00:05
Welcome, welcome everybody. My name is Jeff Towson and this is the Tech Strategy Podcast from Tecmo Consulting. And the topic for today, what the SpaceX IPO filing teaches us about both AI and tech strategy. uh Some really great lessons in this. mean, the IPO filing is kind of crazy and a mess, but like in terms of understanding what AI is doing and understanding how...
00:31
Elon really approaches building businesses, tech strategy, a lot of good lessons. So I'm going to go through what sort of my takeaways are and my summary of it, which is not really explicitly laid out there, but it's my what I sort of gleaned. So that'll be the topic for today. Should be pretty fun. It's kind of a crazy document, really. I'm going to that. OK, let's see. Standard stuff. uh If you're interested in getting a free copy of my book, Motes and Marathons, Part One, not
01:00
tooth or seven. You can go over to jefftowson.com or techmoconsulting.com, sign up there for the email list, you'll get a downloadable PDF copy. Yeah, I feel pretty good about it actually. I spent quite a few months rewriting since about really July to December, rewriting almost all of my thinking. And this is sort of the first rewritten version and I just decided to put it up and let people download it. And then I'm going to rewrite tooth or seven, but most of the main stuff's in number one. The rest is kind of details.
01:30
Anyway, so that's there if you're interested. And let's see, think that's it for today. Okay, standard disclaimer, important on this one. Okay, standard disclaimer, nothing in this podcast or my writing or website is investment advice. The numbers and information for me and any guess may be incorrect. The views and opinions expressed may no longer be relevant or accurate. Overall, investing is risky. This is not investment, legal or tax advice. Do your own research.
01:56
So this is about the S1 filing in the IPO. I am absolutely not giving you investment advice about whether to buy this stock or not. No way. This is purely strategy lessons. Okay, with that, let's get into the topic. All right, now there are some strategy lessons in this, or strategy concepts, I should say. And there's really three, I think, that kind of summarize the Elon approach, in my opinion. Number one is...
02:22
the BCG shaping strategy. I've talked about, look, strategy is really dependent on your landscape. You're going to have a very different type of strategy doing Coca-Cola or Starbucks than you are doing, let's say, a fashion company versus a, let's say, an e-commerce company. And then one of these, and there's a BCG quadrant, four quadrant sort of uh matrix for this. And they sort of look at it based on predictability and
02:51
really malleability. How predictable is your space and how much can you shape it? Now Coca-Cola is one extreme. Look, the space is really predictable and you can't really shape it at all. People buy soda the way they buy it. You can't change the structure of the industry. So that's classical strategy. Okay, the other extreme would be shaping strategy. you know, this thing is kind of predictable over the long term. You can kind of see where it could be.
03:22
Which rockets, like, we don't know what people are going to buy in fashion a year from now, two years from now. That's unpredictable. So you have an adaptive strategy. But you kind of know what rockets going to the moon is going to look like eventually. So the future is somewhat predictable. And then it's also basically malleable. And this is kind of the Elon Quadrant. He's a shaper. know, he can kind of see what the future is and then he brings it into creation.
03:48
So we can call what he's doing a type of shaping strategy. And I think that's where he's lived his whole life. So he's got a very good system for doing that particular approach. But your strategy really does depend. So that's concept number one. Concept number two, innovation marathon. This is sort of core to my Motes and marathons. I used to call the marathon smile. It's oh SMILE. I've changed it to slice.
04:14
Basically, the operating marathons that can create a degree of advantage, it changes over time. So now I call it slice. But within that, I, S-L-I-C-E, the I is basically innovation. Long-term, sustained, cumulative, path-dependent innovation, where if you want to copy what someone is doing with rockets or whatever, you have to sort of follow the whole path they did over time. You can't just jump to the end. So.
04:41
That's kind of a big part of what he's doing. he's been my example, Elon, for Sustained Marathon, really since the same book, the first book. And the third one, which I haven't talked about before, is the idea of repeatability. These companies that show sort of a longer-term strategy and a high success rate in what they're doing tend to have a high repeatability in their activities. This is, I think it was
05:09
Chris Zuck used to be at Bain. He wrote a book called Repeatability. He wrote a book about how you grow and ancillary growth options, core growth options, profit from the core, things like that. Very good books, but one of the concepts he talks a lot about is repeatability. When you do something over a set number of activities over and over and over, you can get better at them and you can get cheaper at them, which is a huge part of what Elon does. And also, it increases your probability of success.
05:39
So if you're hunting for new growth opportunities, the key is not to try a big thing every two years. The key is to constantly go after new growth opportunities every three months, and you will get better at it. And therefore, your probability, your hit rate will go up. So repeatability is actually a big part of what he's doing. Those are sort of the three concepts for today. All right, let's talk about, I'll put in some JPEGs from the S1.
06:08
And yeah, it's kind of a messy document. goes between, well, one, the strategies all scattered everywhere. And then it goes between sort of business thinking to sort more sort of, I don't know, almost ideology purpose in life and talks about going to Mars and making the species multi-planetary, which is how he talks. But it's not exactly a traditional S1 by any means. Now, the overall IPO document is pretty weird. It's...
06:37
It's three to really three to four companies kind of jammed together. It's like, hey, we have, you know, this uh rocket business, which gets you access to skate to space at scale. have a connectivity business, which is Starlink. We have an AI business and you know, they're all sort of linked and it's one strategy. And I'm like, I don't really think it's one strategy. I think you got three businesses here. They help each other in some ways, but in other ways, not really, but.
07:06
You know, it is what it is. So as a document, it's kind of strange. Here's some of the basics from the introduction. So founded in 2002, quote, integrated hardware and software infrastructure for space connectivity and AI. OK, there's your three businesses. And then this integration of hardware and software is actually kind of a big part of the strategy.
07:34
So I'll talk about why that's important. And then there's a bunch of factoids, like since 20, don't know, 23 space, you know, they have launched 80 % of the mass to orbit with a 99 % success rate. High speed, low latency broadband, which is deployed by 9,600 satellites in low Earth orbit. Okay, XAI. So there's all these random factoids, which are pretty interesting, to be honest.
08:05
The core business is obviously the Rockets, which although it doesn't make that much money, the money's in the other business, but the foundational technology that enables the other businesses like Starlink, and if he starts putting his data centers in space, well then it's the Rockets. But the money doesn't, the Rockets actually pretty small money. I'm not going to go through the financials at all because it's, they're a total mess. Like good luck trying to value this thing, like forget it, like.
08:34
look at the TAM and then go from there, but the financials are kind of a mess. But the rocket business has the cleanest sort of pitch, which is like this is access to space at scale. It's disrupting an industry that was prone to stagnation, risk aversion, and economically perverse cost structures. That's actually kind of important. I'll talk about why that's a decent segment of his approach.
09:02
that he is targeting a business that has stagnation, risk aversion, and economically perverse cost structures. That's a big part of the reason this succeeded. It's a big part of the reason why Tesla is such a more difficult business, because your competitors are really good in cars. Your competitors in the space business are kind of legacy government contractors with perverse incentives, which I think is well put.
09:31
A lot of this stuff, he goes through connectivity, he goes things like, let me give you a couple, I'll read a couple of JPEGs for you. I'll put these in the show notes of how he describes this stuff. So here's one. The S1 begins with our mission. It's a mix of sort of belief and actual some business strategy. So our mission is to build the systems and technologies necessary to make life multi-planetary.
10:00
to understand the true nature of the universe and to extend the light of consciousness to the stars. Belief, purpose, mission, continuing. To do this, we have formed the most ambitious vertically integrated innovation engine on and off Earth with unmatched capabilities to rapidly manufacture and launch space-based communications that connect the world.
10:24
to harness the sun to power a truth-seeking artificial intelligence that advances scientific discovery, and ultimately to build a base on the moon and cities on other planets. That's kind of a mix of mission, and there's actually a pretty good summary of his strategy there, which is the most ambitious vertically integrated innovation engine. That's really a good summary what his business is. It's an innovation engine, very vertically integrated.
10:54
Super ambitious. And then he's got some KPIs. Here's a little bit about SpaceX, specifically the rocket business. is the only company that has cracked the code on accessing space at scale, revolutionizing an industry characterized by decades of stagnation, risk aversion, and economically perverse cost structures. Okay, that's a business strategy.
11:24
access space at scale, change in industry characterized by stagnation, risk aversion, and economically perverse cost structures. Yeah, that's a good business strategy. SpaceX upended this paradigm through the application of first principles thinking, which rejects industry assumptions and builds solutions based on the fundamental laws of physics. Okay, fine, fine, fine. Let me skip through. Our intense mission-driven engineering first culture.
11:53
and focus on extreme vertical integration, have it propelled us to achieve what many deemed impossible? Again, there's the same idea. Extreme vertical integration. He talks about this a lot. then engineering for his culture, also important. Going on, we pioneered high cadence, reliable, and affordable access to space. High cadence, high speed, high frequency, affordable. You're going to hear these two.
12:23
Ideas a lot using different language affordability basically he makes things cheaper and he goes really fast Now you got to have reliable in there because rockets blow up But this idea of speed plus low cost speed of innovation speed of rocket speed of cadence of launches plus affordability low cost That's a you see that in literally every one of his business. He talks about that Okay, that's kind of SpaceX. I won't go through the rest
12:52
He has a little blurb, this is from the beginning, about the AI business, which is what I'm most interested in, to tell you the truth. For the AI, we were the first company to develop a coherent gigawatt scale AI training cluster. OK, that's Colossus One.
13:10
Gigawatt, so larger, first, early to market. This idea of coherent training I'll talk about, that's actually kind of important. For complex reasoning and agentic workloads, compute is directly correlated with the quality of intelligence and task completion speed. That's actually like his key KPIs right there.
13:35
He's focused on training and inference for complex reasoning and agentic workloads. Those are the ones that require the most compute. That's where you need the most compute, which means the most tokens. uh Agents use a ton of tokens. Complex reasoning, which your AI runs for a half hour, uses a lot of tokens. So cost of token is like his primary metric.
14:02
He says this correlates with the quality of intelligence and task completion speed. So you see the same idea, quality of intelligence, cost, speed. Good. He goes on, in under two years, we have established a dual advantage in both cost efficiency and deployment speed at scale. Same idea. By owning the compute infrastructure and vertically integrating across the full AI stack,
14:31
We can train and iterate our models at lower cost, higher velocity, and accelerate development styles. Same idea. Vertical integration will give us a dual advantage, cost, and speed. And in this case, you'll be better. Blah, blah, blah. I won't go through the rest. Talks about strategic advantage. Then he gives some numbers, Grok. 500 million. This is kind of interesting. 550 million MAUs using Grok.
15:01
That's a pretty good number, 550.
15:07
Okay, that's pretty much what I want to go through. Okay, so you can see sort of you start to hear the same terms and as you go through the whole document, you'll hear these various terms over and over in different vocabulary and context. So let me sort of pull it together and what I think the actual strategy is using more sort of my language. Okay, so my summary of his tech strategy, which I would characterize as a shaping strategy is really three steps.
15:37
Step number one, target a big problem with a massive TAM, total addressable market. The number they put in there is they look for $1 trillion opportunities. And actually what's pretty crazy is the whole business, the three businesses together, they characterize as a $27 trillion opportunity, which is amazing if you believe it. What's interesting is 26
16:06
trillion of it comes from like AI applications and enterprise and things. like the part that's really sort of clear like, okay, rockets going into space and let's say more clear and then, you know, offering broadband service with your satellites, you know, that's a tiny fraction. The 25, 26 trillion of this apparently $27 trillion opportunity is pretty kind of speculative and there's not a lot of details. I don't know if I buy that, but
16:36
Anyways, the point is the same. Look, step one, yeah, go for a big opportunity. Something that's going to change everything, big money. But also within there, I would put a sub-bullet point, target, sort of weak incumbents. Now maybe you're going to create a whole new opportunity that doesn't exist, which is probably the case.
16:59
It's even easier if you're dealing with legacy rocket companies that have perverse economic incentives, stagnation, whatever. That's even better. Now, I think that's a sub-bullet point. I don't think that's his main thing. I think he looks for world-changing opportunities that have a massive TAM, and if they have a legacy problem in there, even better. But you've got to have an advantage. So that's kind of step one. Step two, uh provide a solution.
17:29
that is mostly about having world-class engineering capabilities. Look for a problem that others can't solve, that is fundamentally an engineering problem. And then the idea is, you have world-class engineers that, you know, that's a really specific thing to do. Most companies can't replicate that. Even other engineering companies, for the most part, can't replicate his team. So, but you're looking for fundamentally
17:58
an engineering solution to this problem. And that gives you a tech advantage over time. And that's why he talks about we have an engineering first culture. And you can see all of his businesses sort of fit those first two steps. Massive opportunity, huge world changing, hopefully with some weak incumbents, not in the case of EVs. Your main solution is engineering first.
18:26
That's most of the, you know, he's not a great brand builder. He doesn't do marketing terribly well. I mean, they do it, but it's not their strength. No, the vast majority of value that his companies bring are engineering based solutions to problems. Our rockets get to space, our cars work better. You know, our lower satellites work. It's a functional engineering solution. It's like 80%. Okay. Then we get to the interesting part. Step three, last step. This is where I think the details are.
18:56
His go-to approach is kind of what I've already been saying. It's, go for very rapid innovation and iteration. Quick, quick, quick. know, rocket launches all the time. New updates to Grok every week. New cars coming out all the time. Advance, advance. That sort of speed of innovation is part of it. And then that goes hand in hand with let's drop the cost of things. All of his business sort of speak to that. We're going to make things cheaper.
19:26
It's kind of his main competitive lever. Now you could say, argue Tesla's are a bit expensive, but he's trying to drive those down. Rockets are going down. I'll talk about the AI. Cost is like 80 % of what he's doing. So rapid speed, affordability, cost efficiency, whatever you want to call it. That's kind of the two. To do both of those things at scale.
19:49
Now, the two tricks, the two go-to levers within achieving that, apart from just working really like a maniac, is vertical integration. Why? Because if you own everything soup to nuts, you can make changes much more quickly. If you're doing a modular approach where you're dealing with suppliers and they're sending in their parts, making changes takes a long time because it has to be compatible, you have to work with the supplier.
20:18
But if you're doing everything, hardware, software, the full tech stack, the full car, the full rocket in-house, you can make changes very, very quickly, whether they're to improve performance, increase reliability, drop the cost, which seems to be a lot. So vertical integration seems to be his main lever. The second one is the one I mentioned about, which is repeatability. He's got rockets going every couple of weeks.
20:47
He's got new cars coming out all the time. He's got new satellites being launched, thousands of them every year. If you're doing these rapid innovations and iterations, it helps because you're always deploying more, deploying more. So every round you can make more changes. And as long as you're highly repeatable, those changes will accumulate. If you're jumping into new things all the time, you don't get the benefit of repeatability plus time.
21:15
And so you sort of stay on this one path. And I think that's a lot of the key to his marathon, his innovation marathon is uh vertical integration, high repeatability in his activities and in his products and in his launches and all of that. So there's a couple of cool concepts within all of that. Now, summarize this up. I think that's a fairly reasonable summary of a very experienced
21:44
strategy for a shaping landscape, for an engineering first landscape, for a tech landscape. You know, I have sort of these four buckets on the BCG uh framework. I have lots of good examples for like classical strategy, Coca-Cola, adaptive strategy, fashion, uh e-commerce, and others, but shaping strategy, there's not that many people that do this. So I like that this is like a very detailed approach for that uh strategy. Anyways.
22:14
Now, you kind of need a couple things for this to work. This whole we're going to innovate continuously over a long period of time and drive down the cost. That only really works if you sort of have a long, long runway in terms of engineering progress. If you're doing, let's say, translations with an app, OK, the translation is going to get better. It's going to get better. And then it's pretty much going to be done.
22:41
There's no more improvements to be made. It can translate English to Chinese without mistake. There's no more runway. He needs a long, long runway for this strategy to work. And so everything he's in has these sectors where you can build and innovate in advance for decades. Okay, most businesses don't have that. So you kind of need a long development runway pathway. You also kind of need
23:10
a big cost structure that you can take apart. When he went after rockets, I mean, he was trying to drop the cost of a rocket launch by 95%. If you're doing something that's already pretty cost efficient, there's not a big room there, not a room there to knock down those costs. So you're losing one of your big advantages if you can't really change the cost structure dramatically. So I think the shaping strategy, everything he does to me looks like it's leveraging all of these aspects.
23:40
Anyways, that's kind of my take on a strategy. Now let me sort of jump into the AI side, because the detail there to me was the most interesting by far. So the breakdown of basically the AI business, which has a couple of things. has XAI, it has the data centers, it has their, they're building chips now. And of course it's got X and all that sort of public discourse. But here's how the S1 sort of describes it.
24:10
says they built the first coherent gigawatt scale training center cluster. So that's Colossus in Memphis. And it's focused on complex reasoning and agentic workloads. So again, are the, one, those are the highest sort of compute heavy workloads. And two, those are the ones that are probably going to explode in volume. Definitely agentic workloads are going to go through the roof.
24:39
Okay, so focused on that. Here's what they say. In the past two years, I'm paraphrasing it, so it's not an exact quote. In the past two years have built a dual advantage in cost efficiency and deployment speed at scale, right? Same thing, hearing it again in each of the businesses. By vertically integrating the full AI tech stack, you can train and iterate faster.
25:06
and your models and at lower cost and higher velocity. So again, driving the cost down, speed, and his go-to tactic for that is vertical integration of the, in this case of the whole AI tech stack. Why is that effective? I'm paraphrasing, because it eliminates external bottlenecks and it allows rapid and continuous improvements in model performance.
25:33
All right, but then we get to sort of step two, which is, we're going to do an engineering first solution to this problem, which is intelligence at scale that's affordable. That's not exactly clear what that is. Like we know what rockets go into the moon looks like, but does anyone really know what intelligence is going to look like in five years? It's far less predictable, I think. ah So here's some of the comments that caught my attention, that the models need freshness.
26:04
coherence and contextual awareness. That's an interesting thing. They need to be up to date, have up to date knowledge with diverse viewpoints. They need real time access to human discourse. There's a lot of interesting sort of takes on what intelligence is going to look like and how it has to work. Now, does it really need real time access to human discourse? Whereas is that just in here because
26:34
One of the things they have is Twitter. you know, maybe it's just saying what we got as opposed to something. And if you have all these things, in theory, you can get objective and relevant insights. So you can see the language is sort of developing on what this looks like. Now, I thought that was pretty soft, but I thought the language was interesting. The one phrase that got my attention is, I think their target is...
27:01
the exact same target Tencent has, which I mentioned in the last podcast. This is from it. They're targeting high frequency, high value use cases in uh both consumer and enterprise AI apps. Now that's kind of where Tencent is focused as well, which is we need to focus on one high frequency, but more importantly, is high value because the variable cost structures of these things
27:30
you can't give them away for free. So you got to do something that people will pay for. You know, that's Enterprise applications are two of those. On the consumer side, not as clear. I think people don't really know what the high value monetizable use cases are on the consumer side. There are some subscriptions, but that money doesn't really add up enough to cover, you know, right now what you're seeing is you're seeing a lot of these AI companies dial back what you're allowed to do.
28:00
You're running out of usage very quickly on the consumer side. But if you're paying $200, $300, $500 a month for a coding tool, yeah, they're doing fine. Well, no, mean, it depends how much they use, I guess. OK. So that's interesting. This whole step two thing is a lot fuzzier. But I like that it was very consistent with what I was just talking about with Tencent. And then talks about this has to be vertically integrated. So the apps, Croc, XAI.
28:30
Then you have the data centers, Colossus 1. Now Colossus 2 is being built. I think it's pretty much online. They have a TeraFab, what they call the TeraFab, which is basically a partnership with Tesla to do chip manufacturing. And then they have something called MacroHard, which is funny. It's play on Microsoft, so it's a software company, but instead of Microsoft, it's MacroHard. And this used to be...
28:59
Described as some sort of software operating system now. It was described as an agentic AI platform. I Don't know what that means But the point of that is you can see they're doing the full AI tech stack all the way from chips data centers apps the full thing Okay, pretty consistent with what we've seen it Tencent and Alibaba Then they got to the part that really kind of caught my attention I was on the plane to shaman as I was reading this and this is the part that really stopped me where he
29:29
the S1 starts to go into what matters in compute to make these things competitive, to make them perform. What are the main pieces of the cost structure that have to be attacked over time? And here, I'll just give you the summary. I'll put in the slides from the S1 as well. Well, not the slides, but screenshots. Basically, the argument is, look, intelligence mostly follows from how much compute you throw at it or how much compute you do.
29:59
which is a combination of the quality of the data and the quality of your model. But generally speaking, if you put more compute onto a question, you get a better answer. And it also depends on how much time you give the AI to think about it. If you give a question and get a flash answer, it's not as good as if you give it a day. So intelligence seems to be proportional to the amount of compute you deploy and how long it gets deployed.
30:30
So it's a game of sort of scale. Okay, if that's true, which it looks like it is, then the big question is going to be what is your cost per token? Token is sort of the unit of compute. And yeah, what is going to be the cost? Because if you can get your cost per token down, you can deploy more compute resources. You're going to get more intelligence. You're going to get better intelligence. So how do you get the cost per token down? He talks about
31:00
I don't know why say he, it's the S1, so it's not him obviously, but talks about sort hardware software integration, which I just mentioned, having end-to-end cluster level coherence. That's kind of what I'm trying to understand, this idea of coherence, why that matters for intelligence. If you look at something like, let's say Alibaba or OpenAI or Huawei, they will have heterogeneous compute.
31:30
They will use different types of chips. They will put them in different data centers that are maybe geographically dispersed. OpenAI is putting data centers everywhere. Microsoft is doing the same thing.
31:42
X is really an outlier. They didn't build data centers all over the world. They built one data center in one city and everything is close together and tightly integrated. They didn't use lots of types of different chips. They used one chip, basically the Nvidia, the H and then now they've used the GB300, I think. So there's this idea of end-to-end cluster level coherence. And the argument seems to be coming that, look, that works better.
32:12
I'm not quite sure I understand why that is, but when you look at how other people are building data centers and sort of the architecture of intelligence, X and SpaceX stand out as a very different approach. One type of chip, one data center only that serves the whole world. Everything is physically right next to each other and it's tightly integrated. Okay, I'm not sure I understand why that is, but it's clearly a point he's made multiple, the S1 makes multiple times.
32:42
And then there's the idea, okay, the need for intelligence, the workload, the demand is going to increase dramatically, mostly from agents, reasoning, which are multi-step, and then multimodal. So there's at least three major drivers that the amount of intelligence and compute is going to go through the roof. Therefore, this idea of cost per token sits at the center of everything. And that's what I get from reading this, that...
33:10
That is their number one KPI they're thinking about is cost per token for a frontier level model performance. I you can do cost per token for an old model. No, staying on the frontier and cost per token seems to be their primary KPIs. That's what I got from this. All right, let me read to you how they talk about it. This is from the S1. It's literally called why compute matters.
33:40
We believe AI leadership will be defined by the ability to rapidly scale compute capacity to support exponential usage growth and frontier intelligence. The training and inference demanded by advanced AI models require substantial computational resources. Reasoning models introduced in 2024 demonstrated that allocating more computational resources and giving models more time to process during inference
34:09
directly leads to higher quality intelligence. In addition, compute infrastructure with end-to-end cluster level coherence through tight integration across software and hardware systems enables more efficient, stable, and higher fidelity training and inference at scale, ultimately enhancing model intelligence and performance. OK, that's kind of the same thing, but yeah.
34:36
Yeah, you can see they have a definite business strategy within this in terms of what they're... They're building a different machine and the machine they're building has a very clear approach which seems different to me than a lot of others. In some ways it's the same. Okay. So why does the cost per token matter? They list a couple of reasons. If you can drop the cost per token, you get more frequent use by users.
35:05
because it's cheaper. If people use it more frequently, the model will get smarter. It also means that the, that's inference. At the training phase, if the cost per token is cheaper, you can train more frequently. New models, retraining existing ones. You can basically do more training overall. So more usage on inference, which makes you smarter and more training of the model, which also makes it better.
35:33
You can build, if the cost per compute is lower, I'm sorry, the cost per token is lower, you can build larger models. You can build more sophisticated models. You can do longer reasoning chains and more processing. Basically, cost per token plays out across the board, which ultimately gets you a better model. And if you have a better model at a good price, you should get more adoption and make it more accessible. They argue this is a network effect. I think it's just a feedback loop.
36:02
I think it's a flywheel. If you can make your model cheaper, it's easier to make it better. Better plus cheaper means people use it more, you grow, it gets smarter, it's a flywheel. They call it a network effect, I don't really think so. Okay, then it goes further, which was super helpful. They basically, if cost of compute, cost per token is the primary thing, how do you decrease it?
36:28
they break down the cost structure and tell you pretty much what they're targeting. So they say, look, it's a couple of things. It's the model cost. It's the compute infrastructure cost, which is a data centers, and it's the power cost. Those are the three big buckets of costs that determine your cost per token, ultimately your cost to compute. So they basically say, look, we're going to drive those numbers down. the two that they're focused on, or not that they're focused on all of them,
36:56
The two where they say they think they can get an advantage is infrastructure cost and power cost, not necessarily model cost. They didn't cite that. They said the second, those first two, which is interesting. So how do you drop the infrastructure cost, the cost of compute, the AI data centers basically? Well, you can lower the cost of construction. You put everything in one location and lots of other things. These people are all engineers, so they'll find a way to make it cheaper.
37:24
They can shorten the time it takes to build. You can build more frequently. I mean, this is kind of how he makes his rockets cheaper. One, they're always making the Raptor engine cheaper, but they're also firing very frequently. And so they're iterating very quickly. So they're kind of shortening the time to build, shortening the cost to build, and then decreasing the overall cost of the structure once it's built. All those things are being driven down.
37:53
the compute infrastructure costs. Interesting. And then they similarly talk about power is the same thing.
38:03
We need power plants; we need natural gas. Some of these companies are making nuclear power. And within all of these, you can make things cheaper and you can do them faster if you're vertically integrated.
38:16
It looks a lot like the other businesses. That's how it seems. So anyways, their approach, and I would call this their innovation marathon that they're running, is to stay at the bleeding edge of intelligence performance at the lowest cost per token. That I think is, I think those are their two primary KPIs that they're sort of running on as their digital marathon. I'd call that the sustained marathon, innovation marathon.
38:44
It was that's of what I wanted to go through. But as I was reading this, suddenly all these little press releases I've been seeing started to make a lot of sense. Oh, I get why you're doing this in Memphis. I get why you're building one massive center. I get why you've vertically integrated into chips and you've gone all the way down in the AI tech stack into one integrated hardware software model. I get why you're integrating the compute and the power infrastructure together. You're trying to drive down those two costs in particular. I get it.
39:14
And then the most recent thing, which this will be my last point, is this idea of data centers in space, which has kind of been a talking point in meteorology. I think they're totally serious about it. So they talk about orbital compute versus terrestrial compute. These Memphis things, this is terrestrial compute, data centers on the ground, orbital compute, they're in space. And the argument I get from reading the S1 is, look, the demand for this
39:44
is going to go exponential. There is no way to meet the demand with the energy resources on the planet. It's too much. We're going to have to go up into space and to pull from the sun directly. And also it solves your cooling problem. Now, if you do the AI compute in space, the problem is you got to connect it to everybody's phone on earth. Well, if you have Starlink, you can do that.
40:11
I don't really understand the technology of that well enough to know if that's actually true or if it's just tying the two businesses together and saying, hey, look, one is an advantage for the other. I'm not sure about that. But yeah, that was kind of the idea, that if you're going to get hundreds of gigawatts of power, the sun is probably the only thing that can meet the accelerating demand. Maybe. We'll see. Anyway, so that's kind of a funny idea, terrestrial compute versus orbital compute.
40:42
And that's pretty much what I wanted to talk about. I'll finish up here with some crazy ideas. Now there's a massive valuation on this up, which I'm not going to talk about at all. But if you're going to have a massive valuation, you probably need a big story. So in the S1, there's a section called future markets, which is, what could this business be one day? I thought that was pretty crazy. I'll just read the list. uh Point to point.
41:12
terrestrial travel as a business. Basically, get in the rocket in New York, you'll be in Shanghai in 23 minutes or whatever it is. Space tourism, I would totally do that. In orbit manufacturing, when actually when they talk about SpaceX, which I didn't talk about at all, the rockets, there's actually a very interesting approach where they're building several things to get to space and to build a colony, which is
41:40
You have to be able to manufacture in space and on the moon. You have to be able to do power generation in space and on the moon. You have to be able to logistics on the moon and between the planets. Well, that's the rockets. But they're building three big buckets of capabilities to make this dual. Basically, you need manufacturing, power, and logistics, transport. OK, so this idea of in orbit manufacturing is not crazy because they're going to have to do that to get a colony on the moon or Mars or whatever.
42:10
uh passenger and cargo transport to the moon and Mars, uh energy production on the moon and Mars, manufacturing on the moon and Mars, asteroid mining, which is pretty cool to think about. And it was lots of markets. say this is the largest TAM in history, $28 trillion. You know, who knows? So anyways, that stuff's pretty fun. Anyways, that's it for that one. I'm not going to go into that too much.
42:40
It's interesting. And I thought the tech strategy I thought was very helpful to see the same pattern over and over. I thought the breakdown of the costs of AI was helpful and to see that they're kind of where they're going and why that makes sense. The three concepts, which I think are useful. Number one, sustained innovation as a digital marathon. I think that's totally what they're doing. Repeatability in a business model is quite powerful.
43:10
I think that's something people don't talk about nearly enough. And then this idea of sort of the four quadrants, the four different terrains that will determine your strategy. his terrain is what it is. It's technology and frontier and lots of engineering. this is a pretty good breakdown of a shaping strategy for that quadrant. OK, so some good lessons in here. I thought that was fun. Hopefully that is helpful.
43:36
As for me, it's just Saturday and I'm sort of catching up on podcasts I've missed in the last couple weeks. Articles have been going on. I'm pretty much caught up on articles. I think I have one more podcast behind. I'm going to start writing about CATL this weekend, which is fascinating company. It's all batteries. I had to do a lot of reading about the chemistry of batteries, which was kind of fun, but not really my area. And it's a bit off topic, but the business model is interesting. And the batteries go into everything, robots.
44:05
Data centers super important EVs obviously I mean as the world gets smart and automated in robotics Yeah, it turns out batteries are a big deal and see ATL is the number one company in the world for rechargeable So anyways, that'll be the next content going out next couple days. That's it for me Hope everyone is doing well, and I'll probably talk to you in a couple days. Bye. Bye