Aaron Ginn: US export controls are backfiring — China is demolishing the US on AI efficiency because we gave them Nerf chips

level. You really are brilliant. I'm grand. You really are grand. I have grandeur. Well, our next guest has grandeur. Aaron Gin, welcome to the stream. How you doing? Lovely hat. Let's go. There we go. Let's hit the gong for a hat appearance. Boom. Didn't even warm it up today. Take that. Appropriate as well.

Culturally from sales guys from from from BDRs all over the world. Yeah. What's new in your world? Hs good, bad. Are you happy? Are you pumped? I'm happy. They're happy. They uh Yeah.

They they they choppyl hawks that have a luxury belief system around that Chinese AGI is going to destroy us like a crouching tiger hidden bag, you know, dragon version of open AI is going to attack America. Like we're done. And and that this 15% uh deal which like broke Sunday. I don't know how it broke. Yeah.

Was just amazing. It was just like I mean one like props to props to President Trump who is the ultimate deal man. Art of the deal. It felt like both 4D chess and art of the deal into one. Yeah. And and so you know like the the the sort of average profit margin of the H20 is around probably like high 50s%.

So you know there are still and they also have the write-off they're probably going to roll in as well and and there's some struggles with DSMC like you can't just you know I think it's the one of the issues related to the intel foundaries like you can't just like restart production within a month.

It takes several months. So like probably you're just going to try to find someone's non-destroyed and probably send them to China the the age 20s. Uh but what most Americans understand is that it's unconstitutional to export the tax exports. Yeah. So it's like in the constitution.

I looked it up and I was like how is this going to happen because this So you're saying a revenue share is like a workaround. Yeah. So it's voluntary. Like that's the only way it could happen. It's tipping.

No tax on tips cuz the because the taxes are tips now because like even from a process perspective like Nvidia doesn't apply for export licenses like the OEM does and when you apply for those OEMs there's no license fee cuz that violates the constitution. So you just get it and you can ship it but the bomb is listed.

The bomb is wait really quickly who's the OEM in this scenario? Oh super micro Dell. Yeah similar. So, so, so, so someone in China will go and say, "I want a rack of servers with H20s already on them, some other CPU, some other memory, some other hard drives on there.

Bring those bring those blades to me and they will call Dell or Super Micro for that. " Yeah. And so, that's who applies for BIS. And so, they're the ones that that final PO number is all that crap together. It's not Nvidia. It's not Nvidia. You don't know Nvidia's pricing.

You don't, you know, Dell does cuz I buy it from Nvidia. So, that's why it has to be something volunteer. You can go to the website now like Jordy if you're interested in donating to the Goliath you know the lethon you can go to treasury. gov and you can find the little we accept PayPal, right? VMO. Yeah.

So, so it's it's quite brilliant. I mean, I think many of the faux neo China hawks are devastated. Uh I mean my my oped last week. Yeah. It's just they're just sapuku. They're just like, you know, they're there. They're going to be on the Jefferson memorial just like, you know, going at it.

man because it's a decision but but today yeah it's a it's a decision by the administration to be like you know we don't really care like like we want to do deals we want the American worker to have you know get money from China we want to we want to do what China did to us over the last 20 years we want to take money dig into the faux China hawks the fauxhawks as we uh explain like the uh stated preferences, revealed preferences.

Are they really doompilled? Do they really believe in fast takeoff? Do they really believe uh that AI is a nuclear weapon or is there something else going on? Some other economic consideration that they're working through.

I think and I'm working on the oped right now for a very liberal newspaper, but they I think it's a a jox machina uh sort of fallacy they built in their brain where they think AI is going to be God and America has to build the first god or China China has to build the first god.

And so it's a kind of tower babble moment. And and the problem is like if you remove that from their vote, that's why I call them faux neo China hawks. Like they're not really China hawks. They're actually agi zealots. If you remove it, none of their policies kind of make any sense.

This one reason why you hear, oh, like we have to prevent chips. Oh, we have to block models. Oh, like you know, we have to restrict um, you know, model ways. We have to do tokens. Like they it's just kind of like amorphous blob. But they can't really win in DC with the idea that AI is gonna be God.

Most people are gonna be like, "That's kind of insane. " Like you're go back to to reading, you know, the go back to watching Matrix or, you know, reading the Terminator or some some other related sci-fi novel. And and so they kind of wrap it in something that is like super easy to understand. China. Oh, China, right?

You know, oh, the the yellow country, right? And and and they and oh, they're so scary, right? But but you look at the specific policies, none of it like makes any sense. None of it.

I met with a venture capitalist last week who said honestly none of this stuff's going to matter in two years when we have fast takeoff but he wasn't he wasn't doomer pill he was just saying like fast takeoff will mean that every single company goes to zero because there will only be one company which is a very funny take he has invested in SAS this year and yeah and it's yeah so the revealed preferences is is all over the place um but yeah it is it is weird to fe to feel to make the argument that like uh we're going to develop AGI god but It won't like it it it it will be it will be God, but it won't be subservient to actual God.

But also or or it will sit below the stack of it'll it'll it'll respect the federal government of the United States. It'll be God, but it won't but it won't but it won't but it'll do what the president says or it'll do what the House and Senate say or it'll do what one company says or what one person says.

It's like if it's God, can it be controlled? How does that make any sense? That seems very odd. Yeah. Yeah. It's a and and and as you and I talked about, John, it's like like so much is built into of their presuppositions are defined by the title two debate.

Remember that whole thing of like, oh, free speech is going to die. We have to regulate and all stuff. Many of that same donor class, many of those same policy people are also the AGI doomer zealot people.

You go back even like a like series of steps further like many of those people who are super hot on new atheism and like and like oh this idea of like we can we can prove via science. We just need more time. We just need more scientists. We just need more people.

It It's the same pattern in humanity that we create these kind of albatrosses.

This idea of like creating an an all-in or sort of argument that because they don't have that in their life and then they make this thing and and and as I said like on a on a one of I think times I appeared that if you don't have that metaphysical framework that is traditionally Judeo-Christian or Western, you try to fill something else in that void.

It's a classic Charles Spurgeon quote that that you you basically fill it with something else outside of just the fact of what the place of like God or faith is supposed to fill and technology I would say Silicon Valley probably San Francisco they don't really understand that there's so much faith goes into actual the VC industry that is you're really kind of selling BS to an investor you know like the I know y'all lots of jokes around what is ARR in AI it's mainly you're just selling kind of like nonsense and the idea that like this could be something this could be it is a it is a faith driven investment that's what also makes it so awesome why we why we look at Zuckerberg we look at Elon who he made an idea and the vision the future reality today so it is heavily oriented to towards a metaphysical theological framework but this AGI thing as it relates to China so much of it has just gotten lost in this I don't even know what I'm talking about anymore because the these same people who are criticizing the 15% tip the to the federal or government to the American worker uh for Nvidia are the same people who didn't want to ship GPUs to Switzerland because I don't know they make two good watches or Portugal they have too many good sandy beaches like they they the same people right so it just keeps shifting because it's really trying to provide a luxury belief to a base around the idea of AGI what's your what's your line though because I think that the the critique here that's pretty fair is that if you start applying this to other categories of goods like fighter jets like what what What's what would the you know fair price be to export an F-35 to somebody that nah maybe they're on the on the edge?

Maybe they're kind of friendly now, but maybe in the future uh they they they might not be so friendly. Like you believe there should be we should draw kind of a line and uh we shouldn't want government agencies to kind of apply this uh approach to fund raise.

No, it's it's a difference between it's not a hunk of metal versus like you're selling a system. I have the same sort of concern as it relates to Nvidia as it relates to Apple building phones in China.

If I got an iPhone, no OS on it, none of the the magical, you know, more and more transparency via, you know, the the design, then I then it's kind of useless. It's just like a phone, right? And and that's the way Nvidia actually is. Nvidia is an Appleike ecosystem. It's not just a hunk of metal.

Same with the reason, you know, we we sell 787s, we sell trip 7s, we we sell the trip 7X to China. Why? Because it's made up of this complex the thousands of companies. It's a software platform. We remove certain things which I'm generally okay with like advanced avionics.

Uh we remove we we also kind of limit certain types of engine design. I'm fine with that. So with the B30, H20, these kind of nerf chips, I I don't actually make much of an opinion about it. Like I don't have to have a no true Scotsman sort of argument about like well at what point would you not do it?

I'm like I don't sell fighter jets. I don't rent fighter jets. So you can talk to Lockheed about that or Boeing about that. Like what what what I think is most relevant is the fact of these innate externalities these exogenous factors outside of the machine which that half of all AI engineers are Chinese.

half or actually more 75% half of them are Chinese nationals. Another if you include all the Chinese associated last names in Meta's super team are Asian.

So like we either have to lean into our strengths which is the fact we're good at making platforms, we're good at dominance, we're good at free enterprise, we're good at trade and and the fact that Chinese have something that they don't have the raw production capability of doing which is not Nvidia.

that if I don't take anyone seriously who criticizes export controls if they associate Snick with Nvidia it's like saying Kirkland like Kirkland and Costco like Costco makes Kirkland like that's not what Kirkland is right Kirkland is a rebrand of somebody else's stuff and that's kind of like the criticism it's like oh like how dare Costco make this Kirkland thing you're like it's a rebrand bro like it's not it's I don't actually make this thing like it and so that that's where this constant confusion lies that many of these people criticize this stuff are actually criticizing Smith or or TSMC completely different companies Jensen buys from them.

So if you want to go after the foundry stuff ASML the Japanese like you have to realize one we don't make this technology it's it's Japanese and Dutch but the other is like it's not what Nvidia does if if you want like I'm cool with targeting uh lipography equipment preventing them from getting to 5 nometer and that's that is something that's actually quite reasonable because because we do export control trip 7s and etc.

like we have some limitations where we send that's fine because it's a freaking size of a house and it cost $200 million but the same of lithography so much of what I what I'ming for is pragmatism is like realism like like we have to lean into what our strengths are we have to proliferate and prevent Huawei from spreading its wings outside and if we want to target the ability to manufacture 5nime or 2n 3 or whatever like most Nvidia stuff's on a 4 nimeter process that's cool like I'm totally in favor of that but don't attack Nvidia because that's not what Nvidia Yeah, makes sense.

Um, on the foundry side, I want to jump over there. Let's let's talk Intel. There's been a lot of headlines recently. What's your read on the situation? How important is Intel in your view? Uh, do they need a new board? Do they need a new CEO? I don't want I'm gonna I'm gonna get in hot water here.

Yeah, it's um you know, so like my dad's Chinese and and so there this thing like I am I'm a true China hawk.

I'm OG China hog and and there's many times where like China's the largest country in the world and there's many overlaps that are just completely not you know they're they're just fine and and I don't know anything about the particular criticisms of Intel CEO I think his strategy is not great for Intel so like I mostly would target there but America needs a foundry program and particularly trailing edge uh leading edge I don't think is economically viable uh so that's why I'd rather ship that to like a Brazil or Argentina or something like that like do a newman road doctor and sort thing.

Why is that? I feel like the leading edge is coming into Arizona through TSMC. Samsung's doing stuff like I feel like Yeah. Trailing South and then leading edge here. Oh, okay. Yeah. Yeah. Yeah. Yeah.

Traing edge also economically, but we can import we can import the the semiconductor, you know, geniuses that will run American TSMC. Yeah. And and you also have to understand that like there's lots of reports. You can go read about it of a former, you know, Chinese nationals that work at TMC that takes that back.

Like it's it's a Taiwanese company. Like there there's there's an there's an arm that we can extend as a nation that can only go so far. And that's part of the America first concept of Trump. He's like, "Do I really want to take all of the juice and squeeze in terms of negotiating presidency over a nerfed H20?

" He's like, "No. " Like like we have fentanyl issues. We have 300 billion400 billion dollar trade issues. We have them IP issues, right?

We have Chinese nationals at labs all across like those are much bigger things to pry than the fact of like oh they get like you know basically an equivalent of an A100 uh being deployed into to to their to their fabric.

So the so the Intel stuff I I kind of have this disposition that it feels like it's going to be nationalized. I get kind of like vibes around that as a as a economic conservative. I'm like oh that's kind of scary. But at the same time, like it it is one of those things that just kind of happens.

Like the the Did y'all hear about the uh the federal government invested into a Rare Earth Minerals project in America? The one in Vegas, right? They got like 15% for the for SAM. Yeah.

And so so I think the same is going to happen with with Intel is like they basically it's it's it's occurred through there's a couple different lending vehicles in the federal government. It's basically gets a warrant.

That's kind of the vibes I'm getting from Intel is that something like that's going to happen because you also can't take TSMC process and Intel processes and just mix them together. Like that's not the way this works.

Uh, and also TSMC, uh, I think whenever they originally pitched that idea, I think the president was like, "Yeah, like do Intel thing. " And TSMC was like, "Maybe, right?

" And then they go look at and they're like, "Holy, this is a freaking nightmare of people who don't want to work, who don't have the work ethic of TSMC. " And and that just kind of all blew up.

And I think there's some embarrassment in terms of like that this is America's founding program and this is kind of like how it matches up to the best of the world. So, I just get kind of vibes that it's going to be nationalized or split apart and divided as a company.

Um, and but in terms of bearing on like my opinion about whether or not that guy should be a CEO or not, I would point more towards the strategy rather than these kind of rumors or we had Doug on from some analysis on Friday and he he basically said like look at the board like a lot of the people that are still on the board have just like overseen like numerous CEO changes.

is the strategyy's clearly not been working like you know well they're not semiconductor experts was his main critique yeah they're finance guys finance guys but also politicians and and thinkers like it's a it's a wide variety uh without a lot of focus on like manufacturing expertise uh what was your what was your reaction to GPT5 uh not as negative as John uh and uh I was negative yeah you were like oh like it's over and I was No, I don't think so.

Like see, so so first of all, first of all, so I think I think it's like pure pure numberbased GBT revisions being indicative of of order of magnitude scaling is over. Yeah. Like so the pre-training scaling law is over. I I So one is like a couple months ago we didn't say that because Rock 4 came out.

So like so it it's so I would like maybe take a more like Prudence perspective on this. Yeah. One is that there's a clear decision here by uh Sam that he is a consumer app company. Exactly. Which is which is which is the right bet. It's like I don't sell consumer SAS. Yeah. Yeah. I'm not I'm technical.

I don't remember what all the freaking numbers and crap does. Doesn't matter. Yeah. And so that's the biggest innovation here.

And if you look at the the idea like you know he he expanded the context one open not very much which implies inference right that that he doesn't have enough capacity to run as many user requests.

The the other is that he mostly solve for downside scenarios and ease of use in the UX and those are like really big changes.

the the the benchmark maxers are uh basically creating goalposts that are artificial like if do you do you mean do you all remember your SAT scores your SAT scores like do you remember do you remember y or calculus you know high school calculus you know like no it's like these are all abstractions and metaphors that we create as huristics to understand something's intelligent but many times as you know in real life we don't grade people really by that what have you done that's a good what Can you do product?

Consumers ask like my take away my my big takeaway Thursday was product now and and and and the value you can deliver to consumers is the only thing that people should really care about in the short term and and OpenAI is in an amazing position here. Anthropic is an amazing position here.

They have you know Anthropics got an amazing enterprise business. Open AAI can sell a lot of subscriptions to a lot of people in the US and all over the world. And the real people that were in trouble seeing GPT5 were the people that were, you know, still kind of fundraising against this idea of machine God. Yes. Yeah.

Yeah. Yeah. Dex machine out of the dead. Like I I I'm thankful for that. I'm a Christian. Like, you know, people remember Tower Babel. It's like a it's a it's a it's a both a story to take lessons from, but I believe really happened.

And the the the the point is that uh that humans naturally are drawn in every moment of the cycle which is why that and recre quote I don't know if it's tribute to him but probably somebody else actually said it that we overestimate in the short term we underestimate in the long term I think the god complex is that and and if I go into the also the other kind of more tactical point is that many of these things that happen where people are using T85 if you go to grog four it solves it just fine so so it was just funny that people are like oh my god they can't do blueberries you feel like you I literally took that did in Brock and it dancer just fine.

So this seems like much more of like a tactical decision on Sam. It could be the fact of like Yeah, it's interesting people's desire to like gotcha the model.

So funny like they can they can completely ignore that it'll just like oneshot like a really crazy physics problem and then they'll like get it on some like technicality. Yeah.

And the whole point and that that everybody should internalize is that the models are really smart at some things and then there's going to be edge cases and then where they they mess up just like humans do and then there's going to be areas that they they're not capable at all and that's totally fine.

It's still incredibly valuable. Exact. Exactly. Like like I I try to simplify ways what AI is going to be which is computing that talks like a human.

It doesn't mean it is a human just as if like there there was obviously the famous uh I love this analogy because it shows how much humans are not aware about how signals and intelligence has developed.

So you remember back in 70s and 80s like they said oh like monkeys can talk right they taught in sign language and then they could repeat it.

So they ran the test again and this time they had no facial expressions and they tried to just limit all human human interaction to to uh to to the ape and they couldn't do it because what they realized was actually the the the apes were responding to un known human cues that the that the the researcher was delivering to the ape.

the right they were just mimicking what they thought would actually thing they want but they couldn't actually create discrete knowledge and action intelligence about what the actual language was being was being delivered and and and so that that's where it goes into like when you look at something so much about because humans are so amazing at being able to create it's part of the the original genus argument appeal to the land name the animals we're amazing at that because because it's it's either believe like a god-given talent that we have that only we have as a species but we're so good at it we have no idea how we affect other people or other animals via that that projection that as a as another basic scientific book you're married as another like very simple thing that most people don't understand about the way they pick partners much about the way you pick a partner is driven by the pherommones of your immune system and they've done studies of this where you where they basically created a blind test where they took from different you know like opposite the opposite sexes and they basically gave it as as to the opposite sexes like male and female like which one smelled better and they found that the one you thought smelled better had the most different imological profile than than commonality and that's because of the desire to have a children children that had the greatest amount of protection from disease.

So, so that's so like you know we can go into blondes brunette blue eyes like hi whatever but so much is driven by smell and we're not even like aware of it and and so we're here being like hi and oh like we got to create smell vision first we got to create fund the company right now smell for AI so um for sure last week there there's been different headlines rumors around negative gross margins in various SAS SAS companies specifically in the in the codegen space does that you know that the the reason you would invest in a company with negative gross margins is you would expect inference costs to decline dramatically and and for them to generate you know positive gross margins over time and right now you just need to acquire as many customers as as possible and create products that people love.

So that that that investment case makes sense and then you have GPT5 come out which has you know a significant cost reduction. Uh I understand like cursors are going to be 15% better in the next administration potentially.

Um potentially but but how how are you thinking about is that uh you know and and the um journos are are freaking out about it saying like oh this is all like I saw that rune post I know exactly what you're talking about.

Uh but but what's what's your view um in terms how fast how fast can we actually drop inference costs over the next couple years because it does feel like if it takes a decade to to to you know get an oo of savings like an order of magnitude of savings like we might be in trouble financially. Yeah.

And if but but at the same time if you if you I I think like researchers at the labs have been focusing on raw intelligence more so than efficiency. And so if we're hitting if we're hitting a sort of plateau intelligence wise, the next thing that makes sense to do is like really optimize efficiency. Totally. No. Yeah.

No, I I think I think Jord hit it on on the head, which is that there's two trends. There's the effectiveness trend and the efficiency trend. China is rationing us on the efficiency trend. Absolutely demolishing us. And this is the whole thing.

We we we give them all these Nerf chips and we limit their chips so they just get incredibly efficient. It's like shots in the foot. I mean, bring it back and then inference that. Yeah, there's a lot we can learn. So, so yeah, exactly. Exactly.

So, so I I I think that there it all develops into kind of we don't really fully understand customer preference theory as it relates to these tools and that what is the actual first pass the goalpost sort of goal to to pay versus free. But in terms of the brute natural law that compute gets cheaper, that's bulletproof.

So, that will continue. And and so at what scale, like you said, does it actually frost that? I I I think the bigger concern is more of these mega models uh that that they just have the ability to absorb the cost, be able to spread it across different things.

And if you think about it, it it's a little bit like how Apple will continue to invest in its own apps that you remember when you first got the iPhone, you had a camera app, right? You had calculator app and then they kind of replaced it.

That's the bigger concern for some of these tools is like a mega model be able to absorb the cost like like Google which has a crap ton of code or like maybe llama open source as a coding agent which seems pretty pro probable that that becomes like more of the go to like basically compressing even further.

uh so they have to be able to create the data stream like a I think it was a cognition that's trying to do that where as they uh engineers you know put stuff in in into the agent then it basically re reinforce self-learns from what the act doing that's probably the best path to create some significant amount of of of margin but you should every company can rely on that it's going to get cheaper but as Jordy said there's some protectionist stuff returns our national policy that has prevented that and and I made this argument on another show where I said that well funnily the US government this was pre Trump was applying protectionist measures to these large scale model companies and he like didn't understand that he goes well what are export controls like and like he just didn't think through like the economic consequences of that and that means in our country we have an over supply versus what we actually would pay for versus what the world would pay for which is a protectionist measure which means that you don't have the same market forces being applied to our frontier labs as you would in other countries as you already Yeah, we send them Nerf chips.

That's the cap. So then they then go down. Like that's their only direction they could go.

So if we actually have a more liquid supply, more market driven supply where things can be priced in, then you're going to see OpenAI start doing more open source, more of this downward cost efficiency because they're now exposed to it.

So it shouldn't be lost on us that open eye is pursuing this idea of like how do they add these mega releases cheaper models is because they're starting to be more exposed to these like cost structures that is needed for us to actually be competitive on the efficiency side.

But if we continue to go down this effectiveness range effectiveness is generally like a B2B problem like where you're company versus company. So it's not really like adoption you're going after like we just got to be the best company and and efficiency is the opposite.

Efficiency is like how do I make this as cheap as possible for someone to adopt? Javon's paradox like that that's the direction. So I think over the course of two years the next two years I think you're going to see a a tilting into lowering cost making things more efficient.

So there's going to be multiple times in the show where I go on where people are like comput's so cheap no more wants to buy more compute and like the remember deepseek it's like this that's going to be kind of what continues to happen.

Uh but it's good for it's good for the world like we we want comput to be as cheap as possible and on a token basis and that means optimizing for like hydro energy optimizing for tax regime reducing construction costs reducing permitting all those great things that would make everything cheaper because the more cheaper we get the more people we use and and AI is such an innate good for the world that is something that America should be proud of that we can share with the entire world and the only way that's going to work and they got to make it super cheap to run in Africa.

Not launching a Stargate in, you know, in you know, freaking Tanzania to like that's not going to go anywhere, right? But you have to make it to where it's capable of actually being scaled across the world. Okay. Lightning round from the chat. Pancakes, waffles, or French toast.

I'm keto, so I don't need any of that stuff. None of that stuff. What skills do you think will be most in demand in AI and data infrastructure over the next three to five years? Uh, many of the things that Trump ran on or to still remain true.

electricians, plumbers, building something like the raw work with your hands sort of work as well as there's going to be new categories of jobs that we don't even fully comprehend yet that are going to take that salary of engineer and like it's going to go there. Last question.

What was the most who was the most important hire you made in the early days of your company? Um like what do they do? What do they do? Probably my CFO, general counsel. There you go. Uh do you expect any data center project like a lot of data center projects to blow up over the next few years?

people um there's been some interesting commentary interesting commentary people are basically credit funds or well well so people are like basically forecasting the useful life of their GPUs and then separately what like what they'll actually be able to sell uh inference at and ultimately if either of those equations are out of whack at all you can be completely underwater and not be able to service your debt.

Yeah, I would I would say that in and maybe this for our next time GPUs are not overbuilt, but data centers will and and I lean into not like I agree agree Jordy that the cost goes down and changes in vintages but the key thing is like look at what the money is doing.

There is very tight control still on like getting loans for GPUs. There is almost none for for data centers because they view the use of life 15 20 years.

But as Jord just said is like if the square footage profile of power density versus what you can get is not structured correctly as well as Nvidia is actually making less GPUs next year. There's going to be a lot of places that are just there and nothing's in it and like in China like we see in China.

So GPUs in China are like massively being utilized but there's lots of data centers that have nothing in it or like just can't actually run it. So I think data centers are far more likely to be overbuilt and be a bubble than GPUs because you know one company controls the GPUs.

you generally from market forces don't really see that be a bubble because it's a monopoly generally and and data center lending is very aggressive because they view it as a 20-year asset and the GPUs are short term so there's just not as much money flowing into the GPU side as much as the data center side makes sense thank you for coming on always a pleasure thank you for coming on always guys we'll talk to you soon public.

com investing for those who take it seriously they got multiasset investing industryleading yields They're trusted by millions, folks. Get on