Jensen on Dwarkesh, Cursor x XAI, Netflix Stock Sinks | Diet TBPN

0:00John Coogan:Absolutely wild podcast between NVIDIA's Jensen Wong and Dwarkesh Patel. So many clips, so much debate. We're going to kick it off with the question of whether or not NVIDIA is a car. Let me start off with this vibe reel from Semi Analysis because it is a treat. Wait, and pause for a second. You know why I was asking you? Yeah. Like what app you use to edit videos. Oh, yeah. It was to make this exact video. And then I saw this and I was like, oh, I don't need to make it. I don't, you're not talking to somebody who woke up a loser. And that loser attitude, that loser premise makes no sense to me.

0:37We are not, we're not a car. We are not a car.

0:44John Coogan:Such a great edit. It's absolutely, it's such a good vibe. And it's so funny because, I love to, this is Tokyo Drift. It's funny because it's this new type of vibreel where the semi-analysis team, obviously, they recontextualized the two clips and added the beat to get the video going. But that final edit with the cars sort of morphing together, like it goes from there's a match cut. These are match cuts between one stick shift move and another or one car drifting and then another car drifting. That takes a really long time. Somebody clearly made this video just because they're an enthusiast of Fast and the Furious.

1:25John Coogan:And then the semi-analysis team was able to quickly recontextualize that to be about Jensen's answer on NVIDIA's moat and the CUDA moat and what's happening with NVIDIA. So let's go through this question because it's been humming for a while. And it sort of bubbled up most recently because there was a whole bunch of news that Mythos might have been trained on Tranium, then maybe TPU, then maybe Blackwell. and sort of a mix. And it just feels like more and more of the AI labs are capable of making other chipsets work. In the early days, it was all about NVIDIA. And now it feels like the incentives are really high to figure other options out.

2:07John Coogan:And that creates a different competitive dynamic. So we can run through my thoughts on this. And then we can go into the geopolitical considerations as well, because that was another fascinating segment of the interview. So Jensen, the CEO of NVIDIA, spent over 90 minutes It was almost two hours when you include the ads too. In the ring with Dwarkesh Patel, it was electric. And the key question, at least for the start, was whether or not Nvidia is a car. And I'm only half joking about that. It sort of was the key question. Nvidia has been the market leader for years. During the gaming boom, Nvidia was the gold standard for rendering PC graphics.

2:43John Coogan:There was always decent competition from AMD, but once the AI boom kicked off, the CUDA ecosystem significantly sped up development of AI systems and training of AI models. And for those who aren't familiar, CUDA is a programming model that enables GPUs to accelerate demanding workloads by parallelizing computation. So instead of needing to work on the very low level instruction sets, if you are a CUDA kernel engineer, CUDA engineer, you can access all the power of the GPUs very efficiently while staying up in the more mathematical AI research, more standard Python C++ programming paradigms. You don't need to dip down too low, but it's getting easier to dip down lower and that's what we're seeing.

3:25John Coogan:So that created the CUDA mode because developers were way more productive and the biggest bottleneck to progress was allowing AI researchers to quickly test ideas and scale their experiments up to whole fleets of GPUs and eventually entire data centers. At the time, researchers really liked CUDA and really liked NVIDIA and they did not want to have to spend hours and hours figuring out other hardware systems. They just wanted to run their tests and see if the model was getting better and continue to scale. But recently the biggest cost center for AI labs flipped from researcher time, more or less, to compute capacity.

4:00John Coogan:And this creates a lot much larger economic incentive to figure out a way to drive down the cost of chips used to train AI models. I wrote about this back in on Tuesday, October 22nd of last year. So I said, not every link in the supply chain of AI can be completely commoditized. NVIDIA has an insane amount of power, having ramped full year revenue over the last three years from$27 billion to$60 billion to$130 billion. Absolutely insane top line revenue ramp at that scale. And that's why Jensen is so confident about how dominant this business is. It's the world's biggest company for a reason. It has been growing spectacularly at immense scale.

4:43John Coogan:And not only did they grow the top line, but net profit margin grew insanely. So it grew from 16 % to 56%. Yeah, you do that when you have pricing power and massive leverage because, you know, in this case, demand massively outstripping supply plus developer love and kind of just like. And I believe the forecast net margin was 70 something percent. We were hearing about 80 % potentially. And so NVIDIA, you know, the plan, and the plan is still to make an incredible amount of profit off of these chips because they are incredibly valuable. But all the hyperscalers and the AI labs, they are sort of incentivized now to form a bit of an anti-NVIDIA alliance to commoditize the accelerator market and drive down those margins at least a little bit.

5:30John Coogan:And so today the AI chip market is starting to look much less monopolistic. Like AI coding agents can make it easier to write software that works on non-Cuda chip stacks. And the teams behind competing chips have plenty of resources and economic incentives to bring performance in line with NVIDIA. Even if it's going to be a big hassle, even if you're going to need to spin up a team to get AMD or TPU working, it's going to be worth it because you're talking about billions and billions of dollars spent on chips. For the past few years, NVIDIA... Yeah, another example of like an instance where an AI lab had so much urgency that they were willing to spend whatever it took, was meta rebuilding like SL.

6:06John Coogan:Yeah, exactly. Like they were willing to outspend pretty much any other lab on talent because they didn't have time to find a homegrown talent or go through a normal recruiting process, especially considering a lot of those engineers were like happy doing what they were doing. Yeah, yeah, yeah, exactly. Yeah, the incentives flip when you get to this scale or the incentives just get so big you can build a whole team for a specific thing, solve any problem. And so for the past few years, NVIDIA has sort of looked like SpaceX's launch program. It's an incredible technology with very few viable alternatives, and so that creates great margins as we've seen with SpaceX's launch capacity, and they control something like 90 % of the launch market.

6:45John Coogan:While the products have not degraded, quite the opposite actually, Blackwell and Verirubin are incredibly powerful chips. They're clearly on the leading edge there, and they have an incredible amount of supply chain guarantees from TSMC and across memory and all the other different pieces of supply chain like NVIDIA is ready to make more chips, but increased competition makes the category look a little bit more like the car market than the rocket launch market. And so that I think is where Jensen is pushing back and saying, no, there's a lot more that we bring to the table with our customers that don't think of us like a car.

7:24John Coogan:This is not, you know, the difference between a Ford and a Toyota, they all sort of get you to the same place and you can swap one out. You can be driving a BMW one day, go to the dealer, turn it in for a Mercedes and you're going to have a pretty - Yeah. The other example that I was using with the team earlier was this idea of like, if you're a delivery company like FedEx and you have a lot of like Ford vans and then Hyundai comes to you and says, hey, you've been spending like$50 ,000 per Ford van, but like, would you consider a Hyundai van. We'll sell it to you for$35 ,000. It's just as good.

8:00And it could be like mildly inconvenient for the company because like they're kind of used to using board. They maybe have like an internal team that does some maintenance. But when you start looking at that cost differential, it can start to get pretty interesting to say like, hey, we should, why don't we try out some Hyundai's? Let's like move over some of our routes to Hyundai's and like see how that goes. And they try it and they're like, hey, this actually, this works pretty well. Maybe there's like, you know, higher maintenance, but it actually like masks out.

8:26John Coogan:Yeah. And like, we're going to actually start adding more Hyundai's to our overall fleet. Yep. And so Jensen's point on the podcast is that like, we're not, we're not selling cars, right? This is not like something like you can swap out. Yeah. And Dwarf Cash was obviously pushing, pushing back. And for a lot of companies, swapping out Nvidia for TPU would be very difficult. Some workloads that Jensen focuses on, he says, we're not a tensor processing unit. We're an accelerator. There's a whole bunch of scientific computing workloads that work particularly well with NVIDIA. The problem that Dworkesh was just saying, yeah, well, it's perfectly fine to have a specialized chip for specialized workflows because the biggest companies in the world, the biggest buyers here have a single type of workload that they're trying to do.

9:13And that's why they're using pure competitors.

9:15John Coogan:Yeah. Because the AI buildout does not seem to be slowing down as long as power can continue to be brought online and data centers can continue to find towns that will approve of them. Demand for chips will presumably grow, but every chip designer and AI lab has to be praying that those net margins come down. How quickly will it happen? Very early to tell. The market did not react negatively to this back and forth, although the price of NVIDIA has been basically flat since last August. And I think that's why Jensen is trying to sort of like reset on the narrative, because it's possible that, And there's a lot of movement up and down.

9:53John Coogan:They've been sort of flat. There's a desire to sort of reset and recontextualize. Yeah, inference demand is scaling in a very significant way. And the NVIDIA stock price is relatively flat. Yeah, and since August, there's been so many different moments where fears of the AI bubble, fears of the products not finding product market fit, that revenue might stagnate. Like we've seen a ton of bullish signals for like AI demand broadly. Demand is there. And so if you're a supplier, you should also be going up. But there's been this overhang of what will happen to margins and market structure. And so that is what people are going back and forth on.

10:33John Coogan:Tyler, did you have anything else, thoughts on this? Personally, I think I probably am more on the Dorkesh side where like, yeah, I mean, at some point, like these margins are so high. There's so much opportunity here. We've seen that like actually if you're a big lab, you know, it's a lot of resources. but you actually can just like train a model on a different architecture. You can serve them on a different architecture. Like you actually can figure these things out. As models get better, you can go lower and lower on the stack. You can write the kernels semi-autonomously. These things get faster and faster.

11:01And it's like, I'm probably much more on the DoorCash side.

11:04John Coogan:Yeah. It'll be interesting to see how Grok fits into this, the new CPU integration between different pieces of the puzzle. And then also DoorCash makes this point that the supply agreements that NVIDIA has might be a bit of a moat for the next few years while TSMC line time is so constricted. Yeah, but even then, like, you know, Jensen was saying that kind of supply constraints, this is like a two to three year problem after that, like, you can just solve these things. Like, I don't know how to think about this, because, you know, to me, it seems like, yeah, so much of the value of NVIDIA is just like, they have such an incredible relationship with TSMC.

11:36Yeah. But if, and it's so valuable because of how constrained it is. Yeah. But if that kind of constraint is maybe going to you know go away to some extent tsmc yeah that's what he says like they're going to increase you know they can build a new fab and sure whatever three three years i mean the big the big the big takeaway from the conversation is like you have one person who seems incredibly agi pilled yeah which is dorkesh yeah and then you have jensen who doesn't seem agi pilled in that sense at all right he uh dorkesh asked about he was kind of getting at early on the idea of like will you be able to just like prompt your way to nvidia chips he's like you basically sell software yeah um and jensen was basically like no i don't i don't think that'll happen and then and then when you get to the whole geopolitical conversation too again it's like dwarikesh is like you're selling nukes and um jensen is in my view like i'm selling computers and uh that that was like the big rift it was like these two kind of totally conflicting worldviews and it made for some very...

12:41John Coogan:Pay-per-view. It was a pay-per-view. It should have been a pay-per-view. Shreeram Krishnan summed it up pretty well. He said, every person hears reaction to the Jensen plus Dworkesh podcast can be extrapolated directly from whether they believe in the Frontier Labs achieving short timelines for AGI, ASI. If you believe in the labs achieving RSI and then AGI, ASI for some definition of all three in the next few years, you're probably sympathetic to the frame Dworkesh adopts. If not, you're probably more sympathetic to the arguments from Jensen. And so we can go into the export controls next and talk about that.

13:17John Coogan:Metacritic Capital sort of summed up a little bit of why just Jensen's rhetoric and how he wasn't conceding a lot of things. Dean Ball said, Dworkesh Jensen reveals how inconsistent and unbattle-tested AI acceleration talking points are, especially when they're filtered through the prisms of corporate comms and mass politics. Strategically coherent accelerationism is possible. He says, I try, but not currently prevalent. And Metacritic Capital says, the problem is that Jensen doesn't concede anything. Compute spending going to the moon. One trillion revenue in sight. Models keep getting better.

13:47John Coogan:No unemployment. Software codes are good. Other Western accelerators are bad. Chinese competitors are good. NVIDIA makes token costs decline 90 % per year. But Chinese compute scientists are capable of making all the necessary algorithmic improvements. He also can't be AGI-pilled enough because at the end of the day, he is an intellectual property company in the business of sending a file to TSMC. I think it's part of Taiwanese culture to want to be loyal to all clients and don't have favorite winners. He doesn't want to betray his software co-customers. He has antitrust concerns. Yeah, he was making the bull case for software, which was that AI agents are going to use tools.

14:19So he's like, there's going to be more users of software than ever.

14:23John Coogan:Which is something I'm somewhat sympathetic to, but yeah, it definitely... It's still very easy to take the counter. The flip side of that, yeah. Well, let's play the distilled recap from Dwarkesh Patel of the back and forth with Jensen on export controls. It's about four minutes, and we'll watch this and then discuss. If Chinese companies and Chinese labs and the Chinese government had access to the AI chips to train a model like Claude Mythos with these cyber offensive capabilities and run millions of instances of it with more compute, The question is, oh, is that a threat to American companies, to American national security?

14:58First of all, Mythos was trained on fairly mundane capacity and a fairly mundane amount of it by an extraordinary company. And so the amount of capacity and the type of compute that it was trained on is abundantly available in China. And so you just have to first realize that chips exist in China. They manufacture 60 % of the world's mainstream chips, maybe more. It's a very large industry for them. They have some of the world's greatest computer scientists. As you know, most of the AI researchers in all of these AI labs, most of them are Chinese. They have 50 % of the world's AI researchers. And so the question is, if you're concerned about them, considering all the assets they already have, they have an abundance of energy, they have plenty of chips, they've got most of the AI researchers.

15:44If you're worried about them, what is the best way to create a safe world? Well, victimizing them, turning them into an enemy likely isn't the best answer. They are an adversary. We want the United States to win. Having a dialogue and having research dialogue is probably the safest thing to do. This is an area that is glaringly missing because of our current attitude about China as an adversary. It is essential that our AI researchers and their AI researchers are actually talking. It is essential that we try to both agree on what not to use the AI for. With respect to China, we want to have, of course, we want the United States to have as much computing as possible.

16:21We're limited by energy, but, you know, we got a lot of people working on that. And we had to not make energy a bottleneck for our country. But what we also want is we want to make sure that all the AI developers in the world are developing on the American tech stack and making the contributions, the advancements of AI, especially when it's open source, available to the American ecosystem. It would be extremely foolish to create two ecosystems, the open source ecosystem, and it only runs on the Chinese tech tech, a foreign tech stack and a closed ecosystem. And that runs on the American tech stack.

16:56I think that that would be a horrible outcome for the United States. I mean, I think the concern going back to the flop difference in the hacking is, yes, they have compute, but there's some estimates that because they're at seven nanometer, they don't have UV because of chip making expert controls. the amount of flops they're about to actually produce, they have like one-tenth the amount of flops that the U.S. has. And so with that, could they train eventually a model like Mythos? Yes. But the question is, because we have more flops, American labs are able to get to these level capabilities first.

17:26And because Anthropic got to it first, they say, OK, we're going to hold on to it for a month while all these American companies, we give them access to it. They're going to patch up all their vulnerabilities. And now we release it. Furthermore, even if they train a model like this, the ability to deploy it at scale. You know, if you had a cyber hacker, it's much more dangerous if they have a million of them versus a thousand of them. So that inference compute really matters a lot. In fact, the fact that they have so many researchers who are so good is the thing that makes it so scary. Because what is it that makes those engineer researchers more productive is compute.

17:54John Coogan:We should always be first and we should always have more. But in order for that outcome for you to what you described to be true, you have to take it to the extremes. They have to have no compute. And if they have some compute, the question is how much is needed. The amount of compute they have in China is enormous. I mean, you're talking about the country. It's the second largest computing market in the world. If they want to deploy, aggregate their compute, they've got plenty of compute to aggregate. Very, very tense back and forth. Dean Ball says, it's a shame Jensen mostly fails here because the monoculture on export controls is bad.

18:28John Coogan:If you're a young AI policy researcher trying to make a name for yourself, it's almost impossible to be taken seriously unless you are pro export controls. Monocultures are usually bad. And I am sympathetic to Dwarkech's points there for sure, especially on the inference side. Even if models exist in both worlds, like having a whole bunch of, you know, good guy compute that can go and patch bugs while the amount of attackers is much smaller. It's just a matter of, you know, how many, you know, resources you have on each side. That's a great point. The only thing that they are like talking around is just Taiwan as a particular turning point and how their various positions flow through to Taiwan policy and where the China stance on Taiwan is something that I've always puzzled.

19:20John Coogan:And I wish that both of them had articulated their sort of philosophy on actually war gaming out what export controls do to likelihood of Taiwan intervention or blockade or anything like that. I don't exactly know. I've been trying to like work through it but I don't have a complete thesis but we've been debating it back and forth all day. I don't know if you have a strong take on any of this, Jordy. I appreciate Matt Zeitlin's point. I kind of appreciate that Jensen Phong seems relatively normal about non-business stuff compared to other tech founder CEO types but then when it comes to NVIDIA's actual operations he's a complete sicko.

19:57John Coogan:Yeah, I saw a couple takes around this that he had like some very, very, very, very strong points if you're deeper in the supply chain. I couldn't really assess how he did on that, but some people were, there were definitely people that were in Jensen's camp. It was divided, which I think is why this went so viral. Jensen Wong on Dwarkesh today, most combative interview he's done in a while. The biggest regret not funding him. Has there been a more combative interview ever with Jensen? It seems unlikely. I think so. There is some funny details in here. apparently the Larry and Elon begged Jensen for GPUs at dinner story.

20:35John Coogan:That never happened. We absolutely had dinner. At no time did they beg for GPUs, which is funny. I wonder what happened. Dwarkesh posted tomorrow and it's him and Jensen standing next to each other and Alex Volkov said, hey, Dwarkesh, was this picture taken before or after the pod? Because it does feel like it was a tense situation. Although to both of their credit, Jensen, it felt like he loved being in the arena, getting asked hard questions, working through this, there's this back and forth where Jensen's pushing back and Dwarkesh says, oh, I can drop it. And he says, you don't need to drop it.

21:08John Coogan:I'm enjoying this. Like, let's hash this out. And I thought that was very diplomatic and just good overall. Well, Intel is up on the news, up 4 % today, 10 % over the past five days, almost at all-time I think we're very close to the 2000 peak for Intel, which it was also around where they were trading in 2021. $67 a share,$330 billion company. Clearly with all of this backdrop and just the idea of more chips and maybe the CUDA ecosystem being something that you can work around, can an American fab run by Intel produce a chip that's viable for an AI lab? It feels like increasingly yes. That's certainly the argument that's being put forth by Dwarkesh.

21:58John Coogan:And it would be very exciting. I think it's something that everyone would support an Intel resurgence. There's some news around TerraFab potentially getting involved. But first, let's start with the scoop from Grace K over at Business Insider. She says, Scoop Cursor plans to use XAI's infrastructure to train its Composer 2.5 coding model. Where's the golden scoop, Tyler? According to people familiar with the matter, Cursor will use tens of thousands of XAI's GPUs, they said. And we got a scoop for Grace K. This one's going to Grace. Grace, congratulations. You win the golden scoop of the day. Congratulations.

22:38Interesting to see something we talked about probably midway through last year. XAI has shown a tremendous ability on the kind of infrastructure. Yeah. Data center side spinning up a huge amount of compute very, very quickly ahead of any timeline that any reasonable party would have probably expected. And demand hasn't exactly followed in the way that they would have liked. And so opening that up to a company like Cursor, who has all the demand and what they really need is their own model.