"Jensen Huang's Exit Could Change the AI Industry Forever" - Dylan Patel
- A Jensen Huang exit would put Nvidia’s decision speed, supply-chain leverage and customer financing under pressure simultaneously. Discussing allegations around Huang’s China relationships, Patel argues that replacing his ability to coordinate those functions would be difficult. Moving decisions to committees means “that's just gonna make decisions slow.”
- Outside Nvidia, Patel sees open source as the biggest loser. Nvidia backs American open-model developers, while tighter China restrictions could threaten another major source of open weights. Anthropic looks relatively advantaged: Patel puts its Nvidia exposure at one-third or less, versus more than 90% for OpenAI, and concludes, “In general, the closed source is winning.”
- The venture risk extends from model availability to the financing required to secure compute. Patel asks how inference providers grow from $1 billion to $10 billion in revenue without credit, a backstop and improving open models. He calls Huang “the venture ecosystem's strongest ally,” while questioning whether Nvidia will remain as active.
- Patel favors tighter China hardware controls while preserving some cloud access. He suggests Google renting chips to ByteDance could retain American control and extract margins, while floating possible limits on some chip companies’ rentals. Data generated with frontier models could face scrutiny as another channel for transferring capabilities. Retaliation remains a constraint because American AI servers contain Chinese components: “China can stop it if it wants.”
- Patel has become more receptive to AI 2027 while retaining doubts about diffusion. He says roughly 70% of its verifiable predictions so far that year proved accurate, despite his initial view that it was “a big stretch.” Recursive self-improvement is “looking more and more true,” but China’s response introduces supply-chain risks.
- GDPX shows sharp coding and chip-design gains alongside much weaker progress elsewhere. A host tentatively identifies the tested model as “Mythos 5.2” and cites roughly 16%; Patel emphasizes domain-specific reinforcement-learning environments and data. His call on valuable data suppliers: “It's experts. It's experts.”
1. Huang’s departure would expose Nvidia’s coordination problem
- The hosts frame allegations around Huang’s China relationships as a “black swan,” then question Nvidia’s deal pipeline without him. One had assumed Huang would prevent any AI company valued above $10 billion from failing for at least the next year, partly to avoid a damaging headline. Patel worries about decisions migrating to committees: “that's just gonna make decisions slow.”
- Asked whether Huang actually designs architectures, Patel says he does not; his role is committing Nvidia among competing options and switching “on a dime” if one fails. His example is Volta: Tensor Cores were added about three months before tape-out. Deferring them to the next generation, he argues, would have prevented Nvidia from establishing AI dominance.
- Supply-chain leverage is equally personal: Patel doubts a successor could secure 60%–70% of a supplier’s capacity below market rates. Asked about Nikesh Arora, he questions whether Palo Alto hardware experience translates to competing against Google, Amazon, AMD and frontier-lab chip teams. “These guys are fucking good.”
- Internally, Patel favors Ian Buck and Jonah Alben, with Alben closest to the technical leader; promoting either could dilute work they already own. His preferred outsiders are Broadcom’s Charlie Kawwas and Apple hardware chief Johnny Srouji for their technical and supply-chain command. Hock Tan or Lisa Su would be better, he says, but he does not think either would take the role.
2. Open models and venture financing share Nvidia’s backstop
- Patel says American open source “would not exist without Jensen,” pointing to Nvidia’s backing of Reflection and Thinking Machines. Applications and inference providers also depend on Chinese open models. If restrictions tighten, he questions whether those models keep progressing and whether developers keep publishing everything: “I think the biggest loser is the open source ecosystem for sure.”
- Hardware diversification changes the exposure. Patel says Zhipu supports seven Chinese accelerators and DeepSeek works closely with Huawei; Nvidia-dependent Chinese labs look weaker than diversified peers. The same comparison favors Anthropic over OpenAI: he puts OpenAI above 90% Nvidia and Anthropic at one-third or less. Weaker open-source competition provides Anthropic another benefit.
- Patel agrees venture investors are leaving the major-lab trade as those companies approach public markets; OpenAI “probably just did their last round.” Downstream, financing remains essential. He asks how Baseten, Fireworks and peers scale from $1 billion to $10 billion in revenue without compute access, credit and a backstop—or continually improving Chinese models.
- The application-layer distinction, in Patel’s framing, is control over the underlying model. Companies such as Harvey and Sierra forced onto closed models become wrappers; using open models while owning the customer relationship gives them another position. Huang is “the venture ecosystem's strongest ally,” but Patel does not know whether Nvidia’s investing will remain as active.
3. China restrictions could extend from hardware into training data
- Patel declines a Mercor-specific prediction but says many data companies sell heavily to China. Their pipelines use frontier models to generate, extract and filter material, including environments. Selling that output transfers information indirectly: “It's not like traditional distillation.” He expects greater scrutiny of chip sales and rentals, with data potentially joining an area previously focused on hardware.
- His preferred restrictions target chip shipments, TSMC and memory-company sales, and advanced chipmaking equipment, with possible data restrictions. He suggests not broadly restricting chip rentals, while floating possible limits on chip companies such as Megaspeed; he then qualifies that suggestion. Google could withdraw chips from ByteDance when needed or instructed, while charging margins that make ByteDance less profitable relative to Meta. The objective: “how do you make America win.”
- On the hosts’ “honeypot” framing, Patel says he does not think the woman involved is a spy or honeypot “in the colloquial sense,” describing a well-connected business figure. He argues such connections can bring pressure to cooperate with intelligence agencies. Whether Huang broke laws remains unresolved in his account: “I don't know.”
4. AI 2027 looks more plausible, but China complicates the trajectory
- A host sees capabilities tracking AI 2027 closely and asks about its November 2026 prediction that China “wakes up.” Patel says he originally thought the scenario “was a big stretch,” despite being highly bullish on AI. Roughly 70% of its verifiable predictions so far that year proved accurate, he says. He remains skeptical about diffusion, while recursive self-improvement looks increasingly plausible.
- Another host pushes back that leaving GPUs in Hong Kong could suggest longer Chinese timelines. Patel sees China waking up as a significant likelihood, but warns against an indiscriminate restrictions pendulum: every Nvidia server and Google TPU server contains Chinese content. Beijing could stop that supply, raising the question of what minimum access might prevent retaliation.
5. GDPX makes specialist data central to further progress
- A host tentatively calls the model “Mythos 5.2” and cites roughly 16% on GDPX. Patel says testing Anthropic’s model took about two days and revealed major software-engineering and chip-design improvements. He initially describes other areas as showing no improvement, then qualifies that it “did improve slightly in all domains.”
- Patel’s explanation is that reinforcement-learning environments determine where capabilities scale: without unique data and environments, particular domains will not improve. Anthropic’s chip effort illustrates the shift. Previously, he says, building a chip looked too slow relative to AGI; now “the model can make the chip.” Scaling size alone will not improve every benchmark: “You really have to scale techniques, algorithms, and data.”
- Asked about data obtained through product harnesses versus purchased from brokers, Patel says Anthropic buys very little chip-design data “as far as I understand,” instead building it internally. Biology uses a mixture: its own wet lab collects data while the company also buys substantial external data. Different domains therefore support different internal-versus-external sourcing strategies.
- Generic vendors connecting labs to merely adequate labelers look less valuable as models improve. “It's experts. It's experts.” Patel sees room for a strong business producing specialist material-science data, for example, while leaving its eventual scale unresolved: “how scalable is that actually gonna be?”
Full transcript
And we have Dylan Patel from SemiAnalysis coming back on the show. We missed him. It’s been over a week since we hung out with him. Is that a—
And he’s not—
Is that a foldable?
No, it’s not a Duo. It’s a Samsung.
Oh, Samsung. Living in the future. How long have you had that?
I’ve had foldables for three or four years.
Wow.
Ahead of the curve.
Truly.
Ahead of the curve, always.
Truly. What is your Duo review? Is that gonna be enough to pull you over? Were you only in—
iPhone Air.
iPhone Air? Okay.
iPhone Air.
Okay.
Foldable Samsung. This is the elite class.
Oh, two phones.
You know?
You’re—
The priestly class.
This is screaming a firewall between the investing arm and the research arm here.
One for—
This is secure.
One for work, you know?
Yeah.
One for work.
Okay, okay, okay.
One for personal.
Okay. There we go. There we go.
Good.
Well—
Good.
I am afraid to say that both Jordy and both of you guys, Jon and Jordy, only have this phone number, not this one.
Ooh.
Ooh. Wow.
We’ll have to fix that.
Wow.
We’ll have to fix that.
Wow.
To be fair, I only have this one.
I’m updating my priors.
Three or four months ago. Yeah.
No.
I’m updating our priors on our friendship. I thought I was texting the personal the whole time.
Wow.
I thought we were like that.
Like that.
Look, look.
Is this—
Only Chinese friends have this one.
Okay, okay.
—
Stay safe out there.
1. The Black Swan Nobody Saw
Do you guys try to predict black swans? It’s effectively impossible by definition, but I found myself two weeks ago thinking—
Hm.
Actually trying to rack my brain on what the most unpredictable outlier scenario could be that changes the game on the field. I think we got it. Unfortunately, I didn’t—
What were your guesses?
I don’t know. I think a lot of them were much more around individual players in the financial web of the entire trade, but—
Well, the big one that everyone’s been waiting for is some result from the next training run that shows there’s not another leg up beyond coding, that there is something—that taste is the new mode. For so many white-collar jobs, it really does matter that you can sit across the table from someone and shake their hand. We’re never gonna replace the lawyer or the doctor, or even augment them to the tune of—because you don’t even need to do the job displacement thing. It’s just the fact that a software engineer can deploy $100,000 worth of AI in a year without people blinking. That hasn’t happened for the marketer or the salesperson yet. If that does happen, the revenues do go to a trillion dollars. But if it doesn’t—
I don’t know if it’s a black swan, though. That’s not a—
No, it’s not a black swan.
That’s a—
That… yeah, that—
I think that’s something we can all debate, right?
Exactly.
Yeah, yeah.
Exactly.
No, but another thing I would say is, I was very confident that at least for the next year, Jensen would not let an AI company valued at north of $10 billion fail. Period. He was like, “I don’t—I’m not gonna have that headline. I have the ability to give any company a meaningful up round and keep everyone in the game.” A lot of people want the headline. Even though, in the grand scheme of things right now, a $20 billion company could go to zero and it wouldn’t really have that much impact, it didn’t feel like Jensen would even tolerate that kind of headline, just because the bears would sink their teeth into it and have a nice, tasty meal.
Yeah.
Yeah. I think the black swan is this, right? We’ve all predicted all sorts of things.
Totally.
Even with all the knowledge and insight and rumors. I’d even heard rumors, which I didn’t put much credit to, but people in the industry would be like, “Oh, yeah, Jensen does this with someone at Nvidia.” And I was like—
Whoa.
Palmer Luckey.
No, no, no. At Nvidia.
No, but that’s another—
Like, I—
Yeah.
That’s another thing we’ve been asking.
Well, you would never predict it, right?
No, never.
And you don’t—you hear something—
And certainly not to this extreme degree.
It’s impacting.
There have been plenty of other—
You hear rumors of stuff impacting.
This is so different than the Bezos story, so different than the Gates story, so different than the Elon story. It is so entwined with the actual business and the business dealings and geopolitics and domestic politics.
Well, Elon has slept with employees.
But not to this—
Not with his wife.
Yeah.
But his get-out-of-jail-free card is that he has 15 baby mamas, so you can’t pin that on—
Yeah.
Yep.
And he has a lot more control. I think the question that we’ve been asking ourselves—
Way, yeah, way more control.
We’ve been asking ourselves: on October 10th, 2025—
Hm.
We specifically named—
About Alice.
Megaspeed. We specifically named Alice Huang. We talked about how the whole situation was suspect. A lot of the details were basically out—
Yeah.
In the open, and all these videos were public. You could read into the videos and be like, “All right, this looks like, at the very least, a pretty close relationship.”
Mm-hmm.
Some of these videos have his wife and his daughter in the same room, right? That’s the thing that’s shocking.
Yeah, yeah, yeah, which kind of meant the basic analysis was like, “Okay, well, if the whole family’s there, probably nothing’s happening. Maybe this is purely—maybe he’s the purely rational economic actor. He’s playing it as aggressively as you can to grow the business internationally.” We don’t like that from a chip export position. If you’re a China hawk, you don’t like that. But there is the letter of the law, and we rarely—CEOs don’t get fired for pushing it right up to the letter of the law. If you’re selling the absolute max, like the H20, which was designed to get around the chip exports, a lot of China hawks didn’t like that, but they were like, “Ah, that’s the letter of the law. You found the loophole. You found the edge.” It is just a complete black swan in that it’s so much further than what anyone could see.
Yeah. So overall—
Total lack of judgment, just given that this guy, for as many friends as he has, has seemingly an equal amount of enemies that stand to benefit from this. What do you think happens to the Nvidia AI central bank deal pipeline?
Mm.
I assume—
Yeah, so I mean, I think—
I assume if you’re a number of these companies, you can’t lean on, “Oh yeah, Jensen wanted to do $2 billion,” and now it seems like those investments, when it comes time to actually sign the docs, will probably be going to committee.
2. Nvidia Without Jensen
Some things are gonna change. Yeah, I think that’s the biggest worry for Nvidia: now it’s a committee of people doing stuff, and that’s just gonna make decisions slow. Technically, they could fall behind because of this. Obviously, the other big aspect—there are three main things, right? There’s the technical side. Nvidia’s so ingrained in helping make decisions there. There’s no one clear technical lead. Maybe Jonah is, but there are a lot of people who are also directly reporting to Jensen doing technical stuff, so it’s hard to move as fast. Supply chain-wise, Nvidia wants to make aggressive things, do aggressive things, lock up supply, convince the supply chain to do things they’ve never done before in terms of technology and supply, and lock it up to Nvidia. That’s probably a lot harder to do, given whoever the CEO is probably isn’t as charismatic as Jensen. They’re not gonna be like, “Yeah, sure, Nvidia, you can have 60% or 70% of my capacity,” at a price that’s below market rate.
That’s probably tough. The third leg is that Nvidia is propping up a lot of things in the supply chain. Open source in America would not exist without Jensen. Both Reflection and Thinking Machines are pushing the open-source frontier in America, and both have massive investments and backing from Nvidia, as do many other companies across the ecosystem. All these application companies rely on open models, whether it’s Harvey, Instinct, or Sierra. They all use open models.
Yeah.
All the inference providers, including multiple companies worth more than $10 billion, rely on open models from China. If we’re slamming the hammer down on China much harder now because Jensen and his team were the biggest defenders against restrictions, what happens next? Their credibility is out the window, and they’re being investigated and subpoenaed. There are going to be a lot more restrictions. Will Chinese models continue to progress at their current pace? If they do, will they open-source everything? These are real questions. I think the open-source ecosystem is the biggest loser, for sure.
How do you think about the technical role of the CEO at Nvidia? You see Jensen making deals, at events, on podcasts, and in the news. If there’s deep technical work happening, it’s behind the scenes. How real is his involvement? Is he involved in the architecture of the next chip, or is his technical role more about recruiting and finding the next generation of great technologists?
I mean, he’s no doubt inspiring people and bringing them in. He’s not designing the architecture. But when there’s a decision to make, I think the common thing everyone following Nvidia sees is that you hear valid arguments for options A, B, and C. Nvidia could be doing all three at once, and then at the last second, Jensen puts all his weight behind one.
Mm.
If there are options A, B, C, and D, which one gets chosen? Is it a committee? Who’s deciding? Who chooses A and gets all the suppliers working on B and C to not be pissed off and actually push behind A?
Oh, yeah, that’s the founder mode thing.
And if A starts fucking up, you switch to B on a dime.
Switch to B, yeah. That’s really hard.
Those are the kinds of things.
I remember at one point at Meta, they had two different operating systems for their virtual reality strategy. They had XR and AR and VR or something, and it was just so much extra work. Then Zuck had to go into founder mode and say, “Look, some people are going to get reassigned. There are going to be some layoffs. It’s going to be aggressive.” That’s the benefit of a founder-mode company. Who’s on the shortlist of potential successors? Can Colette not do it? Does she not want it? How does this develop from here?
I’m not sure what Colette wants. I think she’s awesome, but at the end of the day, she can’t lead the company technically.
Sure.
I think the 2 people internally best suited would be Ian Buck and Jonah Alben. Ian is great. Jonah isn’t as much in the public eye, but technically he’s closest to the leader and is in charge of most of the technical organizations. Those are the 2 most important people from that perspective.
Ian is charismatic and probably pretty good at wheeling and dealing, though not nearly as good as Jensen. But will he be as good on the technical side or the supply-chain side? And will Jonah be good on the supply-chain side or at wheeling and dealing? I don’t know if someone can be propped up to take over Jensen’s role like that. If Jonah or Ian is in charge, does the work they were responsible for fall down to some extent, in terms of the focus they gave it and the relentless effort?
My favorite picks for CEO would be external. They would be Charlie Kawwas and Johnny Srouji. Charlie is number 2 at Broadcom and has been for a long time. He’s fucking awesome. Johnny is the head of hardware at Apple, and he’s basically led the chips team and strategy. Those are the only 2 people in the world I think can navigate a supply chain and technical details like that. Obviously, it would be better if it were Hock Tan or Lisa Su, but I don’t think either of them would do it.
What about leaders from the Trainium or TPU organizations? You’re highlighting Apple, but there are plenty of people leading huge semiconductor efforts who are still 2 rungs below the CEO or founder. This would be an opportunity.
Andy Jassy was the one who wanted to buy Annapurna Labs.
Yeah.
He’s not going to leave his job as Amazon CEO to be Nvidia CEO. You take a few steps down, and I think there are really great people there. I love the teams at Amazon. They work and move super fast.
Yeah.
I think Google’s awesome, but I don’t think they have a single person who can do all of the above. If anything, I think Amin Vahdat at Google leads all this infrastructure stuff, but I don’t know if he’d want to do this. I’ve never asked him; I just have a feeling he wouldn’t. I also have a feeling the leaders of Trainium want to stay in the background. I know they very much want to be in the background on all the Amazon chip stuff. I don’t think they want to do this.
Yeah. What about a wildcard? Sean Maguire was throwing out Nikesh Arora, someone who’s already in the CEO seat but isn’t running a trillion-dollar company. Do you have any expectations?
A humble, you know, just a couple hundred billion.
A tiny company, right?
I don’t know Nikesh too well. I’ve met him once. I think he’s awesome, but technically, Palo Alto has shipped hardware. That’s definitely part of their business. But is it the same thing? Is it Nvidia leading edge, doing crazy stuff? Look at Blackwell. Blackwell was amazing. If they hadn’t done what they did with Blackwell, they would’ve lost a lot of market share. They decided at the last second what to do.
With Volta, they decided to add the Tensor Core about 3 months before they taped it out. Who would’ve said, “Yeah, let’s just push out the current thing. Let’s add it into the current thing”? They would’ve said, “Let’s add it to the next thing.”
Yep. Totally.
If they’d added it to the next thing, how much further behind would they have been in AI? Nvidia never would’ve established dominance. These are technical decisions. And it’s not like you’re fighting—no knocks against Palo Alto Networks, but their competitors aren’t that amazing—whereas Nvidia is fighting an OpenAI chip team, a Broadcom chip team, a Google chip team, an Anthropic chip team, and a Meta chip team.
TPUs.
Amazon’s chip team, too. These guys are fucking good, as is AMD. They’re all really good, and so are all these startups.
That’s what I wanted to ask you about.
If you look at the startups, they’re all good. It’s not like they’re incompetent. It’s very different. As great as Lisa Su is, she took down Intel, which was incompetent. What is it like to be the person who defined and created a category and kept all the invaders at bay? That’s very different from being in second place or making smart tactical business decisions and innovating once a category is already established.
It’s just very different.
Mm-hmm.
How do you think Chinese labs are reacting and updating their road maps?
It’s an open secret that the Chinese labs are working super closely with Nvidia. Even if Nvidia isn’t selling them chips directly, engineers from both the U.S. and Nvidia’s Shanghai offices are constantly going to these labs, talking to them, and helping them optimize—
Mm-hmm.
—understanding what they’re doing and bringing those optimizations over to Nvidia’s software stack and open source. I think the Chinese labs have already started updating. Companies like Zhipu, which is sanctioned, and DeepSeek are heavy users of different types of chips. Zhipu supports 7 different Chinese accelerators. They’re not as good as Nvidia’s, but they work. DeepSeek is super close with Huawei.
There have probably been a lot of updates to their priors. The Chinese labs that relied heavily on Nvidia are probably weaker now than the labs that were already diversifying. I think the same applies to Anthropic versus OpenAI. Obviously, OpenAI has its own chip and has been working on it longer, but it’s more than 90% Nvidia, whereas Anthropic is one-third or less Nvidia.
Mm-hmm.
So it’s a completely different level of impact on them. OpenAI works more closely with Nvidia than Anthropic does. This is a positive for Anthropic versus OpenAI. Anthropic is pushing back on open source, and open source being weaker now is great for Anthropic.
Yeah.
These sorts of things flow through not just the Chinese labs, but also the American labs. In general, closed source is winning.
Yeah, yeah, that makes a lot of sense.
Any predictions on how this will impact Mercor and various data companies, given that there’s going to be more scrutiny on basically any type of business relationship with China?
I don’t know about Mercor specifically, but I do know a lot of the data companies sell tons of data to China. It’s actually a really funny way to distill this: even if the Chinese labs don’t directly have the closed-source models, some of their biggest customers are the data companies, like Mercor and so on. They generate data, and their pipelines for generating data involve a lot of language models from the frontier labs for extracting, filtering, and all that. If you’re making environments or whatever with these frontier models and selling them to China, in some way, you’re kind of distilling the information out of them. It’s not traditional distillation.
I imagine selling chips to China is going to get a lot more scrutiny. Renting chips to China through clouds is going to get a lot more scrutiny. Selling data hasn’t had any scrutiny yet, and it probably will. Right now, people focus on hardware, but—
Mm-hmm.
—and Jensen and all the people at Nvidia who were defending Lutnick and David Sacks, and all these people who were sort of preventing some of the restrictions on Nvidia in China—
Mm-hmm.
—are heavily discredited now, especially Lutnick and Jensen.
Mm.
In any argument to allow China to have some stuff, we’ve been in favor of certain types of restrictions for years. It’s been our biggest point of contention with Nvidia. I think they really dislike that. They could understand when we were excited about AMD or Google, and viewed that as, “We can out-engineer them.” But when it came to China, it was a different matter.
Now the cover is gone. I kind of hope the government is smart with the restrictions. Make sure chips of certain kinds don’t get sold. Really ironclad. Make sure TSMC and memory companies don’t sell to China. Make sure high-end equipment for making chips doesn’t go to China. Maybe have some restrictions around data. But don’t restrict renting chips.
There’s a whole waterfall of options around renting chips from, let’s say, hyperscalers. Maybe restrict random chip companies like Megaspeed from renting chips to China.
Yeah.
Or diverting them. But maybe not that. If Google is going to rent to ByteDance, maybe that makes sense, because Google can always pull away the chips when it needs to.
Yeah.
Or if the government tells them to. Google can always charge ByteDance a much higher margin, making ByteDance less profitable than Meta. These are the sorts of things that are relevant for U.S. competition and leadership. Hopefully, all the restrictions aren’t just rash actions as the pendulum swings. Hopefully, the focus is on how to make America win.
Yeah, no, I like it. I remember when you branded yourself not as a China hawk, but as a bald eagle. I thought that was a good way to put it. How are you reflecting on—
The video you guys made is so good. We’ve got to repost that.
Yeah, we’ve got to repost that today. How are you reflecting on the predictions in AI 2027? When I look at the model capabilities, I feel like it’s incredibly accurate. We’re completely on track. Huge congratulations to the whole forecasting team there.
I’m still a little uncertain about 2 predictions. One is that they forecast there would be a 100,000-person anti-AI protest. The midterms were kind of that, so I think the midterms might stave off a little bit of that. But the more interesting question is that around this time, AI 2027 predicts that, around November 2026, China wakes up. I’ve been interrogating that. It’s very hard to get accurate data. How do you feel we’re progressing along the line of China waking up? Is this an example of China waking up because they’re honeypotting our most powerful AI CEO, or are they—
Or—
—still asleep?
The other side of that, John, is that there’s always a possibility that they believe Nvidia chips have back doors in them, and—
Hmm. Oh, they don’t want them there. Okay, yeah.
—and Jensen is the asset.
Yeah, there are a million tinfoil hats—
No, but it doesn’t seem like you’re waking up if you’re saying, “Hey, there are a bunch of GPUs in Hong Kong, but keep them there. We don’t want them in here.”
That’s a good point. Yeah, how have you been processing that?
That makes you think there’s—
Yeah.
—some group that has much longer timelines.
Yeah.
As far as AI 2027, I’m very AI-pilled, and I thought AI 2027 was a big stretch. But it turns out 70% of the things they’ve predicted so far this year that we can verify have turned out to be accurate.
Yep.
Which is incredible.
Incredible.
Right? Wow.
Incredible. Yeah.
I’m still skeptical on some of the diffusion aspects, but I think RSI is looking more and more true.
Yep.
China waking up is probably a big likelihood.
Mm-hmm.
That’s the other thing, right? If we go too gung-ho with the pendulum of banning stuff in China, does China swing its pendulum and prevent the supply chain? Every Nvidia server has China content.
Sure.
Every Google TPU server has China content.
Yeah.
Every one of these things has China content. China can stop it if it wants.
So there’s some level of, “What’s the minimal amount of stuff you can give China in AI, and give them in everywhere else, to prevent them from cutting off the supply chain?” That’s a view of the world that could be construed.
As far as China waking up, it’s probably possible. On another thing you mentioned, China honeypotting Jensen, I actually don’t think she’s a spy in the colloquial sense.
Yeah.
And I don’t think she’s a honeypot in the colloquial sense.
Okay.
She’s a well-connected individual in the Chinese business world, and her family is well-connected. If you’re well-connected in China—hell, if you’re well-connected in America—you’ll talk to intelligence agencies when they tell you to, and you’ll tell them the stuff they want to hear. I just don’t think people are going to say no.
There’s just a lot of pressure there.
Yeah. The CIA was asking me for your personal phone number, and I had to tell them to take a hike. It was brutal. We gotta—
I only got this phone a few months ago.
—we gotta figure out where he got that jacket. John, can you get us Dylan Patel’s personal phone number? And I said, “Look, I’m kind of second fiddle in his world.”
Yeah.
Yeah.
Just a professional connection.
This was my personal and work phone up until a few months ago.
Okay, okay. I’m a little less hurt. No.
But anyway, I think—
Yeah.
—I think they were getting some information, right? She was clearly showing everyone on WeChat that she’s close to Jensen and can get them chips in China.
Yeah.
Jensen—did he know? Probably. I mean, I don’t know, but he probably knew.
Yeah.
But Supermicro didn’t have to honeypot Jensen to ship billions of dollars of servers to China.
Yeah, totally.
Right? I think there’s a lot more to this story than—
That we know of. It still hasn’t come out. They might have been having an affair. We don’t know. Let’s not jump to conclusions.
I’m sure, I’m sure.
Anyway, what does this mean for the venture capital community? We’ve seen a lot of VCs rallying around Jensen. Is that purely because of the open-source ecosystem? What does it mean for the other… I mean, you mentioned wrapper companies, but are VCs effectively out of the big lab trade, and is that why they’re rallying around Jensen a little bit more than you’d expect?
Before I answer that question, Jensen has the… It depends on how bad this stuff is, right?
Mm-hmm.
There’s a lot of information coming out. I don’t know.
Yeah.
But if their China stuff is really gung-ho, what if Jensen just goes to China and starts a—
Whoa.
—$100 billion raid? I don’t know. That’s a wild card. I don’t know how it gets stopped.
But why would he not go to Taiwan in that situation?
Well, because the U.S. still has interests in Taiwan, I guess I would say.
Yeah.
I mean, it depends. Did he break laws?
Yeah.
Is he going to be called in front of Congress for months and months and months, berated, and found to have broken laws?
Sure.
I don’t know.
If he didn’t break laws, I think he’s going to St. Barts. We learned that from Bing Lee.
But anyway, as far as the venture ecosystem, you’re totally right. They’re out of the lab trade. The labs are going public. At least OpenAI probably just did their last round.
Yeah.
It wasn’t that great.
Yeah.
And then Anthropic is going public.
Yep.
They got liquidity.
Yeah.
Some of them will… But then everything downstream, right? Like all of these inference clouds—how can they secure compute without credit, financing, and a backstop?
Yeah.
This is a real challenge. Who’s going to be willing to do that? And how do Baseten, Fireworks, and whoever go from a billion dollars of revenue to $10 billion of revenue without access to compute and without Chinese models being continually updated and sold externally?
Yeah.
This is a big challenge and risk. Likewise with companies like Instinct, Harvey, and Sierra, which are making awesome businesses: can they continue to ramp and improve their product’s capabilities, or do they have to turn to closed-source models? If they have to use closed-source models, their product is a wrapper. But if they can switch to open-source models and they own the customer relationship, they’re not a wrapper.
Yeah.
I think this is a big worry. Jensen is the venture ecosystem’s strongest ally, and he’s propped up a lot of companies.
Yeah.
Nvidia is very active in venture and investing, so I don’t know if they’re going to be as active now.
Mm-hmm. I had one last question. Jordy, do you have anything? I think the—
Yeah.
Yeah. Sorry.
I’ll tell you my phone number.
It’s just—
678—
Okay.
—triple nine, 8212.
Wait, what is that from? Is that a regional commercial from Georgia or something?
No, no, no.
No? What is that?
It’s—
Let me look it up.
Yeah.
Is it that song? There’s a song about it, right?
In L.A., we only know Dos Mantis—
There’s a song. 678—
Dos Mantis, Dos Mantis—
—triple nine, 8212.
—accidentes. Okay. I know 867—
Soulja Boy. Soulja Boy.
Oh yeah, Soulja Boy, who’s now an AI spokesperson. He signed with—
Soulja Boy is so cheap to hire. I went on a boat for some startup’s series launch or whatever. They rented a boat. Hiring Soulja Boy cost $25K.
Really?
And he performed for about an hour. He had his whole entourage.
Amazing.
His own DJ came.
We need him at SemiAnalysis CON 2027. Speaking of GDPX, were you surprised by the moderate jump in progress? I saw the 5.2. Was it Mythos 5.2? It got about 16%. It speaks to the quality of the benchmark that it wasn’t immediately 100%, but take me through your reaction to the results.
Yeah. GDPX in our benchmark we launched.
Yeah.
We tested Anthropic’s model and launched it. It took a day or two to run the test—more like two days. It was such a big improvement in some areas: software engineering and chip design.
Okay.
There were areas where it just didn’t improve at all, which was really interesting. It speaks to reinforcement learning environments being the substrate for scaling.
Yep.
If you don’t have unique data and environments in certain spaces, you won’t scale the quality of the model in those places.
Yeah.
Chip design drastically improved, right?
Okay.
I think it speaks to Anthropic’s focus. We just hired a bunch of chip people, and we’re making a chip. We’re making the model way better at chips.
Yep.
In the past, Anthropic’s view was, “We’re not going to make a chip because it’s not an AGI pill. By the time we get a chip out, it’ll be too slow.”
Mm-hmm.
Now the model can make the chip, so I think they’re really accelerating on that front. Likewise, software engineering continues to gain like mad. Anthropic is so much better than everyone else at software engineering and chip design. It’s incredible.
Mm-hmm.
In other areas, they just didn’t improve. I think it speaks to the importance of data, and to these data companies and vendors. It also speaks to the durability of some of these categories. They didn’t improve in some categories at all, and they’re probably not spending a ton of data on them.
Yeah.
Or maybe the benchmark is just so difficult that—well, it did improve slightly in all domains. But generally, scaling size isn’t going to get you improvement in every benchmark.
Yeah.
You really have to scale techniques, algorithms, and data.
3. The Value of Expert Data
On the data side, how much of it is…
How much value is there in the data flywheel, the data that comes through the harness directly and is given to the company, to the lab for free, versus data bought from a broker specifically? Do you have a sense of how valuable each is?
Yeah. We've established a relationship with a lot of the chip design data companies and AI chip design companies. Anthropic, as far as I understand, buys very little of this. They've been building it themselves. That's the only answer, right? How else would they do this?
That makes sense.
But at the same time, in some spaces like bio, Anthropic started its own wet lab. They're collecting data there, but they're also buying a lot of data elsewhere. I think other folks have different strategies in terms of what they build internally and what they build externally. But I think the data vendors are really important. Given that models are getting so good, a general, large-scale data vendor connecting you to a bunch of people who are just data labelers isn't what's valuable, right? It's experts. It's experts. If you're an expert in material science and you're making data there, you can make a great business. But how scalable is that actually going to be?
Totally. Totally. Yeah. Well, we'll find out next time you're on the show. Thank you so much for coming on. We'll let you get back to work. Have a good one.