA
  2. 02Export-control fallout will be regulatory, slow, and probably aimed at renting rather than selling
  3. 03The Jensen scandal creates an opening for AMD and other vendors
  4. 04Data center permitting, not labor or equipment, is now the binding constraint
  5. 05The data center backlash is largely perception, not experience
  6. 06Product quality is becoming the differentiator, and the labs' product teams are weak
In this episode

1. Key Themes

Jensen is a single point of failure for NVIDIA

Patel argues NVIDIA has no clear succession structure, so a worst-case outcome for Jensen would be existential for the company, not just a headline. The ecosystem-propping role Jensen plays (allocations, investments, backstops) is also at risk from distraction alone.

Dylan Patel: "Jensen is like no longer the CEO. Nvidia is like fucked in my opinion, if that's ever happens. Um, cause like there is no like clear, you know, number two." 00:04:04 He adds: "Jensen has like 50 reports or whatever. So it's like kind of hard." 00:04:32

Export-control fallout will be regulatory, slow, and probably aimed at renting rather than selling

Patel expects investigations to take months, with the real consequences being tighter restrictions, congressional grilling, and a likely pivot toward rules governing who may rent chips to Chinese firms. He sees chips as already widely leaking into China.

Dylan Patel: "We have estimates on how many it's on the order of hundreds of thousands, right? It's not a small number of GPUs that have made it to China." 00:09:01 On the policy direction: "Maybe it's better to not ship chips to China and have them rented from the U S ... Maybe you want to make the rule where it's like only U S companies can rent chips or like hyperscalers." 00:09:52

The Jensen scandal creates an opening for AMD and other vendors

If NVIDIA is toxic for selling into Southeast Asian data centers that serve Chinese clouds, competitors can take that business with little scrutiny.

Dylan Patel: "AMD has free reign to sell to companies that are in Singapore that are renting to Chinese, you know, clouds ... no one's really going to give a shit, but they're going to give a lot of fucks if NVIDIA does it again." 00:14:51

Data center permitting, not labor or equipment, is now the binding constraint

SemiAnalysis tracks every data center globally and has been pushing timelines out because of permits, not supply chain.

Dylan Patel: "We've 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, um, the labor is not there, not because of some like shortage of equipment, but because, you know, permits are just not available." 00:18:48 On Texas specifically, the governor controls "water permits, wastewater permits, stormwater ... and there's like air permits," and "she can still stop them via the air permit." 00:17:50

The data center backlash is largely perception, not experience

Patel's poll shows widespread ignorance about data centers in people's own counties, suggesting the backlash is a proxy for anti-AI sentiment.

Dylan Patel: "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." 00:00:01 And: "I think people dislike AI more than they dislike your data centers. But then data centers are an easy way to point at it." 00:26:27 Patel claims "all of Elon Musk's data centers use less water than three in and outs," counting embedded water. 00:24:00

Product quality is becoming the differentiator, and the labs' product teams are weak

Patel says Meta's Muse and Instinct are better products than OpenAI's despite worse models, and that product stickiness will matter more than model rankings.

Dylan Patel: "Muse and Instinct, despite being worse models than OpenAI's models, are way better products than OpenAI Dots." 00:29:21 On OpenAI's interface: "This is a billion user fucking app." 00:32:49 On SemiAnalysis's own usage: "We've gone 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 back end is still using Claude and opening eyes, just a better harness." 00:30:40

Enterprise lock-in via spend-share discounting

Patel has seen frontier labs structure contracts to make switching costly, even as developer tools like Claude Code and Codex remain easy to swap.

Dylan Patel: "There's like contracts floating around where Anthropics are like, oh, if 70% of your AI spend is with us, we'll give you this discount. Um, and if it's not, you don't get that discount." 00:31:37

Independence is SemiAnalysis's moat, even against its biggest customers

Patel's business model depends on being credible, so he refuses paid opinions and criticizes his largest customers.

Dylan Patel: "We've never done ads. We've never done, you know, uh, you pay us for an opinion ... our biggest customers are the ones we hate on the most." 00:49:43 And: "My biggest customers are hyperscalers and major semiconductor companies and AI labs." 00:50:13

The $500M fund is a hardware-and-infrastructure bet leveraging a 115-person expert bench

Patel invests only where his team has deep expertise and says hardware requires both founder and idea, unlike software.

Dylan Patel: "I raised $500 million, SoloGP, you know, we're focusing on Seed Series A ... in the infrastructure space all the way from like robotics to, you know, silicon and optics and manufacturing." 00:37:22 "I think in hardware, it's harder to just have a great founder and they figure out their way. You kind of have to have a great founder and idea." 00:41:33

Hardware innovation is about to accelerate because AI is getting good at hardware tasks

Patel launched a benchmark to track model progress on chip design, materials science, cyber, bio, and quant, and expects an explosion of hardware companies.

Dylan Patel: "I think in a year you'll be able to like this, the number of companies making things in the hardware realm is going to skyrocket. And the number of companies that can do it well, it's going to skyrocket." 01:03:01

Space data centers are a 2030s story, not a near-term fix

Dylan Patel: "I think space is like a solution still two years away ... I know Elon's trying to do a shitload in 2028. Um, 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." 00:16:01

2. Contrarian Perspectives

Most startups will fail, but the AI app layer has real survivors, and the model leaderboard barely matters

Against the "who's winning this quarter" narrative, Patel says markets smooth these swings. He is skeptical of over-indexing on any one data point while still seeing real traction at Instinct, Sierra, and Harvey.

Dylan Patel: "The market doesn't actually shift the valuations that much ... as people who watch this stuff and are terminally online, we're going to watch this and over-index on any individual data point." 00:27:38

Selling some chips to China (via renting) beats a total ban

Patel resists the hawk-versus-dove binary. He thinks denying China chips entirely wastes profit, while uncontrolled sales are dangerous.

Dylan Patel: "I'm sort of of the belief that like, you don't want China to have too many AI chips, but you should still give them some, um, and still make some money off of them." 00:09:29

Anthropic and OpenAI are "bad at product," and safety guardrails are hurting adoption

Patel pushes back on the safety-first posture as a product liability.

Dylan Patel: "Some of the safety focus of, like, OpenAI restricts them, right? Like, they have this thing called Guardian, and it, like, stops the model from doing certain things in Codex and in Dots ... often overzealously." 00:29:21 And: "If it, like, texts someone incorrectly or, like, books an appointment incorrectly, like, fine, whatever." 00:29:49

Don't invest in what you don't know, which is the opposite of generalist VC

Dylan Patel: "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 shit about." 00:40:41 He also argues investors rarely make hard operator decisions: "A lot of investors have never made tough decisions ... have you fired people? ... Have you been sued by someone you fired?" 00:57:58

Being invested doesn't mean you must promote or soften criticism

Dylan Patel: "I have an investment, large investment in fluid stack. And yet we've still said their data centers are delayed publicly ... a lot of VCs, especially think they have to like be the tail that it's wagged by the dog that is their companies." 00:53:01

3. Companies Identified

NVIDIA

GPU and AI infrastructure leader. Mentioned as dominant but dangerously dependent on Jensen, and as the ecosystem's key backstop. Patel says NVIDIA "shut down even manufacturing of some of their high-end gaming chips just so they could focus on the AI chips" 00:07:40 and has propped up the ecosystem through "favorable allocations to NeoClouds, investing in a variety of startups, helping them with software, building out ... buying Hugging Face and Poolside ... backstopping NeoClouds." 00:12:27 Patel also says some NVIDIA practices toward certain neoclouds are "anti-competitive." 00:51:36

AMD

Competing chip vendor. Mentioned as a potential beneficiary of NVIDIA's China-related troubles. Dylan Patel: "AMD has free reign to sell to companies that are in Singapore that are renting to Chinese, you know, clouds." 00:14:51

Huawei

Chinese chip maker. Raised as a possible beneficiary of the scandal, and one of the chips SemiAnalysis is testing for inference performance. Dylan Patel: "Testing of, um, various chips, including Huawei chips, Amazon's chips." 01:02:08

Meta (Muse and Instinct)

Meta's Muse is praised as a clean, intuitive product. Patel beta-tested it: "I was able to beta test Muse ... it was like clear and good UI." 00:33:04 Instinct is also named, with "legitimate shot of just having some staying power." 00:33:59

Instinct

AI application company. Patel met its founder and praised its traction while questioning its valuation. Dylan Patel: "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." 00:33:59 Notably, it relies primarily on open models. 00:31:07

Perplexity

Powers the majority of SemiAnalysis's persistent-agent spend. Dylan Patel: "It's through, um, perplexity computer, which on the back end is still using Claude and opening eyes, just a better harness." 00:30:40 Patel is unsure of Perplexity's margin on pass-through. 00:35:27

Anthropic

Frontier lab. Patel says it "crushed it in Q2" and is "back to crushing it" 00:28:05 but criticizes product and discusses its 70% spend-share discount contracts. 00:31:37

OpenAI

Frontier lab. Patel praises its being among "the best companies in the world" yet criticizes delayed model releases, over-restrictive safety tooling, and poor chat app UX. [00:28:05, 00:29:21, 00:32:37]

Google

Mentioned as having "got back in" the competitive race. 00:28:05

Harvey

Legal AI company. Patel cites it as an example of growth with real ROI, while noting margin volatility in its public blogs. Dylan Patel: "Sierra and Harvey, like they're growing revenue really fast ... I don't know if you saw Harvey's like blogs, they like showed their margins ... seesawed all over the place." 00:33:59

Sierra

AI customer-service company. Cited alongside Harvey as growing revenue fast with real ROI. 00:34:28

Citrini Research

Macro/AI research firm acquired by SemiAnalysis. Patel praises their approach and independence. Dylan Patel: "They've captured the market's imagination. They've been right a shitload of times." 00:43:45 Examples of their scrappy research: they "developed a scorpion antivenom with AI" and "sent someone to the Strait of Hormuz." 00:46:32

FluidStack

Neocloud in which Patel has a large investment. Dylan Patel: "I have an investment, large investment in fluid stack. And yet we've still said their data centers are delayed publicly." 00:53:01

Iris Energy

Bitcoin miner turned AI cloud. Patel's nuanced take: "Iris energy has terrible technology on the cloud services side ... At the same time we've said, oh, but they have powered land and that's worth a lot." 00:52:32

Gartner and IDC

Used as the foil for a new kind of information firm. Dylan Patel: "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." 00:48:04 On Gartner's pay-to-play dynamic: "They're like, pay us or we're going to rank you poorly." 00:49:03

Elon Musk's data centers (xAI/SpaceX)

Cited as water-efficient and as pursuing space data centers. Dylan Patel: "All of Elon Musk's data centers use less water than three in and outs." 00:24:01

SpaceX

Referenced as having IPO'd and now being "an AI company." 00:27:20

Cerebras

Mentioned as having IPO'd earlier in the year. 00:27:20

Oracle, Microsoft

Named as clouds selling lots of GPUs to ByteDance, supporting the point that renting is legal and profitable. Dylan Patel: "Oracle or Microsoft or Google, um, all of whom sell lots of GPUs to a ByteDance." 00:10:42

Alibaba and ByteDance

Chinese demand centers for GPUs, either via NVIDIA-supplied capacity in Southeast Asia (Alibaba) or rented clouds (ByteDance). [00:06:20, 00:10:42]

ASML, Applied Materials, TSMC, Intel

Named as sources of SemiAnalysis's talent depth. Dylan Patel: "I have people who have worked at ASML and Applied Materials Land Research ... people who did R&D at TSMC and Intel." 00:39:17

Hugging Face and Poolside

Cited as NVIDIA's deals to support the ecosystem. 00:12:53

Brex, Harvey, Zone Frontier, Merkur, VCX by Fundrise, Public, Deel

Ad sponsors of the episode. Zone is described as developing "next-generation data center campuses." 00:23:08

New York Times bestseller (unnamed book)

Cited as overestimating data center water usage by 1,000x. 00:26:27

4. People Identified

Jensen Huang (NVIDIA CEO)

Subject of the scandal. Patel's concern is concentration risk, not moralizing. Dylan Patel: "Bad thing he did, right? Like very bad." 00:02:54 On NVIDIA without him: "Nvidia is like fucked in my opinion." 00:04:04 Patel previously idolized him as a business leader and now hedges. 00:47:18

Elon Musk

A "tremendous businessman." Patel notes he "at least owns" his personal conduct. 00:47:49 His data centers are water-efficient and he is pursuing space-based compute. 00:24:00

Governor Greg Abbott (referred to in transcript as "Governor Hino Hossa" in Texas)

The Texas governor, described as constructive but wielding permit levers. Dylan Patel: "She's not going to stop data centers. She's just going to make sure that they, you know, don't raise electricity prices." 00:16:57 Her power: "She has a lot of levers to fight back on any data center." 00:17:50

James (Citrini Research founder)

Dylan Patel: "James is awesome. He's, uh, you know, he, he's awesome. And the whole Citrini team is awesome." 00:43:45

Kyle (Citrini Director of Research)

Holds the final voice on publications. Dylan Patel: "Kyle, who's their director of research is like, got the most sort of voice on what gets published and not." 00:44:38

David Shor

Pollster who collaborated with SemiAnalysis on the data center poll. Dylan Patel: "We did this poll with David Shore and he collaborated with our team." 00:20:06

Dina Powell

Referenced as having highlighted, at an All In Summit, a county funding its education with data center revenue. 00:23:35

Apoorv Agrawal

Guest who proposed the "four seasons of AI" framework (Google, Anthropic, OpenAI, SpaceX). 00:27:20

Alex (Meta, likely Alexandr Wang)

Reached out to Patel to beta test Muse: "Alex like texted me. He's like, hey, try this new product out." 00:33:04

Dwarkesh Patel

Podcaster with whom Patel discussed credit markets. Dylan Patel: "People clowned us, but then we did the pod ... a month later, interest rates start skyrocketing and we're like, yeah, clown us if you want. We were right." 00:13:53

Doug

Occasional "board" member of SemiAnalysis. 00:56:18

Molly O'Shea

Host of Sourcery and a VC. Her observation that Patel is "building out an empire" frames the conversation. 00:35:42

5. Operating Insights

Rebuild your own AI stack around persistent agents and harness quality

SemiAnalysis moved to 45% of a $15M annual AI spend on persistent agents within months, mostly via a third-party harness rather than the labs' own agents. Test harnesses, not just models, and expect your spend mix to seesaw.

Dylan Patel: "Over the last few months, we've gone from, you know, zero persistent agents to 45% of our $15 million a year AI spend being persistent agents." 00:30:40

Resist single-vendor AI discount traps

Labs are structuring contracts to lock in 70%+ of enterprise AI spend. Negotiate with that in mind and preserve optionality, especially because the best harness may change.

Dylan Patel: "If 70% of your AI spend is with us, we'll give you this discount. Um, and if it's not, you don't get that discount." 00:31:37

Build a product around a credibility moat and never sell the opinion

Separate the objective data business from paid influence. The firm's value comes from being willing to lose a customer by being honest, and sometimes customer teams stay anyway because they still need the data.

Dylan Patel: "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." 00:50:40

Firewall businesses and invest heavily in compliance

Running research, consulting, an acquisition, and a fund under one roof only works with strict separation and compliance spend.

Dylan Patel: "The fund is a completely separate entity and company than semi-analysis and Satrini." 00:53:48 And: "There's compliance, you spend a lot of money on compliance." 00:54:13

Hire deep domain experts across the full stack and use them as a sourcing and diligence engine

Patel's 115-person team spans fab equipment, chip design, kernels, data centers, power plants, and model training, and they refer deals and inform investing.

Dylan Patel: "If they see something interesting, they'll refer it." 00:42:42

6. Overlooked Insights

Primary-source data center tracking is a hidden policy input

SemiAnalysis tracks data centers county by county and ran a poll of counties with 100 MW+ builds. That means they hold both the physical supply-chain data and a unique read on local political risk, a combination that could be monetized for utilities, developers, and policymakers. The "subsidize everyone in a 30 mile radius for nearly nothing" idea is a cheap, concrete mitigation that most operators haven't tried.

Dylan Patel: "We put out a post on, you could just subsidize the power of everyone in a 30 mile radius and it costs like nothing." 00:21:21

The union-labor mandate in Texas is a latent timeline risk for AI buildouts

Mentioned briefly, mandating union labor in a state where electricians and plumbers already earn $200K+ on data center projects could be deeply unpopular and slow builds. Combined with permit leverage, it suggests policy risk compounds quietly even where sentiment is nominally favorable.

Dylan Patel: "Electricians that are making 200 K or more working on a data center or, you know, industrial plumbers making 200 K or more ... All of a sudden, Hey, you have to join a union." 00:17:23