- 2 hours ago
Category
🤖
TechTranscript
00:00There is a lot to discuss. I think, should we just start with Cursor?
00:03Let's do it.
00:03Because it's right in front of us. The focus of this morning was, oh, well,
00:08Elon Musk has been quite honest about X-AIs need to catch up in coding vis-a-vis
00:13anthropic, open AI. What I'm trying to understand is what this means for the
00:17models themselves. Do you have a sort of thesis on that?
00:20Yeah, look, I've said this publicly, but I think about SpaceX as there's five
00:24layers of the company or business. Layer one is launch. This is foundational
00:29it's kind of the future of the company. Layer two is connectivity, which is in its
00:35early stages with Starlink and Direct to Sell, but kind of proven. And I think the
00:41compounding is pretty easy to underwrite. Layer three is compute, which today is
00:46terrestrial and in the future will be orbital as well. I think that the Google
00:52ananthropic deals kind of show that that is very likely to work. And I actually want to spend more
01:00time on terrestrial compute later because I think it's a very, very exciting and underappreciated
01:05part of the company today.
01:08Layer four is the model layer. I think that's where probably rightfully so most investors don't know
01:15how to assess where SpaceX, SpaceX AI is today. And so I think we should talk about that more.
01:22And that's where cursor fits in. And then layer five is all the other bets. It's things that,
01:26you know, people today don't even know how to think about it. That's TerraFab. That's
01:31a moon base someday that has a rail gun launching, you know, orbital satellites in space.
01:38But look, on the cursor deal, I think people are underappreciating just how good that team is
01:45and also how well they work with Elon and the company.
01:48Our last guest called it an acqui-hire. It's just, it's a $60 billion acqui-hire.
01:52I mean, I don't think when a company has a $3 billion plus run rate is the reported
01:57number in, you know, three, as a three-year-old company, I don't think it's only an acqui-hire.
02:03But I do, I do think that this is one of the best teams that's ever been assembled
02:08in the cursor team. And I also think they've proven over the last few months that they can
02:14keep pace with Elon. Like it is, unless you've seen it, it is impossible to explain the intensity
02:21and work ethic of Elon. And Michael and the team have kind of shown that they can keep pace
02:31and fit into this much broader ecosystem, adding value, you know, on TerraFab, planning their
02:39orbital satellites, model layer, you know, cleaning up data to train models, you know,
02:46helping plan infrastructure for the future. It's just like, it is such a strong team that can go
02:52all the way to the bare metal in terms of understanding everything needed to build a model
02:57of the future. And so I like, no, it is not an acqui-hire, but I cannot stress how good
03:03that team
03:04is. And for me personally, it makes me much more optimistic that SpaceX AI will win or at least be
03:11one of the winning players on the model layer.
03:14You're someone that's embedded in the teams. When I say embedded, I mean, you were able to go in,
03:18work with the team, see what they're working on across SpaceX, XAI, X. And when I said that you
03:23were coming on the show, that was pretty much one of the questions that we got for you is,
03:27how do you think Elon Musk takes that cursor team and integrates it and moves quickly? You know,
03:33there's a particular Musk way of doing things, but, you know, this has happened three days since
03:40an IPO. It's pretty fast.
03:42Look, I think there was a couple months, you know, I don't know what the right technical word is. And,
03:49you know, I'm on live TV and there's a recent IPO, but I think there was a roughly two month
03:53trial
03:53period. And I think that only 1% or less entrepreneurs could survive that trial period of
04:02working very closely with... You mean like cursor passed the test?
04:05Yeah, I think cursor passed the test showing that they can keep pace, that they can operate without
04:11making mistakes, that they can actually elevate what's already one of the most talented teams in the
04:18world. And so I don't view this as like three days after IPO, there's, you know, something that came
04:24out of nowhere. I view this as there was, you know, many months leading up to that two month period,
04:31and then two months of like deep, deep partnership. And so like, I personally think the probability that
04:38this is an accretive acquisition is like 99.999% because the trial was so intense.
04:44Sean, before we get to the near cloud business, I think that's like really important and
04:48timely. We're in day three of trading as a public company. And a lot of people will make a lot
04:55of
04:55the milestone of being worth more than Amazon. Maybe that happens at the close, maybe it doesn't.
04:59And I'm guessing you'll tell me that you'll distribute to the LPs when the time comes and
05:05you expect them to hold the shares. You know, what is your kind of longer term position on SpaceX?
05:12I mean, look, for me on a personal level, and I say this as someone, I think I understand this
05:18business very well, like better than, you know, most investors, better than almost any investor.
05:25I am personally going to hold my shares in this company forever. Like quite literally. I think
05:30that, and sure, there's some things that could change politically or, you know, you know, in the
05:37macro 10 years from now. But this is the biggest vision, the biggest mission of any company in
05:43history going after the biggest markets of any company history with the biggest moat. And so I
05:47just think the compounding is going to be like, it's, it's hard for people to even imagine the
05:52compounding. Which part is the moat, Sean? Sorry to interrupt you, but the five layer cake that you've
05:57outlined. Yeah. Which one? So obviously a great question. By far, the biggest moat is launch,
06:05which, you know, we call people rocket scientists for a reason, or at least in the past,
06:09you know, that was an expression. Like the engineering that has gone into building Starship,
06:15I believe it's the, I believe it's the hardest, like single engineering project that any company
06:23has ever done. I think it's, to me, it's meaningfully beyond like ASML, like intrinsic complexity of
06:29building EV machine. I'm happy to go into this. So I would say moat number one is the launch business,
06:34which is just so far ahead of everyone else. Moat number two, and, and for what it's worth,
06:40I think there's this kind of weird thing where I do think that people right now are really
06:45underestimating the terrestrial compute. So let's go there. Well, I'll, yeah, I'll go there in one
06:51second. And, and I think that the moat for terrestrial compute is not launch. It's the quality
06:57of the team that's been assembled and, you know, in kind of this very heavy distillation process over,
07:05over, over 20 years of finding, just bringing in the most talented people from Stanford, from MIT,
07:13et cetera. And only 1% or 0.1% actually survived this distillation process where they proved that
07:19they can work with the level of intensity while simultaneously level of team orientation without
07:24making mistakes, et cetera. So that team has figured out how to solve basically every hardware problem
07:30in the world. And when you apply them to something like terrestrial data centers, they're now like
07:35they're building data centers faster and better than anyone else. And so the team is a second moat
07:41that I know that seems like, Oh, you can just go create a team. No, you cannot because it took
07:4620 plus
07:47years of distillation. Um, but I think that we're in this weird meta stable period where when we go five
07:53years from now, I think that launch will again be the moat because I think that orbital compute will
07:58be dominating net new inference compute, like just absolutely dominating. But I think for the next
08:04five years, like the team feeds into dominating terrestrial compute. So there's now you want to go
08:09to the terrestrial compute. So the state of play is this, uh, the XAI team, SpaceX AI build data center
08:15quickly. And I think your argument is that on a dollar per token basis, highly competitively operate
08:19them. If you go on to a map of, uh, Tennessee, they have Colossus one, Colossus two and Colossus
08:26three. And a big question a lot of people had was, well, hold on. It's amazing to get $2 billion,
08:31uh, a billion dollars per month from Anthropoc and Google, but why is that capacity not being used
08:38to train the latest, greatest models at the model layer or run inference on what Grok's already doing?
08:43seemed like a blinder of a business decision, but longer term, like it's hard to reconcile.
08:51This is my personal opinion. Um, kind of my assessment of the situation. First of all, one,
08:58the way I think that SpaceX in terms of, uh, in terms of terrestrial compute, like,
09:05and you know, you mentioned Amazon,
09:08Amazon and Microsoft are incredible businesses. I have massive respect for those, you know,
09:13founders for the, for the businesses, for Satya running, running Microsoft, but a lot of their
09:23market cap comes from their cloud businesses and, you know, also for their legacy businesses in the
09:29case of, you know, delivery or, um, or the whole Microsoft office bundle and operating system.
09:38But a lot of it is their cloud businesses. We're in a generational transition of cloud from kind of
09:46legacy CPU based jobs to AI cloud. I strongly believe that SpaceX, SpaceX AI is the best in the
09:55world at building this next generation of AI clouds, like not even close an order of magnitude better at
10:04building terrestrial compute in this AI era than anyone else. And it's not to say that they're,
10:08the other companies are bad. It's to say you have the best rocket scientists in the world
10:13are now using their skills to build terrestrial compute. And what that leads to, in my opinion,
10:21is I think people are really underestimating how much more compute is going to come online next year
10:26and the year after for SpaceX. And that, you know, means I really believe SpaceX is going to have a
10:34surplus of compute. And I think it's very rational. And I'm, as an investor, I love that they are
10:40offering compute to Anthropic and Google because one, it brings in a ton of revenue that I think
10:47in an IPO shows the world that all of this CapEx build out is not for nothing. And it's not
10:54all or
10:54nothing based on the model layer. There's this, you know, massive value that the company has a dial
11:01of how they choose what to do with it. Um, kind of next point. I think that people are underestimating
11:09that most of that compute is kind of Colossus one, which is roughly 10 miles away from Colossus two
11:16and Colossus three. It's heterogeneous GPUs. In the prospectus, the team is saying at SpaceX,
11:22SpaceX, as early as 2028. Um, in a moment, we're going to bring up a Bloomberg rendering of the,
11:28of the design of, and that's actually the SpaceX one, right? You, I think you're probably in the
11:33camp of people that, that think they can do this quickly. Um, where do you stand on your,
11:38your data or the center or orbital data center outlook for SpaceX?
11:43I stand incredibly bullish, just incredibly bullish. Yeah. Um, I'm not surprised by that.
11:49Yeah. But, but like, I, I think it's rationally, rationally bullish. I think something that a lot
11:56of people are struggling with here is most, like most people just never thought about what is a
12:01satellite? How do you build a satellite? What are the components of a satellite? What are the
12:04constraints around launching these satellites? And I think that people that have been following
12:09Starlink closely are very well positioned to assess because you can extrapolate out on both the
12:15economics, but also the engineering and physics. Exactly. The orbital compute satellites are very
12:21similar in many, many ways to Starlink satellites. Like I truly believe in Elon said this last week
12:29in a video on orbital compute, that the intrinsic complexity of making a Starlink satellite is probably
12:37a little bit greater than an orbital compute satellite. So I know this sounds kind of crazy,
12:42but Starlink is on the satellites themselves. Starlink satellites are more complicated than
12:47orbital compute satellites. That said, orbital compute satellites are bigger. For them to make sense,
12:53they have to be much bigger. You basically want to put a whole server rack in space just in terms
12:57of
12:57doing inference on a model. You kind of need a minimum size for it to be useful. And you can't
13:03launch these
13:04bigger satellites unless you have Starship. So almost all of the difficulty of orbital compute
13:11satellites is in Starship. It's in the launch vehicle. And so if you want to estimate when
13:17orbital compute is going to be scaling, it's really a question around like when is Starship going to be
13:25We're showing the videos. This is this is file right of Starship. And you know, the basics of it are
13:30that on paper,
13:32you know, 135, 150 metric tons payload to orbit. And I think there's some idea that in the future in
13:38iterations, 400 metric tons.
13:41I think the company said, you know, 200 metric tons is pretty like pretty achievable. 400 metric tons.
13:49Maybe. I don't know. It seems likely, but I don't know. You just make it. There's something with rockets where
13:54the taller
13:55you make it kind of the easier things become the surface area to volume ratio and that benefits you with
14:02drag.
14:04So it seems very likely, but manufacturing at that scale is harder.
14:08Can you compare and contrast then? So you outlined in great detail why you're bullish on the on the terrestrial
14:13data center
14:14neocloud business. Where does the orbital data center business compare over just a like reasonable timeline for you?
14:21Yeah, again
14:23And I love the question. I appreciate the question. I think people are looking at orbital compute and they think
14:29oh wow
14:29This sounds so crazy and I wasn't even I wasn't thinking about it two years ago
14:34So it must take a long time and a lot of Elon's timelines in the past have have been wrong,
14:39which you know, he says himself
14:40Um, but if you look at what an orbital compute satellite is, it's basically
14:47some compute element whether that's a gpu or tpu or tranium or you know
14:51proprietary
14:52He said they have reference designs for all of them. Yeah, but so
14:55There's the compute and you want about a server racks amount of compute. There are solar panels like that's easy
15:02fully kind of proven
15:05uh
15:05SpaceX has huge experience building solar panels and the third main piece are
15:09Radiators so this is what takes all the heat from the compute chips and
15:15Radiates it into space via you know
15:18Via black body radiation
15:22Those are the three main things and then you also have laser links to connect to other satellites or
15:28You know connect to the ground and radio to connect connect to the ground
15:33of these things
15:35Every individual component here is kind of fully
15:39Fully
15:40Fully proven by spacex except for
15:42The compute side
15:44But that is not it's just not that hard
15:47And so the satellites themselves are pretty easy to make that's a rendering again a bloomberg rendering of what you're
15:54describing
15:54Yeah
15:55So to me
15:56It really is a question of when starship is flying
16:00I don't I don't I don't understand why spacex wouldn't be able to have an orbital compute satellite up within
16:06you know
16:06Six months or so of the first
16:09Store starship payload flight
16:10We have 90 seconds left in the show. I can't believe it
16:13Um, we need to address the open ai and anthropic elephant in the room, which is on paper
16:18They are very similar if you read the prospectus and the TAM that spacex outlines
16:22Is there room on earth and outside of the earth for all three is I think
16:29intelligence is
16:31You know the most important it's the defining
16:34Capability of our time and I think there's unlimited demand for intelligence
16:40I think it's very clear that spacex will be in my opinion will be
16:44The dominant provider of compute for intelligence and I don't know what's going to happen on the model layer
16:49But there's absolutely space for all all of these companies to just see 15 seconds
16:54If you had like a kind of black swan memo right now for the firm sequoia, what would the headline
16:58be?
17:00What's the biggest focus for you?
17:03I think it's probably
17:06Regulatory risk
17:07Regulatory risk. Okay, let's leave it as a cliffhanger and you're always welcome here back on the show
Comments