Sharp Tech with Ben Thompson - Nvidia Backlash and the China Conundrum

Episode Date: September 27, 2022

Why the internet is upset after Nvidia's Ada Lovelace announcement, the perfect storm of the past 12 months, a bet on AI inside and outside of gaming. Then, the calculus for Apple and Nvidia as tensions rise between China and the U.S.

Hello and welcome back to another episode of Sharp Tech. I'm your host, Andrew Sharp, and on the other line, Ben Thompson. Ben, how you doing? I'm doing well. I'm anticipating this episode because for our first recording post-announcement, which was sort of this, you know, you're going to have a more normie perspective, help me sort of explain tech without getting too deep into things. I decided to throw you into the absolute deep end by talking about like how graphics chips work.

and things along those lines, which I struggled to write. And it ended up being very dense and maybe a little overwhelming from my perspective, where strategy is already like that. So we're going to have quite the chore on this episode to see how well we can explain things in a way that makes sense. Yeah. No, I love it. It's funny you say that because earlier today, as I was putting together the outline for

this discussion, I was thinking to myself, we kick things off with a discussion of TSMC and the history there, sort of a mini history for our first ever episode. And that was sort of like an intro class, like what you take freshman year as you're learning about the semiconductor industry. And this feels more like a 200 or 300 level course, more dense and more complicated than what we were dealing with with TSM. Well, let's see all you do. We'll give it a shot. Let's dive in. And just as a reminder at the top, listeners can send questions or comments to email at sharptech.fm and we'll answer them on our second episode each week. And you can ask, which by the way, I could not have been more happy with how the first one that went. In retrospect,

I was a little nervous. So I actually had an Adobe take to start out with. And afterwards, like, we should just have done all Melbaid questions. This is, this is super fun. So yes, definitely send your questions. That one is a blast. But we have to get through the boring stuff first. So let's take our medicine. Totally. Ask us anything. And it doesn't have to be about what we've discussed. It can be about stuff you want Ben to write about. You're frustrated. Ben hasn't written about. This is your chance. No, the best question was the or the ones that Terryating you. Like, why are you a dirty thief stealing content? Exactly. The whole spectrum is in play with the mailbag episode. But for now, NVIDIA, you've written about them several times in

the past couple months, including a big article on Sretti this week. You also had an extended interview with CEO Jensen Wong last week. So for those who don't know, Nvidia is an American company and a major player in the microchip economy. And their stock, just for reference, jumped from $66 in March of 2020 to $303 by November 2021. And it has since come back to Earth and is down nearly 60% on the year. So quite a ride for Nvidia over the last two years or so. And as for last week's news, I'll start with this note from Bloomberg. Invidiacorp, the most valuable semiconductor maker in the U.S., unveiled a new type of graphics chip that uses enhanced artificial intelligence to create more realistic images and games. Code named Ada Lovelace, the new architecture underpins the company's G-Force RTFRTX-40 series of graphics cards.

The top of the line, RtX4090, will cost $1,600. and go on sale on October 12th, other versions that come in November will retail for $900 and $1,200. So you sent me a Twitter thread over the weekend highlighting various gaming influencers who have been irate in the wake of NVIDIA's announcement last week. It looks like NVIDIA is currently going down in a hail of Reddit memes. And so I want to get to that. But before we do, I want to sort of anchor. us, why should anybody who doesn't play video games care about technology advances in the video game space? That's an excellent question that is probably, you know, an answer that we can

give to benefit all our gamers in the audience to justify, you know, the hours they spend working on this stuff. But it's funny you mentioned the TSMC history bit because I think an interesting way to think about and understand Nvidia and graphics chip is it's one of the earliest and first manifestations of the entire, the reason why TSM was so impactful. And it used to be you just had these centralized processors, you know, Intel obviously being the dominant player there. And they did everything. And so they would not just process the game, for example, but they would also process the graphics

and they would, you know, put them on screen. And it turned out that, you know, this job of processing graphics, number one, there's this bit that was going to come up again again, where it's sort of this highly paralyzable job, embarrassingly parallel is the term that's used. And what that means is, you know, if you take a screen, you cut it up into a bunch of little squares, you can have one processor that's working on each of those little squares. And the more processes you have, the smaller each square is, which means the faster it can do it. And there's no real performance penalty for doing that.

That's why it's called like embarrassingly parallel. You can do the same operation all at the same time instead of sort of sequentially or serially. You can do it in parallel. And it turned out to get a chance. that was dedicated to doing this made games way more performance, you know, made it possible of the entire 3D revolution. And all this, though, was predicated on being able to actually manufacture these chips. And so in a world where there was like CPUs and you sort of had to make your own thing,

well, now there was a place where you could just design the chip and you could get a company like TSMC to make it. Now, today, there's a whole host of custom chips that do all sorts of specialized functions. And there's a bit about the Nvidia story where that's actually now a threat to them, which we can sort of get to in a little bit. But graphics chip were the first real manifestation that I think resonated with people broadly where having a specialized chip to do a specific job is just way, way, way, way better. And you see this tradeoff in technology a lot where, you know, the advantage of a generalized processor is it can do anything. And you put software on top of that. And software is sort of infinitely malleable.

And so you can design it to do anything, and then the chip can handle it and it can sort of run it. But if you ask something that's dedicated, that's meant to do just one thing and to do it very, very well, number one, it's going to be faster at doing that thing. And then number two, like, it's in some respects easier because you're writing to a specific sort of job. I mean, the easiest sort of goes back and forth. The one hand, it's easiest to not care at all. You just write the program and the processor figures out. But there is an aspect that's dedicated to do this one thing and it does it better.

And now there's chips that do all sorts of specialized functions that, you know, if you go back to a computer in the 80s or 90s, it was just the one chip that sort of did everything. Well, and just to jump in on that front, as far as specialized functions, when you're connecting the dots to TSM and saying that TSM sort of changed the ball game there, what you're saying is that because TSM only specialized in manufacturing, they were able to offer a more customized manufacturing process than had existed in the past. Is that right?

Yeah, that's right. Well, that's part one. And then part two is you could create a new company where all you did was create specialized chips and you didn't have to create your own manufacturing operation because TSMC already did that part. And so the barrier to entry was dramatically lower because you could just work on designing the chip

knowing that TSM could make it. I mean, you drew the analogy a few episodes ago to AWS, which is the exact right one to draw. Like, if you want to start a software company or a cloud company, you don't need to build servers, right? You don't need to, like, get them up and running. You can just use what's already there, and that's exactly what TSMC enabled here,

and graphics was one of the first sort of manifestations of this. Now, as far as why it matters, I mean, there's definitely a bit where people like games, and like computer games are something completely new that wasn't possible before. And so you've seen games be on the cutting edge of lots of technology. So, you know, games started out just like text-based games. And, you know, maybe, I don't know, do you play Oregon Trail when you were here? Are you too young for that?

I was a big Oregon Trail guy and a big Wolfenstein guy. I might be getting the name wrong there. But Oregon Trail all the way in my household. And then Sim City. Yeah, I was a computer game kid until I grew up. I would say like 13 years old, I moved beyond it, but sure. That's why you're boring today. But then you go to, you know, so Wolfenstein was interesting because Wolfenstein sort of simulated 3D.

It wasn't a real 3D game, like through some really neat tricks. They made it feel sort of 3D. Doom was really the first 3D game. And there was this chip that came out called the voodoo card made by 3D FX. And this was a wild card because you would put it in your case. computer, then you'd have a little like two-inch cable that took the output from your regular graphics card that put it into the voodoo, which then like did its 3D stuff and then pushed it out. And like that was the first accelerated card. And Doom was sort of the first 3D game. And

Nvidia came along and they produced the first sort of all in one graphics card where it did both 2D images and it also had a 3D processor on board. But in these early days, and I'm going to try to see if I can do this big history, which I think is important in about two minutes. Let's see how I do here. Okay. In the early days, 3D cards, there were specified instructions in how to render 3D. There was the main language at times called OpenGL. Then Microsoft came with their own called Direct 3D.

That later became known as DirectX. And they had specified instructions in how to render. And so these cards were custom built to those instructions. They would take that specific instruction and they would execute it very quickly. That was how graphics cards work. InVIDIA around 2000, 2002, somewhere around then, they come up with a new concept where instead of their card being custom

to these specific commands, they made it slightly more generalizable. And they had all these sort of like cores that were very fast, but they made it so you could program those cores individually. And these little programs you would put in these cores were called shaders. These shaders then were programmed to run those instructions quickly.

And what was really interesting about this was it was slower to start. And it actually almost killed Nvidia as a company because they were spending a lot of money to build a more complicated card that was slower in some respects. But once it was programmable, you could program it to do more than those instructions. Now, Nvidia was thinking, oh, we can have future 3D instructions and you can upgrade a chip in place. Like the old cards, once they were set, they were set. If there was a new OpenGL instruction or a new direct X instruction, they couldn't run it. But the new ones could be upgraded, number one. But then number two, it turned out they could be programmed to do anything.

And, and Nvidia realized, well, okay, there's this capability to have this embarrassingly parallel capability where you do the same thing multiple times altogether. But no one can actually program it. So we have to build an entire programming environment around it. And that was called Kuda. And there was a compiler, an instruction set. And it turned out that a lot of the stuff that goes into AI and machine learning is doing the exact same thing, multiple times in a row, right?

We talked last week about this pulling an image out of noise, right? That's just doing the same instruction iterating on it as fast as possible until you kind of figure out the right direction and then honing in on that one thing. That is an embarrassingly parallel process. And it turned out that these Nvidia cars with this sort of multiple cores and the shader architecture were very well suited to any problem that's embarrassing parallel, particularly machine learning. And so, Nvidia has become a dominant player in the space where half their revenue is still gaming, but the other half is data center.

And that data center revenue is mostly all people doing this machine learning. And it runs on sort of Nvidia chip. So there was this real transformational moment where they became about more than just games. And they now call it accelerated computing. And the idea being, you know, Jensen Huang is a big, you know, believer that Moore's Law is dead. This idea of we're going to get this free increase just by chips getting smaller and faster is slowing down dramatically and, you know, isn't going faster. So to get faster, we have to build different approaches to computing. And one of these is leveraging these chips.

He wants to go even further up market to where there's entire systems that do this sort of thing. And that's why this keynote, you know, one reason why I think gamers got put off before we even get to the price stuff is there was like 15 minutes on this new graphics card and a couple of, game demos and then the rest of it was like an enterprise keynote. And I think there's some aspect where it's like, I thought you were a gaming company. But they are much more than that. And it's almost like they backed into this thing by going to the shader model and then building this ecosystem around it. And that's a bit how we get to today.

I think that was more than two minutes. But I feel like I did pretty well. No, that was a very helpful summary. That's exactly what I was looking for. And I'll be completely honest. One of the reasons I need that context is because I was sitting there over the weekend reading some of the dramatic reactions to Nvidia's announcement. Like, Nvidia betrayed us. Do they really think they could get away with this?

And as I read all that, it's hard for me not to be kind of a snarky asshole about it all. Like, really, no person over the age of 25 or 30 should care this much about computer games. Like, I have some self-spects. Hey, we have to build up our audience before you start driving them all away. I'm only here to share my truth. But alongside all that, it's helpful for me to remember how much gaming has driven the entire industry forward and the entire tech industry that is. And the progress in gaming eventually trickles out to mainstream society. And it sounds like that's what Nvidia is hoping will happen here.

So I guess my question is right now, who are Nvidia's most important? customers because they have gamers over here. I've read about hyperscalers like meta, alongside hyperscalers in China that they can no longer sell to. And then they used to be supplying chips for Ethereum mining and that's no longer happening. So I know this is a pretty abstract question, but short term, long term, who do you think is most important to their business? No, it's actually a really insightful question because I think you cut right to the nut of what is challenging for NVIDIA right now. So their bread and butter has always been gamers, right? And you look at that, you look at their financial results and their gaming division is still huge.

Now, part of that is some of the crypto stuff, which we, you know, we can touch on. But that wasn't reflected in the keynote, right? The keynote, the gamers were kind of shunted to this little bit in the front. and then everything else was about all this other stuff that they want to build, the sort of accelerated computing thing. And I think there is a real tension there where the financial results of their business today are not necessarily aligned with their vision of where they're going. And that causes problems.

Right now it's about 50-50. And I think a challenge that you're putting your finger on is the, customers that Nvidia has in 5, 10, 15 years if they're successful may end up looking very different than the customers that they have today. And making that transition is what's going to be a problem. So gamers, you know, gamers for sure are still super important. And in VDIA, you know, for all the KVecching, no one is denying that particularly the 4090, the top of the line card, is without question the best going to be the best gaming card. Like there's no, there's no dispute there. There are questions as to whether there are any games that can take advantage of that,

to a way that actually matters. And that's an aspect we should get into as to why that might be the case. But on the other side, on the data center side, it is these big players. It is, you know, the Facebooks of the world, the Amazon's of the world, big companies like that. But that's where this bit about making custom chips is a problem for Nvidia. Like their chip was much more suited to machine learning because it was massively paralyzable and programmable in a way that a CPU wasn't. But what's even better? Well, Google makes their own machine learning chip.

That is not just tuned to being paralyzable, but also perfectly tuned to Google's own software stack. Like it's, you know, this idea that we talk about with Apple, where they integrate the hardware of the software, that applies to other places as well. So Google's built their own chip that is married to their software. And it makes perfect sense for Google to do that. Meta is working on building the same thing. And there's two aspects. Number one, it's that much more perfectly tuned to what they do.

So they can probably get more performance out of it. And then number two is invidia charges a lot, right? And this is a thing that I appreciate about Nvidia. It makes it risky in some regard, but I think they're continuing on this path, is Nvidia really has driven things forward. They have been a very innovative company. and they make up and they charge for it. Right? Like people complain about it, but they still pay because they have the best chips.

And again, I've long maintained, Nvidia and Apple are actually very similar companies. I was just going to say, that sounds like Apple to me. I mean, I'll pay anything to Apple because all of their products have been so high quality over the last 10 or 15 years of my life. Yep, and that's kind of what Nvidia is too. And actually, very famously, Apple Invita do not get along. Apple has not shipped an Nvidia chip in like over a decade. and there's a lot of griping about that because, like, I mean, gaming on the Mac is a disaster in general, but particularly when gaming, like, if you had an Intel Mac, you could boot into Windows,

but then you couldn't get an Nvidia card, you know, like, so you had to use AMD, which is, which is, you know, kind of second place as far as that goes. And I think a reason they don't get along is they're very similar, right? Like, they're both like, we're going to tell you how the world works. You're going to complain about it. You're going to whine about the price. and then you're still going to pay it because we make the best stuff. And it's kind of almost admirable in a way, right?

They just lean into it. Like it's like, no, like we know how it's going to be it and it's going to be the best and you're going to be pay for it whether you want to or not. Okay. So as far as their pricing is concerned, on Friday of last week, I read an article headlined Ada Lovelace GPUs shows how desperate Nvidia is. And that was from Dylan Patel at semi-analysis.

And he was writing primarily about. the pricing and like the opportunity this creates for AMD because Nvidia chips are going to be so expensive and there's not really value being created alongside the extra expense I think was some of the argument there do you share that view I think desperate might be a little strong so one of the big controversies here is the top of the line chip I think everyone agrees like they they want charge six hundred dollars for it it's like an incredible chip it has everything does everything I don't think people generally are really griping about that I

I think it's more the, so they have the 4090s, the highest, the 4080s is the next level. And they have two 4080 chips. One is 60 megabytes memory and one's 12 megabytes. I was like, oh, just a little bit less memory. You know, I don't run super high resolution screen. Like that's fine with me. I'll take it. Turns out no, they're totally different chips that are on there.

Like lower memory bandwidth, less capability. And that 4080 would have been branded a 4070 previously. Some people would argue even a 4060. So that like that's how they sort of like tier their chips. And the problem is that. the 3070 was like 500. I'm pulling out of the top of the time. I'm like $500 or $600,

whereas this new one is $900. And so there's a lot of surmising that Nvidia is playing games with the naming to hide the fact that they raise the price of this chip by like $300 and they don't want to sort of like admit to that fact. So number one, there's a lot of griping about Nvidia sort of being deceptive and, you know,

tricking people into saying this chip is better than it is. So that's sort of like some of the YouTube comments. What Dylan is getting to in this article is there's a direct connection between size and cost when it comes to making chips. So you buy these wafers, right? Like it's pretty crazy. Like they grow these silicon crystals out of like 99.99% purity that are these long, you know, and so they're 200 million or 300 million liter wafers. And then they slice them super thin.

And so there's a couple of functions of why size matters on how many you can put on. on that wafer. So number one is the larger a chip is, the fewer you can fit on there. And it's not just fewer in terms of taking up area, but you're putting square chips onto a circle. So like there's going to be more waste sort of the larger something is. I mean, this is like going back to like geometry or something on those lines. So that's number one. Number two, the larger a chip is the more chances there are that there's going to be a flaw on it. Right. So there will be flaws in the manufacturing process on the wafer. They'll just throw that chip away. Well, if the larger a chip is, the more likely part of that chip will be flawed,

whereas if there's smaller chips, the more likely a chip can sort of avoid a flaw. So there's, and this comes, this becomes a yield issue, which is your sort of profitability in manufacturing is how many good chips can you get per wafer? And if you have to throw too many away, sort of your cost structure starts to fall apart. So one of the challenges in Vida has is they're making a really big chip. And one of the reasons they're making a really big chip is they're not just adding on that traditional graphics functionality and that shader functionality I talked about, they also have this AI functionality,

and they have dedicated parts of the chip that all they do is the AI parts. And in this case, the AI is for gaming where, like, they will actually predict the next frame and they'll pre-render it, and so you have to have, like, calculate it every single time. You think about if a game 60 frames per second, the next frame is to be very similar to the previous frame,

and if they can understand what's moving in that image and sort of pre-render it and only need to make small changes, they can get more efficient, right? And it works very, very well, actually. It's quite impressive. They also had this bit about ray tracing, which is the way we, you know,

how they calculate light. And they can do it sort of dynamically and on the fly the way they do for like a Pixar movie. But Pixar movies, they render, it takes days to render it. Right, right? In this case, they're doing it in real time.

So it's not as full fidelity, but the results compared to a traditional model where you render a scene, then you lay over like a shadow map or a white map on top of it, and you calculate the light. It works, but it's a hack. Well, and also, the comparison to movies is probably instructive, right?

Because movies, you've got a set runtime and a set universe that you're having to render. And with some of the video games that Jetson Wong was talking about during your interview last week, I mean, the possibilities are multiplied like a hundredfold or a thousandfold in terms of what needs to be rendered. Oh, it's basically infinite. Yeah, exactly. Like if you want to play like your traditional like World of Warcraft or something, right? And you go in and you destroy something. The next person that logs on and goes there are the things back, right?

Like it's not there's no persistence. If you want to have a real online world where things you do last and then people go there and it's different, that doesn't work with the way games are rendered today because there's the light would all be wrong. Like all the textures would be wrong because they didn't pre-calculate like everything's pre-calculated in some respects. And this is one of the reasons why games have gotten so expensive as the resolution's gotten super high. All that stuff has to be drawn by hand. Whereas this new model where you're dynamically calculating the lighting and the shadows and all this sort of stuff based on the viewpoint of who's looking at it. And like you trace like where does the light go from the eye bounce off an object and does it hit a light source?

Then you cut like it's pretty wild the way the way this works. To do this in real time is really, really expensive. So Nvidia has dedicated parts of their chip that. all they do is do this. So, Nvidia's leaning into this, right? They're actually making even more specialized parts of their chips to handle this bit to create this new possibility. The problem is that all these extra pieces of the chip make the chip bigger. Making the chip bigger makes it more expensive. And AMD, they have ray tracing, but it's like built into their shaders. It's kind of like a half-ass version of it, right? It's enough to like support it and so they can do it, but it's

It's nowhere near as good as what Nvidia does. But the payoff is AMD has much smaller chips. And what that means is their cost structure for this generation, their chips are coming out next month, but we do have leaks of what the specifications are, is likely going to be lower than Nvidia. So you're going to have a situation where their chips probably aren't going to be quite as good.

They're going to be very good. And AMD's been making good progress. But their costs are going to be a lot lower. And Nvidia, meanwhile, their costs are getting much higher. And they're getting much higher, not for existing gaming functionality. Exactly. But for games of the future. Paying the higher cost doesn't seem worth it.

Like the juice isn't worth the squeeze because there's not enough games out there that will allow you to take advantage of this crazy, ambitious technology. But here's the thing. And this is like this is why it's worth looking back at Nvidia's history. That's what they did with shaders. Shaders were not necessary. And ever at the time is like, why are you doing this? You're killing performance.

If you just did a dedicated graphics chip that executes instructions, it would be cheaper. It'd be higher performance. It'd be better. In the long run, though, that bet transformed in video from just being a sort of commodity chipmaker to being like that entire keynote, all this accelerated computing, all this machine learning, all this huge ambition that's in that keynote is all downstream from having a near-death experience of introducing a new technology that no game supported,

that nobody understood why they were doing it, that they paid a performance cost, they paid a margin cost, but it paid off in the long run. And that's why I find this very, very intriguing because it feels like a little bit of a repeat. It's going to be very, very painful for this generation because they are going to be

at this cost challenge versus AMD. But if we do get to this future, because the other thing about ray tracing is not only do you, get better, more realistic sort of lighting. But also, number one, the same concept of ray tracing applies to physics, right? You're just dealing with, like, how stuff moves around in an environment. Like, if you want to ray trace, you need to know, is this light hitting a rock? Is it hitting grass? Like, that affects how it bounces around. That's our, you can apply that to how things interact

in the environment, right? So you can have these crazy realistic simulations. And more broadly, if you want to get to this world where you have this deeply immersive, persistent environments where people can experience completely new things, not because it doesn't have to be pre-rendered or pre-wit, but can be done dynamically. It's actually, you actually get this lighting for free, right? Because instead of you having to draw all these chips, the chips figuring it out for you on the fly. Why does it work? Because our processes are so fast, it doesn't matter. And Vindia is making this bet that we can make our processors so fast and so. good at this stuff. It actually makes gaming more accessible, more open to developers, because

we're taking care of a lot of this stuff that had to be done by hand. And when you say gaming is more accessible, more accessible to people who are making video games, not people who are paying $1,200 per chip or whatnot. That's right. Because I mean, gaming's in a little bit of a, one of the problems that Nvidia had is they didn't have any good demos, right? Usually you want to have a new chip. You want to have these new games. It's another thing people were pissed off of out, absolutely. There's this huge cost problem in gaming, which is, again, all this stuff has to be done by hand. Like, there's a reason, like, a lot of the big gaming companies are in Eastern Europe, in part because the labor is cheaper, right? Because you have to actually create all these assets

that are, like, everything's hand drawn in the game. People don't, I think, appreciate this. Like, it doesn't just come out of nowhere. You have to act. And so you've had this dynamic in gaming where all the cost used to be in the actual, like, code, creating the game, and then you throw some assets on there, right? Like those, you know, pixel art or whatever it might be. Today, all the cost is in the asset creation. And so to the degree, you can take away some of that cost, like, for example, doing the lighting, doing the shadow maps, like the degree to which you make games more accessible.

We're in a world today where you're either a AAA game that takes years to make and has to recoup its investment. And so you get tons of sequels. Like, it sounds like the movie industry, right? You get a call of, you know, call of a call or God of War, whatever it is, like number 47 because there's so much investment in asset creation, they have to get that money. They have to get that money back. And on the other hand, you have indie games.

What are all the indie games? Like, oh, look, we did this cute pixel art. Why did they do cute pixel art? Because it's easy to make, right? They don't have the pocketbooks to make anything sort of bigger and more realistic. What Nvidia is, what is hopeful about this offering is what if you could get compelling immersive 3D environments without necessarily needing all that asset cost? And this is where the AI bit comes into, not just the pre-running of frames, but what if AI is generating those textures?

What if it's creating sort of stuff? And then you can white it all. And so there's where they're going, I think, is pretty clear and it's pretty compelling. But that's why I framed it as like, you know, Pilgrim's Progress style, you have to go through the valley of the shadow of death because it's going to be tough to get there. Yeah. And can I give you a sports analogy? I want you to grade a sports analogy that I came up with thinking through this, this, this, weekend and then reading your article on Monday. Is that cool? Absolutely. Okay. So the way I see this,

if all of this were sports, I would say that Nvidia is the quarterback and the market is the receiver. And instead of Nvidia throwing the ball to where the market is now, Nvidia is throwing the ball to where it thinks the receiver is going to be in four or five steps. And what's tricky about that is sometimes the route will unfold perfectly in a way that only the quarterback could see and the QB will look like a genius. It's like magic when that happens. But then there are other plays where the receiver trips and never finishes the route and there's no one within 10 yards of where the quarterback through it.

And so that sort of seems like what we're dealing with here. And Vinty it definitely has an idea of where all of this is going to be in four or five years. but I'm curious, like, what happens if the AI takes longer than we expect, and they are just sort of twisting in the wind indefinitely? Yeah, well, they need people to actually make these games, right? They have a dependency on people, and people need to leverage rate tracing, right? And, you know, and it's as, you know, the more share that AMD gets in this generation because they have a better value, for example, in some respects, works against this long run

because AMB's ray tracing isn't as good. So there's going to be less of a motivation sort of in, best in doing that sort of thing. And I don't know. Like it's tough because at the same time, you have this whole data center division that mostly sells to these big companies that are, you know, that are going to make their own chips. And so Nvidia needs to get everyone else on board, right?

They need to get the enterprises using AI. Like one of their big announcements was a partnership with Delight, which sounds hilarious, but it makes sense because what does Deloite do? It helps people who have no idea what they're doing, use technology. And they're saying, like, look, this is. is good enough for you to start using it. But there's a real business need because they need to backfill the demand for their data center stuff as these hypers move away and build their own chips.

And so that's a similar case where they also have an external dependency. They need enterprises to grok the value of AI, to leverage it, to apply it, to bring into their business. Now, I think that, again, like gaming, I think that value is real. And I think companies that get it and apply it are going to be able to make a whole bunch of money doing it. But it's hard to get that started to get that off the ground. And it's just like they have all these pieces where they're really investing heavily to your point and where the receiver is going to be. Or as any Apple fan would say, where the hockey player is going to be since Steve Jobs quoted Wayne Gretzky about skiing where the fuck is.

I'll have to read up on that. Oh, no. That's from the that's from the iPhone. You know, people are going to kill us here. And they'll probably make it. but it's definitely going to be hairy. Yeah. Yeah, well, and I want to know what it looks like if they make it also,

because in his interview with you, Jensen Wong alluded to all kinds of different use cases for these chips. It was honestly a little bit overwhelming, try to keep track of everything. Like you had the omniverse, the metaverse, autonomous driving, the aforementioned partnership with Deloitte. Like, there's just a lot going on. So beyond gaming, What use cases do you find most compelling and what's a realistic timeline for some of the more ambitious use cases? Well, all this AI stuff is is all running on video chips, right? So to the extent that a lot of the things we talked about last week, whether it be, you know, generating law briefs or stock photos or whatever, all that sort of stuff is a big market for Nvidia.

And, you know, the more, if you're a Facebook or a Google, of course you're going to design. your own chip. But that's a still a very major investment. It costs a lot of money. And can they do it well? I mean, that's another question, right? Yeah, I mean, I think there's a this is more of sort of a known problem. Like, like it's, you're very clear what you're designing for, what you're what you're, what you're building for. It's, you know, what makes Nvidia really compelling and accessible is, is Kuda. It's the software stack on top. And there's a lot more people that know how to write software than know how to design this way the right chip. And also, once you make a chip, it doesn't change, right?

That's like, so if you screwed up the chip, you're stuck with it. And whereas if it's software, you can change software, right? I go back to that example of VD making 3D accelerator chips that follow the exact instructions. If a new instruction came along, your chip was obsolete. Whereas, you know, once you had shaders that were programmable, the chip could basically be upgraded in place. So the reality is for the vast majority of use cases for AI, it's going to make more sense

to do more work in software. and have a pretty good chip that's fairly generalizable, which is where Nvidia is. And it's really just the biggest companies where the investment will make sense to sort of do their own thing. And that's why their, their, their keynotes, like, there's a big part of it. It does feel overwhelming. But Huang's job here is, it's a developer conference.

He's trying to inspire people to create all these sorts of things and all the things that might be done because, InVIA's like, we'll give you as many of the tools as you can to do it, but we need you to actually go on. Yeah. So as their pie shrinks, if meta or Google peel off and build their own ships, their thought is let's grow the pie by bringing in different white collar industries that will come to rely on what we do best. And suddenly we'll have gaming and, you know, 30% of the American economy.

Right. No, exactly. And that's why. And the great thing is this is where they do have the advantage. of being a generalizable chip and doing it all in software, right? How can they support all these crazy use? They talk about the cars. They talk about robotics.

They talk about like the AI stuff. Like it's just like overwhelming. But it's overwhelming because all that software enabled. All of this is enabled by one chip, right? And this is the power of software. Software is infinitely malleable. You can make it do sort of anything you want. And so Nvidia's broad array of offerings and a keynote that's sort of all over the place

is downstream from having built up this CUDA ecosystem on top of the chip such that they only need to invest in this one chip, right? It's like there's the seed or the root of the tree, which is this chip or the trunk of a tree, and then all this stuff branches out. And that's sort of the bet. And I think it's a reasonable bet, right? Like, you know, go back to like servers at WS. Like there is an extent where if you build your own custom servers that are perfectly tuned to your use case, it will be cheaper and it will be more performant. But that costs a lot.

If you get it wrong, it's a big problem. Yeah, I mean, that's why Amazon's so valuable is they allow people to start their own small businesses and websites. And like we have a more diverse internet as a result of Amazon web services. Right. And like who like if you spend a lot of your resources on your server, like is that your differentiator? Like no one signs up for your service because you run your own server. Right.

If anything, it's kind of a downside, right? I'd rather you be depending on Amazon like everybody else. So my final question before we move to part two here, as we look at the tree branches, this is just, I'm putting you on the spot here, so feel free to pass. But how far away do you think we are from autonomous driving, which Huang alluded to a couple different times? Pretty far. I mean, I think I've been a pretty consistent skeptic just because if we were, if there was no

driving and we were sort of creating it from scratch. And, you know, I think they would be much closer and much more accessible. There might be an aspect where we need like roads that are tuned to drivers. So there's like sensors around. I just think there's a couple challenges. One, there's so many edge cases, right? And so how do you deal with these edge cases? One, one possibility is you actually like program the edge cases, right?

But then it's very brittle. It can sort of break, right? Like you might not get everything. Number two is you go through enough simulations, where it weren't to handle it. But the downside of getting it wrong is so high, right? It's like, you know, like we all pay massive amounts of attention to airplane crashes and they make huge news, even though the number of deaths in airplane crashes relative to

like miles flown or whatever metric you want to use is drastically lower than like all the people that die in car crashes right now. And so people in tech will be like, well, look at all the people that dying car crashes. We can reduce that. The problem is that every single autonomous drive. accident is going to get be national news drive such a new cycle and make such a big thing that it has to be sort of beyond perfect um and so i i think it's probably going to take longer take longer than than people hope but i'm not super deep in the space but i think it's it's definitely

been a space where it's going to be here in five years every five years has been the pattern i was just going to say it's been five years away for 10 years And so I'm curious as to how long that timeline is actually going to play out because it's like, all right, there's so many regulatory challenges. And that's actually a really good point with the edge cases. Every edge case is going to be like a national referendum on whether this technology should exist and how to implement it. And there will be people lobbying to slow it down.

So that's certainly an obstacle. I appreciate that. Yeah, it is frustrating to folks. because it's like, don't just look at the math, but that doesn't work. Like, no one looks at the math. Okay. So to move to part two, I want to talk about the present day calculus for companies that are doing business in China. And this is going to be related to part one.

We'll kick things off with Nvidia here. At the beginning of September, the Biden administration announced restrictions on the sale of high-end NVIDIA processors to Chinese customers like Alibaba. and these restrictions, they appear to be designed to inhibit China's AI progress. I don't know if there's been like an official statement on the reason for these, but I'll read this from the verge on September 1st. It says the U.S. has not given exact details on what criteria it's using to target chips, but the A100, H100, and MI 250 all occupy the top end of the AI chip market. These systems are used to train a range of machine learning applications from

facial recognition to text generation and the biggest U.S. tech companies use them to create in-house supercomputers for R&D. Meta, for example, has built an AI supercomputer powered by thousands of Nvidia A100 chips. And then later on, they say in a report published last year, Google CEO Eric Schmidt claimed that the U.S. was not prepared to compete. in the AI era. However, other experts have said AI competition between the U.S. and China does not constitute an arms race and that such rhetoric is damaging to both diplomatic relations and the safe development of machine learning technology. So with that report as context, where we're projecting out and thinking about the long-term outlook for NVIDIA, how concerned should they be

about the future of the business in China. Like, Huang seemed to downplay the Biden restrictions when he was talking to you. But what's your read? I don't know. It's his job to sort of keep the everything is fine posture through all of this. It is his job. And the one thing you weren't talking to Huang is that he's like the most inveterate optimists in the world, right?

And so, like, everything's going to be great. It's all going to work through. Not a big deal. It'll be fine. So that definitely makes it challenging. I was surprised at how, you know, fairly blasé he was about it. And, you know, basically saying he is talking about like,

it's certain like issues on like memory bandwidth and interconnects and stuff that that's really the defining characteristic. You know, like it's noted here. It's the AI chips that are targeted, not like the gaming chips, for example. And that's the solution is they could use gaming chips to do this. Like it's still at its core the same chip. There's different, again, like technical piece. but that make them different. So I think that's sort of the solution for now is, you know, it can be worked around.

Like there are real, you know, this excerpt frames it as just supercomputers, but these chips are used in everyday computing operations, right? Like all like the recommendations, like what do you actually see? Like an e-commerce site, like, oh, you've been here before, show you stuff. All that stuff is increasingly driven by machine learning, which is all using these sort of chips to do. So I think I'm definitely a little more pessimistic than Huang. Number one, China is a huge market. Again, they're selling to hyperscalers. Some of the biggest hyperscalers in the world are in China. Number two, the pattern as far as restricting technological transfer to China has been

a ratchet where it goes in one direction. Yes. And there's some stuff that's locked up and then it gets more and then it gets more. And, you know, if we found out that, oh, they're buying all this other stuff. I mean, the goal of the order, I presume, is not to stop a Tencent or an Alibaba. It's the assumption that this is being acquired and used by the Chinese military. And obviously, in China, you'd never really quite know who's actually using anything, right? And, you know, we've talked to like, bite dance has the Chinese government on its board, right? Like, like, so who knows where this sort of stuff ends up? And so from an Invidia perspective, I think it's, it's definitely worrying. And it's also worrying not just from

a addressable market perspective, but a big part of what makes Nvidia chips so valuable and why they can charge so much is CUDA. It's a software ecosystem on top of it. And to the extent that China doesn't have access to Nvidia

chips, that means they don't have access to CUDA, which is going to be a spur to develop sort of an alternative ecosystem or more open source ecosystem, which will spill over from China to other markets, like decreasing the value. of CUDA sort of over time. And again, this will take time to see how it manifests, but it's certainly no good news

from that perspective. I do question these excerpts that act like there's, you know, no harm here. I mean, the reality is China is a rival. They may very well be a military rival. And this is the most powerful sort of technology in the world. Like, it seems, it seems, you know, it's hard to fault the administration. for having a concern about how this stuff might be used. And to the extent the restrictions are a problem for Nvidia now,

they could be a much bigger problem as the U.S. broadens the restrictions. And like you said, when you look at the last, even just the last 12 months, it's hard to believe that the announcement in September is going to be the full extent of restrictions we see. And so given how much business they do over there, I don't know if there's a solution to the problem, but it seems like a pretty big problem that exists alongside some of the other existential questions about what their future is going to look like.

This is probably a more urgent thing. Well, who knows? It may not be that urgent, but it certainly seems like something that's going to be an issue in the near term. That's a good way to think about it, right? Let's Grant Huang, the matter of the doubt, and agree that it's actually not that big of a deal, they'll find ways to work around it to serve an Alibaba, to serve a Tencent, to serve a bite dance, et cetera. It's still this cloud that is going to be hanging over the business,

right? And it's hanging over the business, both from a future financial perspective, where if more stuff gets cut off and from a long-term sort of ecosystem perspective, where, you know, the price entry is an Nvidia chip and, you know, they might have to make some hard choices. Is Kuda going to support anything other than Nvidia? Or are they going to let another ecosystem and be born. So it's not great. It's, and it did sort of pile on this perfect storm. We didn't mention the, the Ethereum bit, but Ethereum, like these crypto problems where you're solving these cryptographic challenges, they're embarrassingly parallel. You just run this, like you're running the exact same thing again and again and again. And Bitcoin is so valuable

and is so particular that Bitcoin has long since been taken over by specialized chips. Like, all they do is do Bitcoin calculations. Ethereum, though, was almost all in Nvidia chips because it was very well suited to their shaders. Inveda made the best fastest chips, and so a ton of of Nvidia cells were going to Ethereum.

Invidia tried to stop this to the extent, like they would put things in their drivers so that would look for Ethereum mining and try to make it not work. Then they got hacked, and so that source code got released. And so Ethereum, though, switched to this new system where they don't run those calculations anymore.

It's people who, called with Stake Ethereum. They basically like walk up on which Ethereum for the right to validate a transaction and then they get rewards if they're chosen. And that doesn't use any power at all. That's why Ethereum's power consumption dropped by like 99.8% or something crazy

because they're not doing those calculations anymore. So all these cards are now worthless on the secondhand market. Also, they made all these cards because there's a shortage of the pandemic. They all came to market this summer. And so they're trying to watch this new card and they have massive inventory issues.

It's a real, it's a real sort of perfect, again, perfect storm. And then you weigh around this china bit on top. And that's why their stock is down by like, you know, 180% or whatever it might be. It's brutal. The real twist of the knife in my eyes is having the Ethereum graphics cards now flood the secondary market and drive the price down even further or drive demand for new chips down even further. I mean, it just, it sucks. There's really no other way to spin it. But Judson Wong did a good job.

This is actually an interesting tidbit that did come out in the interview. Because a lot of people are thinking, like, how did InVidia not see this coming? How did they screw this up so badly? And one thing that Huang said was, well, one thing that's challenging is the lead time for chips is getting super long. And so, of course, you know, where he's like, you know, we have to make capacity calls like 18 months in advance. And if you rewind 18 months. So let's say everyone sees this coming and say June, right? you rewind 18 months from June

2022, it's the beginning of 2021. It looks like the pandemic is like it's still going on where sales are still crazy. And during the pandemic, like, Nvidia cards were selling for multiples of their MSRP because you had all the Ethereum miners trying to buy it because crypto was going crazy.

You had all these gamers who had were stuck at home and had nothing better to do, people with all this discretionary income. And there weren't enough Nvidia chips out there. So they put in an order for new chips which don't show up for 18 months. Then they show up, and this Ethereum thing's happening.

And the other thing with the Ethereum merge, it's like they've been talking about the Ethereum merge for like five or six years. And it's always been delayed and pushed back. Like it's a pretty crazy technical achievement. It's like changing a jet, you know, changing a plane from a propeller plane to a jet plane

while it's flying without it falling down. Like it's a very impressive achievement. And there was understandably skeptical. It wasn't going to happen. And so Nvidia's leaving all this money on the table. And so again, Jensen Wong, inveterate optimists. Like, no, let's make more chips.

Let's get it. And then it all hit all at the same time. And they're sort of paying the price. Well, they're sorting through it, but they're undeterred and ambitious as ever. So I have one final question. It's about Apple. Apple and Nvidia, yin and yang again. They're another company that has a lot of exposure in China.

and J.P. Morgan said in a research note last week that it expects Apple to move about 5% of its iPhone production to India by the end of the year, and 25% of its production will happen in India by 2025, according to that J.P. Morgan note. So from an analyst perspective, whether you're talking about Apple or NVIDia, what are the obstacles to reducing ties to China? Like, is it just short-term hits to revenue? Are there longer-term concerns as well? Because obviously, it's easy for everybody to see this as a red flag, but I'm not sure there are good answers for either of these companies. Yeah, I think we talked a couple episodes ago about China, you know, being that bodybuilder where, you know, they're all arms, no legs.

That all-arms is super valuable and irreplaceable and no one else in the world can do it. So there's this bit where the iPhone is the most. tremendous logistics is, you know, it's not just the best, best products from a business perspective ever. It's from a logistical perspective, it's one of the most amazing achievements ever where Apple turns out hundreds of millions of these things at tremendously high quality and consistency. And they do it in a relatively short amount of time and distributed it all over the world and it mostly just works. Like it's really, it's really impressive. And it's very deeply rooted in and interlocked in in China.

I mean, Tim Cook made his bones by shifting Apple's manufacturing from the U.S. to China. And in iPhone, the value of the parts are from all over the world, right? The chips, obviously, from TSM. The software is from Apple. The glass is from like, I think Pennsylvania according. There's bits and pieces that are from Germany, from Japan. If you go through the bill materials, China's not close to being the most valuable.

Because all those bits are the high capital parts. I was talking about before, where you have to actually, like, create this high precision stuff. Like, there's this company in Taiwan called Largan Precision that makes camera lenses, and that company is, like, almost as valuable as Foxcon, which is well known as making the iPhone. What making the iPhone entails is taking all these components and actually putting them together into the final piece. And there's still a big labor-intensive portion to that. And China, that's why China would, you know, really excelled at that. And, you know, China, it's got a great workforce.

They're well-trained. They're well-educated. very industrious for lack of a better term. And they make great, they make great products. They, they do. And,

and so Apple's deeply integrated into that. And what's happened is, even though today, it's actually, China doesn't have a labor cost advantage, but the entire infrastructure around that, whether it be the capacity to hire and manage a million workers, or to have the infrastructure to move pieces around.

And then China's been coming up in the component place, where more and more pieces of an iPhone precision screws or whatever. Like, they're made down the road, right? And, and so it's one of those things

where the, the, even if Apple wanted to pull out. Yeah. And from what I've heard, Apple didn't even see it, saw all of this as nothing, as none of an issue until the Shanghai lockdowns happen.

And they're like, because that killed, all their backs were made in Shanghai. And there, and that was, from what I heard is, was the real wakeup call. Like,

to get to this scale to make an iPhone just takes years. and years and years that they already put in China. And to make all that investment, to make that possible somewhere else, number one, it's going to be worse. You're going to have lower quality. You're going to have bigger yield problems from a phone perspective, more going to fail because you don't have the highly trained workforce. You're going to have bigger challenges and logistics, moving stuff around.

It's all money down the drain if nothing happens, right? And so it's a massive sort of commitment, and it's a multi-year, decades-long commitment to actually meaningful shift production. Yes, India is making iPhones, but a lot of that stuff is just brought over from China and the final assembly is done there. But that's good. That's a place to start. That's a place to start to start to actually start building up this alternative capability. But for the near to medium term, if there were an actual battle or a fight between the U.S. and China, Apple's screwed.

Completely and utterly screwed. There's this, you know, Tim Cook doctrine that people characterize, which is one of the big things is we need to own and control the primary technologies. behind the products we make. And to me, one of the primary technologies when the greatest triumph of the iPhone is the ability to manufacture it. And it's a complete failure by Tim Cook that they ended up in this position

where they don't control it. It's in China. Now, this is why Tim Cook is one of the great political operators in tech. That's why he was kissing up to Trump, right? Like a lot of people in tech were upset about that. Guess what happened when, you know, Trump imposed tariffs on China goods. What was excluded? All Apple products, right? Like, so that, then on the flip side,

same thing with Beijing. Why is Tim Cook on the board of a business school in Beijing? Or in, I think it's not in Beijing, but in China, Sun Jun. Because, like, he's got to kiss up there. Like, they got to get along with there. And Apple does employ so many people and is the shining example of excellence in Chinese manufacturing that that's worth a lot to China. And it's striking that the U.S. can put these sanctions on Huawei, but China doesn't touch Apple. Yeah, like Apple. Good point. Has played it very, very well from a political perspective, but that doesn't change the fact there is this black swan risk in their business. And it's going to take a lot of money and a lot

of time to diversify away from that. Yeah. I mean, it reminds me of how I used to feel in like 2015 when people would be talking about Netflix and say, this is the future of the television industry and the future of entertainment. And I would sit there and be like, well, a huge part of their business model relies on excellent content, like some of their most valuable content being acquired at a huge discount and a discount that's not going to last forever. And so what happens to the business model once NBC takes away a show like the office? And it's not a one-to-one analogy because it was inevitable that NBC was going to take away the office. And it's not inevitable that there's going to be global conflict between the U.S. and China that really affects Apple's

business. But it is the sort of looming threat that I find difficult to ignore with both NVIDIA and Apple. Absolutely. And I can understand it, though. Like, it's difficult to overstate what a challenge and costs and expense it will be to diversify away from China. and again, to your point, if nothing happens, which is the, I think, the default assumption. And, you know, there is an aspect where Apple's immersion in China and in general, the entanglement of these economies, that is what prevents war, right? It's mutually assured economic destruction that if there were, you know, Russia's one thing. Like, Russia's this relatively small economy off to the side. But even there, like the destruction in Europe is very, very striking that's happening right now.

China would be that times 1,000. And it would be bad for the U.S. It would be worse for China. Like people don't appreciate the extent to which China is so dependent on exports. Like part of that only working the arms and not the legs, China's consumer market is not remotely large enough to support their manufacturing capability. Like if China didn't have the ability to go abroad and to sell abroad, the entire system would fall apart.

And that's a good thing, right? Like, yeah. There is an aspect where good for Apple for being dependent, good for China for being dependent, because that sort of keeps, you know, that keeps us away from our worst impulses. We will take all the checks we can get on global conflict. So Apple, keep doing your thing. Especially me. Yeah, of course, yeah.

You're in the eye of the storm out there. I love your China as bodybuilder analogy, working the arms and the manufacturing industry and neglecting its legs and the capital intensive industries. No, I mean, China is Apple, but as a country. Like, they're so dependent both in terms of market and also for super high precision manufacturing on these other countries. And honestly, that market is going to be the harder part. I mean, it's going to be devilishly difficult to replace some of this super high precision manufacturing.

Like, we see it in chips. It's been a struggle. But you can see, like, the stuff's been invented. Like, they kind of know where they need to go to actually bring their consumer market up to a level where they could handle being shut off from the U.S. I mean, people don't appreciate with the Trump tariffs. The U.S. didn't even notice. The U.S. economy didn't slow down at all, kept growing. And China almost went, like, basically went to recession.

Like, it was crushing to them. And I think that was a real wake-up call to China that crap. Like, like, we, like, it's easy. All the headlines are U.S. lost manufacturing. The U.S. so dependent on China. But that dependency runs both ways. And it's a, it's a real break on whatever China might want to do with Taiwan or anything else. There you go.

Well, final question. Do you think my NVIDIA as quarterback analogy lived up to the standard you've set with a Chinese bodybuilder analogy? No, I thought I was, I mean, I think you can do better. Okay, do better. Hashtag do better. I'll work on it for the next show. Mailbag coming later this week. People should drop their analogies in the mailbag.

That's what we should be looking for. We'll choose the best Nvidia analogy you guys can come up with and it should be a lot of fun. Until then, Ben, I will talk to you soon. Talk to you later this week.