Transcription
Well, good morning, everybody. It's Thanksgiving in the US, and I'm very grateful to have my friend CERN on the channel. And by the way, another surprise for everybody. CERN has a new YouTube channel. It's linked below.
"Good morning."
"Good morning, James. Happy Thanksgiving."
"Yeah, you too. How are things?"
"Fantastic. Thank you. Yeah, we have a lot to be thankful for, and, uh, a lot that we will be thankful for in the years ahead."
"I know it's, it's, you know, I have problems sleeping at night at the best of times, but with this brand new path to trillions, it's been keeping me awake for the last few days, and I'm glad you were able to take the time and try to see if I'm completely mad or not. This is what I call the 13th revenue stream, and it's basically Tesla silicon, and how big that story is, and all the other stuff happening behind the scenes. We'll talk about limiting factors. We'll do some financial analysis, and, uh, this is going to be wild. So, once again, congrats to CERN for his new YouTube channel. His X and YouTube are linked below. And, uh, hopefully, we'll do more of this because CERN and I actually modeled years ago, um, an easy path to $8,000, $10,000, but so much has happened since then with the comp plan where we've got a very specific roadmap as to how we get there. And now we have a brand new revenue stream. It's just a very wild time. So, let's get into it."
"Sounds great."
"If you're ready."
"All right. So, uh, boom, we have a deck. This is called Tesla Chips Trillions. Real simple. But, uh, let me go through some of the homework, the housekeeping first as we go through. This is not financial advice from CERN and I. We're just financial modelers on the internet. That's it. Although he is a financial advisor, so maybe it is from him. Who knows? And happy Thanksgiving to everybody. And this, the inspiration from this came to me on Saturday, uh, last week, and it was such a good question. I actually created a separate deck. I started building it in my Sunday Q&A, kind of short answer, but then it became a monster, and I was like, 'Oh my god, this is too big for just to just gloss over in a Q&A.' But this is from Hobbit 44. 'Thanks for staying with us even after your five-year graduation. Can you shed light on the Tesla AI5 chip being a direct competitor to Nvidia's GB200, H200, etc.? I wonder if the AI5 architecture can be used widely for AI infrastructure like Nvidia processors, or if the AI5 is optimized for FSD and Optimus to an extent that it's less competitive for broad-based sales. If a direct competitor like Tesla would be more energy efficient and give Nvidia some serious competition.' There is, this is one of these brilliant questions you get, like brilliant advice that triggers so many answers, and we're going to try to tackle that today in great detail with CERN. But let me set the stage first. And when I saw this tweet last week, a few days ago, it blew my mind. And I know that Elon Musk sometimes talks with a little bit of hyperbole, but he always delivers. So don't fade Elon. He said, 'Most people don't know that Tesla has had an advanced AI chip and board engineering team for many years. And that team has already designed and deployed several million AI chips in our cars and data centers. These chips are what enable Tesla to be the leader in real-world AI.' Like he's pounding his fist on the table. But the most important part is the next sentence. 'The current version in cars is AI4. We're close to taping out AI5 and they're starting working on AI6. The goal is to bring new AI chip design to volume production every 12 months.' So that means a year from now, they'll be deploying maybe AI6, a year after that AI7, a year after that AI8. And I've got the roadmap for all the chips and what they do. But we expect to build chips at higher volumes ultimately than all other AI chips combined.' Read that sentence again. It's like a movie. I'm not kidding. It's like, 'Read my lips.' CERN, what, what did this mean to you when you read it? Did it blow your mind like it blew mine?"
"Well, it absolutely did. I think there's a couple things that are worth pointing out is Tesla is not new to this game, as Elon mentioned in that tweet. And I think many of us kind of tend to forget that the chips in the vehicles were designed by Tesla."
"Yeah."
"Um, at least for, for the last, last, uh, last few years. Um, and that's very important. So they have a lot of experience in this. And then, of course, the numbers, it's just absolutely staggering when you think about how many chips this could be, given the projections that you and I and many others have made for Tesla's businesses. When you look at the amount of chips that will be required, it is absolutely staggering."
"So there's no doubt in my mind that if, if Elon and Tesla goes down this path, Tesla would be a chip powerhouse."
"And by the way, we're going to get to those numbers of chips too, in a very short while. So let me get on with the story because I really want to set the stage before we start crunching the numbers, and we're going to approach it from two different directions. One, I'm going to show my caveman math, and then we're going to have CERN's financial projections, and we'll see if we are both crazy or if we even come close to each other. So this is, I, I always love to approach, you know, financial valuations from multiple, uh, directions, and I've done that with Bitcoin since 2017 and other things. So, second part is, uh, inference, and this is very important what people overlook. People think AI is all training. No, AI is not all training. And this is a quote from Jensen himself. He said, 'AI inference will be 1,000 times bigger than training.' Okay? AI inference is the token generation business. We'll get into that in a little more detail. But the easy way I frame it, and let me know if you agree with this, because this is the crux, and you and I, you and I have been, if people are having their 'aha' moment about valuing Tesla inference and cars, we did it two years ago. Remember that?"
"And we put actual numbers on it. But training happens once, inference happens millions of times per day. Again, real-world AI, cars, robots, etc. And this is where the rubber meets the road, excuse the pun. This determines everything, you know, where everything happens at the edge, and decisions are made in milliseconds. And we have that live today. It's not, it's not a hypothetical dream. But tell us once again, your view of why inference is so important, and forget training and LLMs."
"Well, the simple way I think about it is, if you think about going to school, how many, how many times or how many hours do you think we spent in school versus hours in our life that were not in school? So, think of school as training, and think of hours out of school as inference. And so, it's probably a thousand times, right? Or something like that, right? It's at least a very large number relative to the amount of time we spend in a classroom learning. So, this is exactly the same thing with AI. I think that's a kind of a nice, kind of a layway to think about it."
"I would challenge you there. I think it's even more extreme because I spent too much time in school."
"Too many."
"You might be the exception, James."
"Yeah. Very well-educated person, as opposed to the, the masses out there that perhaps aren't as educated as you are."
"Yeah. But, but I think if, if I go back to imagine me back in the '80s thinking about school, if I had AI in my pocket, I would never have touched university in a heartbeat. Never. You know, because I'm self-driven. I believe, actually, this is going off on a tangent here. What, just want to pick your brain on this, and I tell people this, uh, I call it education advice. If you are self-driven, and if you are creative, and if you can ask the right questions, you don't need to go to school."
"Yeah."
"Would you agree with that?"
"I, I would agree with that. Uh, I think in this day and age, if you are self-driven, I think you can educate yourself, and certainly with the AI tools that we have at our disposal today."
"Yeah."
"I don't know, I don't think that a formal education is that important, and I think we're seeing that now in the hiring of many of these companies. They're looking for talented people, whether or not you graduate from college."
"Absolutely. Well said. Thank you for that tangent. Uh, let's get into where we are. And this is kind of, I'm just setting the stage because this is, it has multi-dimensions. One, uh, Nvidia is a general-purpose chip, and it does have so-called tech debt, which is holding it behind, and it's designed to service the needs of every Tom, Dick, and Harry out there in data centers, not edge devices. Also requires a truckload of power. That's one of the reasons why Tesla chose, he, you know, if he can buy his own things, he'll buy them as long as the price is right. When his, what he's looking for doesn't exist, he has to build it himself. And Nvidia chips were too power-hungry and not quick enough. A lot of people don't get that, and they've been doing this for many, many years. Um, and again, they prioritize FLOPS over real-world latency where, in a car, a decision has to be made in milliseconds. And I know Ashock shared something about how each car takes in millions of pixels every millisecond."
"I think it was a million or something. Was it 12 million? It was some staggering number."
"Yeah."
"Yeah. Every millisecond. So we're not talking every second, but that's what's required, and that's how the Tesla vehicles can anticipate stuff and get out of trouble real fast, as you know from your FSD experience and your Cybertruck. But anything to add here regarding the, just to share for the audience, the delta between what Nvidia does and what Tesla AI inference chips do?"
"Well, again, again, the analogy here, I think, is Nvidia's chips are kind of general purpose. They can do the job, you know, with everything, but what we want is highly specialized chips, right? Again, it's a bit like someone going to college and just getting a, a general degree versus someone getting a very specialized degree and knows everything about that field. Now, in this case, as you mentioned, it's about the speed of doing that specific task. And when you're driving in a car, the real-world latency is the key issue, right? So you don't want to have a chip that's burdened by all these other things that it doesn't need to be doing. You want a chip that is designed specifically for that task. And Tesla very clearly is, is an expert at doing that."
"Uh, and so that should serve them very well as they embark on this journey to, to develop these chips."
"And, and this is where people get tripped up, okay? Because there's a separate layer to this onion that people get confused by. It's like, okay, they've built a chip. It's only for driving cars. It's useless for anything else. But no, it's not. Because they're able to do, they use the same chip that is in a vehicle for autonomy, for driving cars, in a robot for manipulating an environment, in a factory, or whatever else, and also can be plugged into a data center. The same chip has multiple applications, and I think people overlook that as well. Like, if, if that's why he says, 'Give me as many chips as you want because I, I will use them all. And if I don't have enough demand for my bots, which there'll be infinite demand. If I don't have enough demand for my cars, there'll be infinite demand. I can stick them in my data center.' That's why he's so supremely confident. And I got that from that slide. What was it? Slide five. Um, where he basically pounded his fist on the table. 'We expect to build chips at higher volumes ultimately than all other AI chips combined.' I think that's quite profound. Um, but you want to touch on the versatility of these chips and how it's magical and getting away from Dojo from inference and combining the stacks, combining the silicon."
"Well, I think the thing I would say about this is Tesla has, uh, really specialized and perfected these chips that can absorb visual information, visual data, right? Pixels, whereas Nvidia's chips do some of that and a bit of everything else is the way I kind of think about it. And so when you think about how AI has been trained today, the large language models, we've ingested all of the information that humans have ever created, basically, right?"
"Yeah."
"But what they haven't ingested yet is all of the visual information about planet Earth and everything that is contained in it."
"Exactly."
"And that's what Tesla's chips have been designed to do."
"Yeah."
"Right. So Tesla, in this sense, has a massive leg up over anybody else at this point."
"Yeah. It, it's, uh, it's so interesting as well because what you said, you know, LLMs, etc., have absorbed all the world. But remember, the old world, the old AI GPT world we're coming from is text."
"The new world is all photons. And Elon's been saying that for years, and people have yet to drop that penny. So we're dropping a lot of, uh, really good concepts here that take a while to wrap your head around. Now, if Nvidia, if, if Tesla's threat to Nvidia, just over the last, I think 24 hours, this came out too. Uh, this is the leanness and meanness of the Google Tensor Processing Units. And this little story, I thought, was very fascinating itself as well, on how Google's been making TPUs for a decade, and they're lean and mean. It's an example of a large firm making their own silicon. But what really is important is they can do it at such low cost, and it's also very versatile. And when you look at the Nvidia tax pushed out by Nvidia, they make 70% margin. And you analyze the margin, a thousand ways to Sunday."
"For Nvidia, isn't it 70% plus?"
"Gross margins."
"Yeah. Net margin is like 55, 56%."
"Yeah. Which, which is, you know, the most, one of the most profitable products you can sell on the planet. But Google is the only hyperscaler that doesn't need to profit on the chip because they've got so much demand in their own house. And they have this concept called zero margin stacking, which is incredible. But the important thing about this is, like, Google has gone down a similar path to Tesla. And Google drives the cost per token to zero because they own their own TPU and they've got aggressive cloud pricing. And there's another fact as well around advertisation. And people say, 'Oh, Nvidia chips, they're obsolete in a year.' Well, Google's still using TPUs from 10 years ago in their data centers because, uh, I can't remember the exact thing I used to say is like, when it comes to inference, when it comes to training, you might need a Ferrari, but when it comes to inference, a pickup truck will do fine. You're just moving stuff from A to B. I don't know if that analogy is too crude or makes sense, but, uh, this, the big point I'm trying to make here in a long-winded way, Google has the ability to build the whole electric grid, whereas Nvidia is just a generator, and the grid is far bigger than the generator. And so, as an, an analogy, I believe AI inference at the edge from Tesla could be even bigger than the whole electrical grid, too. Again, it's a big mental gymnastic piece here I'm throwing your way, but."
"Well, there's a couple things I think I would say, James, is that when you think about inference, there's inference that needs to be fast. Cars need to make split-second decisions, but there's a lot of inference tasks that can be slow. There are some inference tests that you could run overnight and wake up in the morning, and there it is for you, and that would be perfectly fine. So a lot of these old GPUs will serve that purpose beautifully, right? Um, in terms of the competitive market dynamics, I think the right way to think about this, and everybody's talking about this AI bubble concept, but this is something I think that's really important to wrap your mind around. With with every other economic good out there, you've got the supply and demand curve, right? And as demand goes up, and if supply stays flat, prices go up. If supply starts to surge and demand stays the same, then prices come down, right? You've got this supply and demand relationship. And we all know it very well. But when it comes to intelligence, when it comes to intelligence, there is no limit to the demand for intelligence. I think the traditional supply and demand curve for intelligence falls apart. And we haven't begun to apply intelligence to everything on the planet yet. Humans have intelligence inside their heads. We're now developing this artificial form of intelligence that we've barely begun to apply to things. We're applying it now to self-driving vehicles. We're applying it now to certain tasks inside a computer. We're going to be applying it to humanoid robots, but as we'll discuss in a minute, we're going to apply it to absolutely everything. So, this idea that that Google's going to take market share from Nvidia, I'm not worried about that. They're, they're both going to grow their, their chip supply as fast as they can, and it's not really going to matter."
"That's so funny you said that because, you know, we always talk about markets being irrational. When this concept dropped, oh my god, Google has TPUs. If anybody with half a brain, they've known they've had those for 10 years and they know they work for inference and AI. But of course, what happens? AMD stock price tanks. Nvidia stock price tanks. And this comes out. I was like, 'Oh no, they're dead.' But you maybe take a second to tell the audience here how irrational markets have become. They literally swing a stock price 10% around news that is completely meaningless. It's a very narrative-driven market, isn't it, James?"
"Um, it, it seems to hinge on the latest story, the latest fear."
"And you've just, you just lived through the MicroStrategy FUD, too."
"That's right. And as Michael Saylor likes to say, volatility is, is a gift to the faithful."
"Volatility as well."
"Right? Yes. Volatility is vitality, right? So, it's a matter, I think, of of understanding these deeply enough so that you don't get swayed by the daily winds."
"Yes. And having the cards of your conviction and using the volatility to your advantage."
"Exactly. And I, I am, uh, I hate to say it, but I am, I like to pride myself. I'm a very good bottom fisher. So whenever things take a dip for silly reasons, it's a dream. People are freaking out. Say, 'Oh my god, Tesla fell to 220. This Trump guy is terrible. The tariffs are hurting us.' And I'm there rubbing my hands together saying, 'Let's go. Bring me more.' Nvidia falls to $88 for the same reason. And it's like, this is exactly the most important thing to say today, apart from all the other good stuff, is have the courage of conviction."
"Mhm."
"Okay. Build, make a plan, five-year plan, stick to it. Build a bag, have a goal of number of shares, and when you see these dips, and as long as the fundamental asset is still solid, grab it with both hands because that's so important. So thank you for sharing that, brilliant, brilliant."
"And I would, I would just add, James, that if you don't have conviction about what you're investing in, then either you haven't done your homework, or you're trying to follow too many companies."
"Yep."
"That's the problem I see with a lot of people is they're just spread so thin, they don't understand anything that they're invested in."
"Yeah. And in fact, uh, Warren Buffett, he had many, and Charlie Munger as well, had very, very good sayings and his, and expressions. And I've grown up with those guys where they're a little older than me. But I remember, I think it was either Charlie or Warren, 'If you don't understand something, don't touch it.' I mean, and this is the tricky part because the stuff that we analyze like Bitcoin and AI and Tesla, they're the two hardest things to understand, possibly, as well. But they're also fascinating because for me personally, I'm just hungry every day to learn more about these things because I know how powerful they're going to be in the future. So, it's kind of like almost like an addiction, which is probably a safe addiction. But let's get, we got to get to the numbers, but we got some more ground to cover first. The structural advantage, I want to stress this as well. Not only the versatility of the chip to work across different things like data center, Optimus, FSD, but the low power and low latency. Low are so important. And this is why Nvidia has no shot in this space. They're trying to compress things and make them smaller. And this is also why Google doesn't have this ability. But for years, Tesla, from the ground up, are maniacally focused, and Elon Musk as well, on that low power draw because you can't have something in a machine, in a car, that sucks power, or a robot where it can only work for half an hour a day. And also the speed at which you can think. I, I think these are very important things as well regarding he talks about the, uh, softmax, MLP, transformer bottlenecks, and how everything is optimized around speed and using literally no energy. And this is what makes it very dangerous in the future world for places like Nvidia and even Google with their TPUs because they do suck a lot of energy. And the two limiting factors I see for AI, and correct me if I'm wrong here, it used to be chips, then people talked about energy, transformers, but Tesla's found a way around those two. And it's no accident that they have an energy business and a silicon business and an autonomy and a robot business all at the same time. Um, any quick points to add regarding that and the importance of power draw and latency, low latency?"
"Well, it's obviously very critical, and I think we're going to see with Tesla's Hardware 3 cars, um, how they're able to take state-of-the-art, uh, FSD and make it available on a, on a far slower, far inferior chip than AI4 and AI5. Uh, I think Tesla will, will achieve that, and I think that that'll be a testament to that. In terms of power draw, I think the, the, uh, inspiration here is the human mind. We actually use very little amounts of power to power our own brains. And so if humans can do it at low power, then why can't computers?"
"Exactly. The human, in fact, uh, this is kind of a weird thing, but, you know, when Elon does things from first principles, he always goes back to the human, how the brain works. You know, the eyes, the brain, the energy draw, the hand, and Optimus, putting the tendons in the forearm, not in the hand, and all these little things."
"He goes back to how the human has developed over, what, hundreds of millions or billions of years. Uh, I think that's kind of fascinating, too. Again, why reinvent something when we already have something that's pretty good? Let's say."
"We should probably understand ourselves first before we try to improve on on humans."
"Exactly. And, and human is by no means, no means simple. All right. Next is the roadmap. We know AI3, Hardware 3 since 2019. Uh, AI4 came out 2023. AI5 is expected, depends, late 2026. And the difference, you can see here between AI3, it tops out at 144 TOPS. Uh, for those in the audience who don't know what TOPS are, they're trillions of operations per second. Okay? These things operate at extreme speed. I'll repeat that. TOPS, trillions of operations per second. That's AI3. That's not quick enough. AI4, up to 500. AI5 will be a thousand trillion operations per second. AI6 and 7, we don't know the exact scope of those just yet. But the other thing that's even more mind-blowing is AI8 apparently will be radiation hardened for what? Space, data centers in space. It's, it's, and this has become a real thing. And it's not just Elon talking about this. So any extra chips he has, they're going up into space. You have got no real estate costs, no energy costs because they're all solar run. They can be pointed at the sun 24/7, etc. But the, this space thing is another thing that blew my mind, and all this Carterv talk, and it's just."
"And by the way, James, does Sam Altman, Mark Zuckerberg, Google, Microsoft, they, they own a rocket company?"
"No, Jeff Bezos is trying to copy. There's a, there's a funny meme. You got Sam Altman and Zuckerberg and Jeff Bezos sitting on a bench, and they're trying to copy Elon's homework in an exam. It's very funny. But yeah, having a space company is again, not by chance. It's very synergistic. The synergy across the Musk industries is crazy. And for years, I've been saying I was really worried about data centers being above ground. I think they should all be buried. What happens if somebody decides to."
"Throw an RPG into a data center and blow up all those machines? Who's going to ensure that?"
"I don't know."
"They need to be underground."
"Or at the very least, up in Canada somewhere where it's cold so the cooling costs are far less."
"Exactly. There's a, I remember years ago, Bitcoin miners used to be in Siberia, you know, there's little sheds out in Siberia. Uh, wild. Okay, now the superpower. Another angle that's been on my mind. You got the custom silicon, custom software, custom hardware, the stack that is built from the ground up to be ultra effective and efficient and draw very little energy."
"And nobody has that edge. And we've covered already that they have applications across many different areas. Not just perhaps they're training. We know they work for training, but they like to train on the Ferrari, which is Nvidia for now. But there could come a day where their AI6, 7, 8 could could be used for training as well, which I think is very clearly on the roadmap. But the key here is the crown jewels is exactly that, that vertical integration, everything built from the house. And the other factor as well, where Elon commits to buying everything, everything from Samsung and layering in TSMC on top of that. Uh, people, again, we'll get to the numbers in a second, but people don't really understand what type of signal this is to the market. Like, I don't see, I don't hear analysts talking about how they're tying up supply."
"Do you?"
"Yeah. If, if one of Tesla's competitors wants to get chips, they're going to have to essentially foot the bill for an entire semiconductor fabrication plant."
"Exactly."
"Or, or at least agree with TSMC or Samsung to buy the entire output because why would those companies open a new plant at that point?"
"Yeah. And another point that supports that as well, TSMC has 70% global foundry share. One player in Taiwan, which is also a risky spot. So that needs to be diversified, big time. And then Samsung is kind of small with their two-nanometer chip as well, in comparison. So it's as if, when I see this layout of kind of the duopoly between these two players, there is room for another player. And the other player could be, and this is pure speculation, uh, and it would be nice if Intel got in the game so that we had at least three big foundries in the world. Um, but I think it's also highly likely that that Tesla goes down this road themselves because the challenge that Elon has is that he has to build this fabrication plant at a pace that none of these plants have ever been built before."
"Yeah."
"He doesn't have five years to build a plant."
"No. He said if something takes more than three years, he's not even interested in talking about it."
"Yeah. It's too long."
"If we start something today, it needs to be finished in two years."
"And."
"And as exciting as it would be to get Intel in the game, I, I think the timeline would be far too long."
"Yeah, exactly. Well, let's jump into some numbers and let's put all of this together after setting the stage. And I'm going to give my rendition of how I believe this could get really big. And I'm going to put us at the bottom or maybe even turn our cameras off. But first of all, if you look at the AI chip market size starting today, it's very, very small, but it's growing very fast. Tesla's market share of it is tiny considering where they are today. In fact, that 5% might be a little bit generous. AI chip revenue, $7.5 billion. 5% of their $150 hundred billion market. And I'm, I'm thinking about the next 12 months here. Nothing too big."
"And the, the real piece is what the net margin could potentially be. And if Elon can create all the chips, more chips than any other player out there, and you calculate how much that could be worth, that could be worth a revenue amount of $135 billion. And what I mean by that, that's 25% of total revenue for the whole space. But the biggest player, that $135 billion is revenue that they don't have to pay to anybody because they build it themselves. Or they could sell them, or they are selling them because they're embedded in vehicles. There's many ways to skin this cat. So think of that as a very crude $136 billion number. But if you apply a 60x PE on that $335 billion, and with the high, very profit margins, you get to very large numbers very, very quickly. I.e., up as high as $8 trillion. That's the first part. Second, uh, mapped all this out, the AI chip market, the size of it, the amount of Tesla chip revenue, the base revenue, the total revenue, and net profit. Again, that $136 billion at a 60x PE ratio, we get to $8.1 trillion, which coincidentally is the plan to get to $8.5 trillion market cap. Now, when I started running these numbers, and I fall back into nearly $8, $12 trillion, that was not, I wasn't trying to push a round peg into a square hole. It just happened that way."
"And that's why I reached out to you and I was saying, 'This is, this is pretty crazy.' Um, keep me honest here, buddy."
"Yeah."
"Where do we go next?"
"Well, if you can pull up my slides."
"Yep."
"Um, I'll walk you through sort of my thought process on this."
"Oh, yeah. By the way."
"Want to cover this first, perhaps?"
"I do. I do have a bare case, too. So, the bull case is that, you know, $8 to $10 trillion. Bare case is $4 trillion. Bare case if they pull off their AI6, AI7. So, it's still absolutely crazy. And we'll come back to this when we sum up, but here's your first slide."
"Yeah, this is actually a discussion on Elon's, uh, comp plan. Maybe we'll, we'll hit this, um, after this. If you just fast forward to the, um."
"There's several slides here. It's, um."
"You'll see a table."
"Okay."
"No, keep going. Keep going. This is still part of that section."
"Okay."
"Sorry, there's six or seven slides."
"Uh."
"Here we go. Okay. Okay."
"All right. So, let's kind of build this up from the, from from the base up, right? So, Tesla Silicon, by the way, TSMC, Tesla Silicon Manufacturing Company. It could be another TSMC in the world."
"Um, okay. So, Tesla's going to need chips for their vehicles. That's the blue line at the top."
"And I've modeled out anywhere from 2 million to 20 million vehicles. So, pick the column that you fall in."
"Okay. Now, that, that would be, well, 20 million is the plan, the comp plan. 2 million is pretty much where they're at today."
"Yeah."
"So you're saying pick a number on this continuum."
"However many vehicles you think Tesla's going to produce on a kind of a run rate basis at some point in time, and pick your time frame, pick the number. Let's just arbitrarily pick 10 million. Every vehicle has two chips in it, right? So right there is 20 million chips. And then the, uh, purple column, the number of Optimus robots starting at 5 million down to 100 million. Pick the number of Optimus robots you think that Tesla will produce. Okay, let's just pick one in the middle, say 50 million. So the intersection between 50 and 10 gets you to 70 million chips that Tesla needs to produce for internal use. The green shading is kind of where we need to be to get kind of a 100 million chip number."
"Yeah."
"So a staggering number of chips that would be needed. U, they could do it by producing, for example, you know, 100 million robots, um, and 2 million cars, or you could produce it by 60 million robots and 20 million cars. However you want to slice and dice that. Okay. So that's Tesla internally. Now let's look at Tesla externally. So, what could Tesla's chips be used for? Data centers on Earth, data centers in space. Okay, that's the blue bar at the top. Again, between two and 20 million. This is chips this time, not number of data centers, but number of chips between two and 20 million. And then inference on the vertical, on the purple. Again, not that difficult to get to a demand for 100 million plus chips."
"Yep."
"Right. Just with those two uses, but it's even bigger than this. James, if you go to the next page."
"So, if we look at the inference chip market, my premise here is that every single computing device becomes an AI computer at some point, right? So clearly every car, every humanoid robot, but also we're talking about smart infrastructure, cameras, drones, forklifts, things in warehouses, energy and grid optimization, Powerwalls, Megapacks, solar controllers, space and Starlink, right? We've got Starlink terminals and satellites. Got data centers. We've got consumer electronics, smart fridges, dishwashers, ovens, AC units, everything that consumes electricity can become an AI device. Logistics, delivery vans, drones, ships, tractors, mining equipment, wearables, sensors, virtual reality, augmented reality, wearables, sensors of all kinds, and then space robotics. We're going to put hundreds of millions, if not billions of robots into space. So what, in the crude ways, think about this is we will need one inference chip almost per human on the planet."
"At least. It'll be multiples, actually, multiples."
"So I did a quick rough and dirty low and high, and I come up with 2.3 billion infant chips needed in the low case and five to six billion on the high case. Take the midpoint, you get 4 billion. Again, really quick and dirty. And of course, you know, there's no time frame associated with this. So I'm not saying this is by 2030. This is at some point in time. But you're, this is the kind of scale that we're looking for, I think, for the inf chip market. And actually, I think I've probably undersold this."
"Yeah."
"It's probably 10 billion plus at some point."
"And this explains the Elon sense of urgency about building his own fab because he's not going to be able to satisfy his need from TSMC and Samsung."
"Yeah."
"Okay."
"So let's just keep this 4 billion, or if you want, 10 billion number in the back of our minds."
"Okay."
"Because I think that's where we're headed to longer term."
"Yep."
"Okay. And then if we roll forward, uh, to my next slide, although it looks like you may have inserted, uh, another one here that's that's really important."
"Yeah, I, I think we need to do, we need to go back because I had somebody put these slides in to, they're in numeric order. This was of 11."
"The one that's after this is a gross, gross margin chart."
"Yeah. Gross margin dollars."
"We could be missing one. Let me, let me try get the deck. Bear with me here. And, uh."
"So while you're, while you're looking for that, I'll just sort of verbally describe what we'll see on the screen in a minute. So if we take the number of chips, right, and on this chart, you'll see I start with 10 million and go up, goes up to two billion on the vertical scale. So we can look at a number of different points in time, number of different amounts. And then the way I approached it in terms of what's the opportunity for Tesla is not so much on the revenue side, but what I was looking at gross margin per chip. And I modeled out between $50 up to $500 just to see at different amounts of chips that are produced and different gross margins because obviously as you start to produce 4 billion of something, the gross margin on each item will come down dramatically. So at 10 million chips and if the gross margin is $500, you have, uh, $5 billion of gross margin. So that's a pretty sizable business. If you get to 500 million chips and let's say at that point the gross margin drops down to $300 a chip in gross margin, which is far less than what Nvidia makes, you're looking at $150 billion in gross margin."
"And then if we get up to, uh, two billion chips and let's say gross margin at that point drops to $100, you're looking at $200 billion in gross margin. So either way, we're looking at potentially a very sizable business for Tesla if they do sell any of these chips externally."
"Yep. I got that now. Um, remove this. Hold on. Uh, remove. And then let me pull this up. This is where StreamYard really sucks. It's very difficult to juggle between different images."
"And it's always fun to do it live."
"Yeah. Hold on. Share screen. You gotta, you gotta remove one and then you gotta, if you already preloaded the other, you gotta move it back in. Here we are."
"Here we go."
"All right."
"So, this is the table that I was trying to talk through. It's obviously easier to see it visually, but, uh, gross margin per chip in blue at the top ranging from $50 to $500. And again, I don't know what the gross margin per chip is going to be. It may well be initially, you know, $500 or more, right? So at low volume, the $500, I think comes into play. We're looking at a $5 to $25 billion gross margin opportunity for Tesla. As the volume ramps up, you move towards the center on the blue bar, maybe to $300 or $200 in gross margin per chip, and you're looking at a $50 to $150 billion gross margin business if you're producing 200 to 500 million chips. And then if you get to the billion, one and a half billion, two billion range, gross margin probably comes down, and you're looking at $100 to, you know, $200, $300 billion in gross margin. That's the kind of opportunity that we're looking at here, I think, for Tesla longer term with, with Tesla silicon. Now, again, if you remember the previous image, we were talking about potential annual sales of four to, you know, potentially 10 billion. That's not even on this page. No, that's completely wild."
"And then we have, we have evaluation."
"And then so you take this, then take evaluation. Now, I used half the PE that you did. I used a PE of 30. If you don't, if you want to use a PE of 60, double the numbers that you see on this next page."
"Okay, hold on. Share screen as we cook. Here we are with the valuation. And I'm gonna make us small so people can kind of see this. All right. Um, walk us through this in great detail because I backed into $8.1 trillion market cap if."
"I'm using a PE of 30. So if, if you wanted to sort of normalize it and use yours and double all these numbers. So, if they sold, for example, um, you know, 500 million chips and the gross margin was $500 per chip."
"Then you're looking at, uh, almost an $8 trillion valuation just on that."
"Which is the comp plan, which is exactly what I came to, but yours is heavily sandbagged to a factor of two."
"I'm using a lower PE. Now, initially, obviously, the PE would be higher depending on, you know, how quickly they can grow, but I'm assuming that they would ramp up pretty quickly. So the PE would be very high, but I'm looking at kind of a longer-term PE of this 30. Uh, Nvidia is, you know, in that range, a little higher. Um, but at the point, let's say they're they're making, you know, a billion chips. Um, and let's say the gross margin per chip is is $250 to $300, you're looking at, you know, four and a half billion dollar valuation on the, on the chip business at a PE of 30. Um, again, I could easily make the case that the PE should be higher than that."
"Yeah."
"And so you can apply whatever PE you would like to to this, uh, double that if you want to, but either way, depending on your assumptions here, you can see that this business could be a very significant, uh, part of Tesla's value at some point in time."
"I mean, it's completely mind-blowing to think about. And also let's talk about the PE for a second because you're very heavily sandbagged here. When you're talking about things like AI and knowledge, information manufacturing, you know, first of all, PE is not what matters. Free cash flow does. Yeah."
"But the amount of money that can be minted and generated, doesn't that command a much higher PE in reality? And let's throw Nvidia out the window because a lot of people don't understand how much demand they have for their products. But, uh, and I know that the Google valuation is pretty cheap too. But if, if the PE, the PE is a function of the growth rate, and if the growth rate trajectory continues, I know you have some financial models where it levels off in 2035, etc. But."
"Talk to us about how that PE should be a lot higher, especially during the high trajectory growth phase, which I don't think will slow down."
"Well, I think you're right, and again, going back to my comment earlier about how there's no upper limit on the amount of intelligence we can use."
"I think we may well be surprised at how high and how long-lasting the growth rates are for AI and for for chip demand. Um, and and that's to Elon's comment about there's no upper limit to the GDP growth of of Earth. There's no ceiling on that. And again, I think it stems from the fact that there's no limit, upper limit on intelligence that we can use. And that's, that's why he's able to say that. I think that's kind of the base, the base for that. So if you have a business then that's growing rapidly and that has massive upside in terms of amount of units in the billions, then the PE that you should apply to that should be very high, probably triple digits."
"Right. And we're seeing that today already with Tesla stock. The PE is very high, and it partially reflects some of that future growth being valued in today. Now, when you and I look at the models that we've made, we still think that Tesla is massively undervalued even into today's high PE. And what I've done is, uh, looked at Nvidia, uh, before it had its hockey stick growth, and today where Nvidia is trading today. You could have bought that at less than one times earnings just a few years ago."
"Yeah."
"Because the growth has been so high that even a few years ago, the PE might have looked high, it wasn't because the company grew that much, right? And I think we're going to witness the same thing potentially with Tesla as they really start to accelerate these, these AI-powered business models."
"Yeah, this is, this is very interesting, and there's a lot of stuff spinning around in my head right now. I did make one comment in my valuation model that I didn't take into account capex."
"Or energy."
"But cashy, here is a very interesting question. What about the energy needed to produce and run this?"
"That's a big part of it. So, for example, they've got this commitment to spend $16.5 billion with Samsung for the AI5 chips. Where's all this money going to come from? Is things like Optimus, Robboaxi going to fund a lot of."
This capex, or where does this fit into the model for the return on capital?
Yeah, I mean, I I think certainly it will. If Tesla needs to spend 25 billion, 50 billion on a semiconductor fab, they'll do so over time. They should be able to fund that with their cash flow. The cash flow from Roboaxi and Optimus will be insane, off the charts. So they'll have plenty of cash flow to fund the growth of these businesses.
Um, the challenge is just the time frame, right? It it takes again, a few years to build a semiconductor plant. It's not like you put it up in six months. It takes takes a while. So I think Tesla would be able to fund this. Um, but even if they couldn't, we have something called the capital markets that that Tesla could raise capital from if they needed to.
Yeah, right. And that's, you know, we'll have to see whether or not they go down that route. If if they really want to accelerate this, if if you want to get it really hit of the curve and build this capacity in advance of needing the chips, then they may need to raise some capital. And then the other angle is from cashier is the energy required to run all of these things. Now, one of the beauties of things like the cars, they have their own cooling and energy pack in the vehicle.
Yeah. So basically they've outsourced all of that to the owner of the car. If we go to inference at the edge, same thing with Optimus in space, who knows? Uh, but it's handy that as as Elon said, handy he has a space company too. And uh, the other angle is the raw materials for silicon aren't that rare compared to the magnets in the robot that's required. That's much more of a limiting factor. So, and Elon also said that that he could if the US strategically deployed mega packs in the US, they could double the energy of the grid because so much energy is wasted.
That's right. So, we we definitely have some issues, right? We we definitely have an energy shortfall that hopefully over time is addressed and it can probably not be addressed fast enough. There there will be some constraints. Um, generating energy in space I think is an intriguing thing. Elon has talked about the idea of of not really sending that back to Earth is you're just using that in space to power the data centers.
Um, and eventually probably manufacturing in space as well. So hopefully we can reduce the energy demand on Earth for some of these energy intensive things. AI data centers being one of them. Um, and Elon talked about that just in the in the past week, right? needing 100 terawatt uh hours uh sorry, 100 gigawatt hours of of energy for data centers. He says we're not going to build that on Earth, at least at least in the US.
Um, but these are these are some of the short-term issues I think that that they're going to have to work through and there's no better person in thinking through some of these problems than than Elon.
Exactly. In fact, we can open up to a couple of quick questions. I do want to be respectful of people's time during this Thanksgiving holiday in the US and we're very early today because I think Sarah has to go cook later, hopefully. But his link is below for his new YouTube channel. It's going to blow up. I I told him he'll have 10,000 subscribers in no time.
Uh, but this is another key quote which when you start thinking about Cardartesef to scale and what this all means. Um, but the way Elon's mind works is a little bit different. It's beyond exponential. But he says think in terms of KartV 2 and the path becomes obvious. Ela Musk and he thinks eight years ahead and he puts all the building blocks in place to make sure he can deliver against his vision.
So uh, you probably heard about this. I think he said it yesterday. Uh, this is pretty mind-blowing when you dig into the whole cart too and the amount of energy on the planet and how much we actually use versus don't and look at all the stuff that China is doing with solar etc. In fact, Elon is building a massive solar charger, a second one in California as we speak, you know. So for the people out there that say, "Oh, solar doesn't work." H yeah, it does.
So any thoughts on what this means and how his mind thinks? And back to our story, we converged from different angles to easily an 8 trillion market cap from silicon alone which has to be reiterated. It's all pretty crazy. How does Cardishv dovetail into this?
Well, you thinking?
Yeah, just like Elon's been thinking about chips, he's been thinking about energy generation probably for just as long a period of time. We haven't yet seen the details of Master Plan Part 4.
Yep. And my guess is that a big part of that is solving this energy generation problem. And of course, you know, Elon talks about solar. So that that's where I think his focus is going to be. The good news is we have all kinds of innovation and development on on nuclear and other sources of energy as well. So it's very promising, I think, in terms of what's coming for for for the planet in terms of energy generation. This is a critical issue to be solved.
in case anybody missed that little piece, how he thinks and how big he thinks.
Yeah. Again, this is something Elon's been thinking about for a long time. It's obviously very critical we generate enough energy to power everything that Elon wants to do and and other companies. Um, and so I think we'll we'll hear more about that from Elon very very soon.
Yeah, exactly. Wild times. Um, I want to thank you once again, CERN, for taking time on this US holiday. I hope you and your family have a brilliant day. I do want to do this again. You're going to get your wiring up so we can co-stream to our channels together at the same time.
Uh, but I do want to one of the things that you and I have had on the cards for a long time is a base case worst case scenario for Tesla priced in the year 2030. And the reason for that is there's a lot of people who are planning around their retirement and people want to be out of the game in five years. And how much can they expect or how high can they expect Tesla to go? Like really a sandbag to hell version? I've got mine. I've we've seen yours, but I think we can take another cut at it too.
And weave it into some of our models. I think that'll be very very important for the audience as well. And thank you once again for spending time with us. Links below for this guy. And a big thank you to our mods in the channel. And let me see, we've got Panda Bando and Green Candle. Anjo Junka Gamble. Gamble again. We had over 3,000 people watching live for this early is very unusual. And for an AI Tesla video is also very unusual. So CERN, you you attract the big numbers. Big thank you to TNT Tesla as well. As well as Shauny and K8. Appreciate you, buddy. have a great Thanksgal remarks before we wrap.
Well, there's a lot of things we could add, but but thank you for having me on today. This has been a fun discussion and there's a lot more, I think, to talk about over time, but uh I really appreciate the invite, James, and happy Thanksgiving to you.
Oh, no. I appreciate your modeling. But let's do the same thing again in a week or two. Uh, we'll come at it from both sides around the ultra sandbagged price prediction for Tesla.
And is there a faster horse? because that's what I'm always looking for to balance riskreward. Is there a better opportunity out there? Should we just yolo into Nvidia or Google or
Micron or Bitcoin or Micro Strategy with the NAV discount? Who knows? We'll do that, too.
Sounds great.
Friend. Happy Thanksgiving, everybody. Bye-bye. Bye-bye.