Latest / In the Loop AI / THIS WEEK IN AI: SpaceX IPO TODAY, AI 1 Satellite, Claude Fable 5

Transcript

THIS WEEK IN AI: SpaceX IPO TODAY, AI 1 Satellite, Claude Fable 5 · In the Loop AI

0:00
Josh: The question on everyone's minds, should you buy the SpaceX IPO?
0:04
Josh: By the time you are hearing this, chances are SpaceX is about to go public or
0:07
Josh: they have just gone public at what is going to be the largest IPO in history.
0:13
Josh: And it's not even close. I think we have very strong takes and opinions on where
0:16
Josh: SpaceX goes. I find the question that everyone is asking me is,
0:19
Josh: should I buy? Is this overvalued?
0:21
Josh: It's going to be trading, I mean, right now, close to $2.1 trillion beforehand.
0:25
Josh: That's an outrageous valuation.
0:27
Josh: And I feel very confident that we have some pretty good takes that are going
0:30
Josh: to point in the direction of
0:31
Josh: actually yeah not only is it a good buy in fact it's undervalued and a lot of
0:36
Josh: the news that we got this week from the s1 from the new spacex releases backs
0:41
Josh: that up so there's a lot of really strong prevailing wins pushing spacex forward
0:45
Josh: and we are going to now address the question should you buy spacex shares should
0:49
Josh: you get invested in the ipo and more importantly why.
0:54
Ejaaz: Yes, the answer is yes. But before
0:56
Ejaaz: we get into it, let's talk about the numbers of the IPO very quickly.
1:00
Ejaaz: So they're raising $75 billion, which is three times larger than the largest
1:07
Ejaaz: IPO that has happened so far. That was from Saudi Aramco back in 2019,
1:11
Ejaaz: and they raised just under $30 billion.
1:14
Ejaaz: This is roughly around 3x that, but the demand for the $75 billion raise is
1:19
Ejaaz: already 4x, as you're seeing on the screen.
1:21
Ejaaz: There's over $250 billion worth of demand with one institution or a couple of
1:26
Ejaaz: multiple institutions offering up $10 billion chunks just to get involved here.
1:30
Ejaaz: Now, if you're looking at like the price per share, roughly,
1:33
Ejaaz: SpaceX's S1 prospectus has put it at around $135 a share.
1:38
Ejaaz: Now, what's interesting about this is typically with an IPO prospectus,
1:41
Ejaaz: you provide a range, you provide a range and you try and find where the match is.
1:46
Ejaaz: SpaceX just went right to the high end of that
1:48
Josh: Range and said it's.
1:49
Ejaaz: $135 per share, which values it at $1.75 trillion.
1:55
Ejaaz: But recent reports from Financial Times and Bloomberg says that this number
1:59
Ejaaz: is actually now close to $162 per share, which gives them a crazy $2.1 billion valuation.
2:07
Ejaaz: Now, the demand for this thing, Josh, is absolutely insane.
2:10
Ejaaz: They reserved, I think it was around 30% of the share allocation that they're
2:15
Ejaaz: IPOing with for the retail, for the regular average person. And then that was
2:19
Ejaaz: distributed to the likes of financial institutions and banks like Charles Schwab,
2:22
Ejaaz: as well as brokers such as Robin Hood and
2:26
Ejaaz: The demand initially required you to have like something crazy,
2:28
Ejaaz: like half a million dollars in your bank account in some cases to get access to this thing.
2:32
Ejaaz: But they reduced that kind of like limit. But it is the most in-demand thing.
2:36
Ejaaz: SpaceX is the most in-demand company in general, like whether it's just to buy
2:40
Ejaaz: the IPO or whether it's just to like work there. It is a crazy scene that we
2:44
Ejaaz: haven't seen before. Largest IPO raise ever.
2:46
Josh: Yes. And it is also by far the most accessible IPO ever. We mentioned this in
2:50
Josh: a previous episode, but 30% of the entire share float is being made available
2:54
Josh: for the public, for retail, as a part of this IPO.
2:56
Josh: This is three to five times more than the average, and that means that a lot
3:00
Josh: of people are able to participate. So there's a chance if you're watching this,
3:03
Josh: you've actually been allocated IPO shares of SpaceX at the listing price.
3:07
Josh: You don't have to pay that premium.
3:08
Josh: We're going to see the market making happen today as you're watching this to
3:11
Josh: see what it's actually going to launch at.
3:13
Josh: But I suspect a lot of people watching
3:14
Josh: this episode are actually already shareholders and owners of SpaceX.
3:17
Josh: And that's such a cool thing that everyone is able to participate.
3:20
Josh: Now, in terms of the valuations of SpaceX, why we believe this is undervalued,
3:24
Josh: why we believe SpaceX is going to be such a remarkable company, there's a few pillars.
3:29
Josh: The first is Starlink. It is internet and space from the sky.
3:33
Josh: The second is manufacturing, their ability to manufacture Starship and get payload
3:38
Josh: to orbit at a very low cost per kilogram.
3:40
Josh: And the third, which is one that I really want to talk about right now,
3:43
Josh: is AI data centers in space.
3:45
Josh: We've talked about this a lot. what does this look like? How does it work?
3:48
Josh: Well, earlier this week, we finally got an answer, and it comes in the form
3:52
Josh: of a satellite named AI-1.
3:55
Josh: AI-1 is the satellite that SpaceX is going to park on Starship,
4:00
Josh: send into orbit and actually put these things into space. And they're going
4:03
Josh: to be training data centers. They're going to be beaming data back down to earth.
4:07
Josh: And now we finally have a rendering and we have some specs on what it's going to look like.
4:10
Josh: In terms of power, there is 150 kilowatt peak with 120 kilowatt average on each
4:16
Josh: one of these satellites for reference.
4:17
Josh: That's about 100 American homes worth of power per satellite.
4:22
Josh: And then we're getting about 70 kilowatts per ton. So that's the efficiency
4:26
Josh: part of the launch metrics where,
4:28
Josh: The more energy we can get per ton, the cheaper it costs. It's lowering those costs.
4:34
Josh: We have compute slots that are interchangeable. And I think this is a really
4:38
Josh: underrated part is the GPU modules that you're seeing on the screen on the animation
4:41
Josh: we were just showing, they're actually hot swappable.
4:43
Josh: So you could have NVIDIA chips, you can have Google chips, you could have Amazon chips there.
4:49
Josh: What they're going to test this with initially is going to be NVIDIA's Vera
4:52
Josh: Rubin platform. That's going to be on version one of these things.
4:55
Josh: And in terms of the radiator because clearly no one believes that spacex is
4:59
Josh: capable of figuring out thermodynamics in space given the fact that they have
5:02
Josh: how many thousands of satellites currently operating in there well they figured
5:05
Josh: it out they said they're going to have 110 square,
5:09
Josh: meter liquid radiator and basically it's just gigantic sails that we're seeing on screen right now,
5:15
Josh: to dissipate the heat into deep space deep space is a huge place even if you're
5:19
Josh: in low earth orbits geosynchronous orbit and the wingspan of those will extend
5:23
Josh: to about 70 meters wide which is wider than a Boeing 747.
5:26
Josh: So this is AI data centers in space. When we say that, this is what it looks like.
5:31
Josh: This is how it's going to function. It's a really big deal.
5:34
Ejaaz: I want to shine a bit more light on this satellite because previously,
5:37
Ejaaz: like Elon has just made bold claims saying that, you know, the valuation that
5:41
Ejaaz: we're raising SpaceX at now at a $2.1 trillion valuation is primarily based
5:45
Ejaaz: on AI data centers in space.
5:47
Ejaaz: And then they kind of signed a few deals recently with Google and Anthropik
5:50
Ejaaz: saying, yeah, they're going to be buying these data centers in space,
5:53
Ejaaz: but we still didn't know how. Now we have the answer and we have a bunch of
5:56
Ejaaz: the numbers and how it's built. Now, to be clear, what does this satellite look like?
6:01
Ejaaz: Picture a refrigerator that you have at home with 70 meter wings.
6:05
Ejaaz: This thing is 95% wing, basically.
6:09
Ejaaz: And that refrigerator is essentially a GPU server rack.
6:13
Ejaaz: It stores around 72 NVIDIA GB300 GPUs. And the total compute power that it has
6:20
Ejaaz: is around 150 kilowatts. Now, immediately after announcing the satellite,
6:25
Ejaaz: Elon also stated that he's filed to launch 1 million of these data center racks,
6:31
Ejaaz: these satellites, these AR-1 satellites into space.
6:35
Ejaaz: If he is able to pull that off, the total compute power that will be orbiting
6:39
Ejaaz: in space will be around 120 gigawatts.
6:41
Ejaaz: That's around 3 to 4x of the total data centers that currently exist on the
6:47
Ejaaz: entire surface of Earth right now.
6:49
Ejaaz: So that is a hell of a lot of compute, but more so currently there's around
6:53
Ejaaz: 14,000 satellites in total existence across every single other company in low orbit space right now.
7:00
Ejaaz: So Elon is like aiming for like multiples more of that for this particular satellite.
7:04
Ejaaz: Now, the wingspan is really important. You must be thinking like,
7:07
Ejaaz: you know, 70 meters is absolutely ridiculous.
7:09
Ejaaz: Radiation, you need to disperse the heat. It's kind of like hard to do that
7:12
Ejaaz: in space. In fact, it's impossible to do that in space. So you need some kind
7:15
Ejaaz: of like really ingenious cooling mechanisms. That's what that wingspan is for.
7:20
Ejaaz: Yes, it is large, but also
7:22
Ejaaz: You're in space. You have all of that space to be able to do that.
7:25
Ejaaz: And the way that these satellites are designed is different to the ground data
7:28
Ejaaz: centers in the sense that they're going to be beaming lasers between each other
7:31
Ejaaz: to form an interconnected grid that they can transfer data between each other.
7:35
Ejaaz: And that is the sole mechanism that is required to train and inference models,
7:39
Ejaaz: which is the entire point around SpaceX.
7:41
Ejaaz: Now, it's important to understand when you're looking at the valuation,
7:44
Ejaaz: when you're looking at the price, you might be thinking, well,
7:46
Ejaaz: I don't know if this is overvalued.
7:49
Ejaaz: It kind of seems like overvalued, even though I like what Elon is doing,
7:52
Ejaaz: you have to understand that the whole premise is you're getting three companies
7:56
Ejaaz: for one share. It's SpaceX, yes, to be the kind of like highway to space.
8:01
Ejaaz: But then you have like the payloads, which is like all these GPUs that they're
8:04
Ejaaz: going to be training on the sun's energy, becoming a Kardashev type one civilization,
8:08
Ejaaz: which all sounds grand, but like
8:09
Ejaaz: we now have the numbers and the mechanism to be able to pull that off.
8:12
Ejaaz: And then you have the model itself under XAI. And then you have the distribution
8:15
Ejaaz: through social media platform, through app stores and all that kind of stuff.
8:18
Ejaaz: And it's not like Elon is depending primarily on his own companies to service
8:23
Ejaaz: all this demand for Grok entirely.
8:25
Ejaaz: He's signed major deals in the last, I think, three weeks with Anthropic,
8:29
Ejaaz: Google, and Cursor, which equates to around $56 billion per year.
8:33
Ejaaz: So the previous critics were like, I don't know if SpaceX is going to be profitable in general.
8:38
Ejaaz: Now we have the answer, like all this money that's been coming from essentially
8:41
Ejaaz: neocloud payments for Elon's existing colossus one two and three data centers
8:46
Ejaaz: on earth is paying for itself it's paying for this entire expedition
8:49
Josh: Oh it's an absolute monster of a company and now i want to talk about a second
8:53
Josh: pillar because we've seen ai1 what that looks like what data centers in space
8:57
Josh: look like we also have visual graphics of
8:59
Josh: the new startling satellites and we've been talking about these new startling
9:02
Josh: satellites for a long time now version two and most recently version three which
9:05
Josh: is going up on starship and the
9:08
Josh: stats and analytics between these two comparing them is pretty remarkable the
9:12
Josh: the first generation of this, version 1.5 of Starlink, had 24 gigabits per second download, which,
9:18
Josh: pretty small. If we look at version two of the satellites, which are what's
9:22
Josh: available today, that's what's going up in the Falcon 9 rockets,
9:24
Josh: the smaller ones, those have about 96 gigabytes per second, and they're able
9:29
Josh: to deploy about 27 of those per launch.
9:32
Josh: So that gives you about a total per launch of 2,600 gigabits per second.
9:37
Josh: Now, when you look at version three satellites of Starlink, that's where the
9:42
Josh: numbers get pretty silly.
9:43
Josh: It's a 20 times multiple per launch on bandwidth relative to version two this
9:49
Josh: is enabled with starship we talk about starship all the time.
9:51
Josh: They've been doing really successful test runs they're going to continue that
9:54
Josh: program once it starts becoming reusable they're going to put these satellites on it
9:58
Josh: and it's going to deliver 61 000 gigabits per second to the internet that is
10:03
Josh: 60 satellites that is a 20x multiple that is how you get to a true worldwide
10:08
Josh: web that exists in space and when you think about what spacex is doing here
10:11
Josh: they're creating a high bandwidth clone of the internet that exists here except
10:15
Josh: in space What happens with the internet here?
10:17
Josh: We actually run wires underground through the oceans to connect the world.
10:22
Josh: If we have satellites, you remove a lot of the threat vectors,
10:26
Josh: you remove a lot of the sensorability, you create this truly open and accessible
10:31
Josh: network that's not only available to anyone with infrastructure,
10:33
Josh: but available to anyone with access to the sky.
10:36
Josh: And building that and making that available to everyone is such a remarkable
10:41
Josh: business. And we've seen that reflected in the Starlink subscriber numbers.
10:44
Josh: We have a chart here that shows the Starlink growth.
10:46
Josh: And it has been the most perfect exponential chart I think I've probably ever
10:50
Josh: seen in my life. And it is the Starlink subscriber growth.
10:54
Josh: As of when is this? May 23rd.
10:58
Josh: It is now sitting at 12 million users. And you could see they've been adding
11:02
Josh: an exponential increment. I mean, the latest 2 million users were added only
11:06
Josh: in the last 81 days. They're averaging 25,000 and climbing.
11:10
Josh: So Elon and the Starlink team, they just recently revealed on an interview that
11:13
Josh: they published a few days ago, the new version of these Starlink terminals,
11:17
Josh: which are even smaller, even higher bandwidth, even more accessible in terms of price.
11:22
Josh: And it's just this unbelievable service that it's going to be a no brainer not
11:25
Josh: to have. If you ever require Internet anywhere that doesn't need fiber optics,
11:29
Josh: you should be a Starlink subscriber. And that's what we're seeing in the growth.
11:32
Josh: So it's really it's an unbelievable company.
11:34
Josh: Starlink is one of those pillars that's just like it's it's crazy what they're
11:38
Josh: doing and no one else can come close to this we
11:40
Josh: we have blue origin which huge fan of i really hope they figure it out but their
11:43
Josh: rocket just exploded on the launch pad last week and their whole program set
11:47
Josh: back and they're the only ones that were close
11:48
Josh: even if everything at blue origin worked as perfectly as they hoped for they're
11:52
Josh: still five years behind at minimum probably 10 this is a full decade lead they
11:56
Josh: have so there's essentially a monopoly on space that spacex owns
12:00
Josh: and as it becomes more on demand to send mass to orbit, there's only one company
12:04
Josh: that can do it. And they can do it damn well.
12:06
Ejaaz: So this might be a crazy take, but Elon's superpower
12:11
Ejaaz: actually it's not a crazy day this is obvious is scaling hardware from zero
12:14
Ejaaz: to one better than anyone else on the entire world he's done it with electric
12:18
Ejaaz: cars he's doing it with robots and now he's doing it with space rockets now
12:22
Ejaaz: there's another advantage that he has which translates directly from that
12:26
Ejaaz: Starlink right so you just mentioned that um you can repurpose starlink's production
12:30
Ejaaz: lines to make this new ai1 satellites in fact arguably it's easier see with
12:37
Ejaaz: when you're creating these SOLID-X satellites, they have this phased
12:40
Ejaaz: Array antenna, which basically is meant to speak to the dishes that are on Earth
12:44
Ejaaz: to communicate and beam this internet back down to Earth.
12:48
Ejaaz: That is a really tricky thing to produce, but a satellite itself is fairly easy.
12:51
Ejaaz: Now, with the AI-1 satellite, it doesn't need this phased array antenna.
12:55
Ejaaz: In fact, you just need the laser projections to be able to communicate and talk
12:59
Ejaaz: with each other. I mentioned it earlier, like transferring data,
13:01
Ejaaz: training your model, inferencing the model, that is arguably easier to do.
13:05
Ejaaz: And so Elon already is not starting from scratch here. Normally, he does zero to one.
13:10
Ejaaz: He's kind of starting from 0.5 to one to be able to pull this off.
13:13
Ejaaz: Now, in terms of timelines, I think he's aiming for the end of 2027 to have
13:18
Ejaaz: a full gigawatt worth of these AI-1 satellites out in space.
13:23
Ejaaz: So this is a very real thing. The timelines are, in my opinion,
13:26
Ejaaz: much shorter than we think.
13:27
Ejaaz: And he already has some of the world's leading AI labs committed,
13:31
Ejaaz: at least in theory on paper and contractually, to pay and purchase access to
13:37
Ejaaz: these things. Like you mentioned earlier, these AI-1 satellites are GPU interchangeable.
13:41
Ejaaz: It's not just built for NVIDIA, although that's the primary kind of case right now.
13:45
Ejaaz: You can fit TPUs, you can fit Amazon Tranium in the future.
13:47
Ejaaz: He's meant to be the tall man, the transporter for all these GPUs to be put
13:51
Ejaaz: out there to kind of absorb all of that sun's energy and train whatever model.
13:55
Ejaaz: He's kind of like agnostic when it comes to that. And I think That is one of
13:58
Ejaaz: the biggest bull cases for SpaceX as I see it today.
14:01
Josh: Yeah, think about how difficult it has been to build a data center.
14:04
Josh: You need the, what is it, land power shell in order to get a data center set up.
14:08
Josh: And it is so remarkably difficult to get the approvals, to get the public sentiment
14:13
Josh: to turn positive, to figure out where to source the energy from.
14:16
Josh: It is this incredibly complicated process that's difficult and no one wants to deal with.
14:19
Josh: And if you could just go over to Elon and say, yo, bro, I got a couple gigawatts
14:23
Josh: that I need in orbit. Can you send them for me in the next few months?
14:26
Josh: That is such an easy outcome and one that i think everyone will want to to use
14:31
Josh: because why wouldn't you and
14:32
Josh: i want to circle back to the ipo and why i said earlier in the show that
14:36
Josh: a spacex should be the most valuable company in the world we we barely even
14:39
Josh: talked about xai and we barely even talked about the other entities that exist
14:42
Josh: underneath it but the idea is that as as ai continues to expand and continues to take up
14:48
Josh: more of what we do on a day-to-day basis it's very much accelerating in that
14:51
Josh: world of bits that we talk about
14:52
Josh: a lot it's able to do ones and zeros very effectively and very efficiently
14:56
Josh: the difficult thing now the bottleneck shifts to the physical world this world
15:00
Josh: of atoms where you have to actually build things you have to create the physical manifestation of.
15:05
Josh: This ai to to allow it to be embodied or to give it enough energy and resources
15:10
Josh: to train itself to become this agi the super intelligence and when it comes
15:13
Josh: to manufacturing when it comes to real world engineering
15:16
Josh: there is really no better team in the world than spacex in terms of hardware
15:20
Josh: engineering and when i look at other engineering teams who are really good at
15:23
Josh: creating physical objects in the real world manufacturing factories which are
15:27
Josh: actually the product and the factories produce these goods and services but
15:30
Josh: building the factory is the hard part
15:31
Josh: i also look at tesla and tesla has done a remarkable job at building and manufacturing
15:35
Josh: these objects at scale and what is tesla now working on they cut down their model s and their x line.
15:40
Josh: In exchange for Optimus, the humanoid robot, that humanoid robot is going to scale very quickly.
15:45
Josh: That humanoid robot paired with Grok and XAI as the orchestrator is going to
15:49
Josh: build this like very autonomous capability of actually physically manufacturing
15:53
Josh: things in the real world.
15:54
Josh: SpaceX and Tesla are probably going to merge together.
15:57
Josh: In the case that those two companies come together, there is no instance in
16:00
Josh: which they do not become the most valuable company in the world so long as they deliver on the promise.
16:05
Josh: And I think the way you look at these companies, the way you validate these
16:07
Josh: companies is you kind of like look at the roadmap and you say,
16:09
Josh: are the products and services that they offer in the future going to be significantly
16:13
Josh: better than they are today?
16:14
Josh: And if the answer is yes, then the idea is that the company should be significantly
16:18
Josh: more valuable to match the added value that they're going to increase.
16:22
Josh: I very strongly feel that way. We have Starship. We have Optimus, the humanoid robots.
16:25
Josh: If just those two things work and nothing else works, the amount of value that
16:29
Josh: will be unlocked is unimaginable. It's really hard to fathom how large these
16:34
Josh: companies can get. And we have this perceived limit of what a market cap of a company can be.
16:38
Josh: NVIDIA is close to $6 trillion or something. But to imagine a company reaching
16:41
Josh: $10, $20 trillion, $30 trillion doesn't seem unreasonable when you think about
16:45
Josh: the scope and scale of a company like Tesla, of a company like SpaceX merged
16:50
Josh: together under the rule of Elon Musk.
16:53
Josh: It's just this unbelievable opportunity that finally the public has access to.
16:56
Josh: And that's why I am particularly excited to invest in the SpaceX IPO.
17:00
Ejaaz: I think for the longest time, the SpaceX IPO has been based on a lot of visionary
17:06
Ejaaz: ideas, kind of belief and this undying faith in Elon.
17:11
Ejaaz: But I do want to touch more on like what the potential numbers might look like
17:16
Ejaaz: to make that multi-trillion dollar valuation that you're talking about,
17:19
Ejaaz: Josh, an actual reality.
17:20
Ejaaz: So there's two kind of perspectives I want to share here.
17:23
Ejaaz: So number one, Hamid over here
17:26
Ejaaz: shared his breakdown of what each of these satellites basically equate to.
17:30
Ejaaz: So if you launch one of these AI-1 satellites out into space right now,
17:34
Ejaaz: today, it's the equivalent of basically half a rack of GPUs that you currently
17:39
Ejaaz: find in a data center here on Earth. So it's around 120 to 150 kilowatts, it's around half a rack.
17:45
Ejaaz: Now, today's data centers, let's say it's worth like one gigawatts worth of compute.
17:50
Ejaaz: That's the equivalent of 4,000 racks at around 250 kilowatts.
17:54
Ejaaz: So you would need to launch 8,000 of these AI-1 satellites into space to be
17:59
Ejaaz: able to equate to one data center, one gigawatts worth of compute.
18:03
Ejaaz: And Elon is targeting that by the end of 2027, which is a lot sooner than most...
18:09
Ejaaz: People would have expected. Now, there's a lot of costs that go in this.
18:13
Ejaaz: The average launch costs, actually, you know, if you're being kind of like medium
18:17
Ejaaz: to kind of like conservative is around $10 billion just for like the launch
18:21
Ejaaz: costs and not the materials at all.
18:22
Ejaaz: Now, Elon's whole idea is that he's going to bring these launch costs down pretty
18:26
Ejaaz: dramatically over the next three years.
18:28
Ejaaz: And at three years in, he's saying that it'll be more cost effective than building
18:33
Ejaaz: a data center here on Earth.
18:35
Ejaaz: And so if we do napkin math with that in mind, right, three years from now,
18:39
Ejaaz: or maybe less than three years from now, a Vera Rubin rack right now,
18:42
Ejaaz: which is like NVIDIA's latest GPUs that they're rolling out,
18:46
Ejaaz: is worth around $8 million.
18:48
Ejaaz: Now, if you couple that with the launch costs at around 200k launch cost per
18:53
Ejaaz: rocket, which is what Elon's aiming for, we are going to hit at around $10 million
18:57
Ejaaz: to put a single Vera Rubin rack out in space.
19:00
Ejaaz: So $12 million on the ground, $10 million in space.
19:04
Ejaaz: You can see a world where the cost economics end up becoming favorable,
19:07
Ejaaz: especially if you're not just using Vero Rubin racks, you might be using cheaper
19:11
Ejaaz: Google TPU racks or Amazon Tranium.
19:13
Ejaaz: So the idea is this is a big bet. Now, if you're purchasing this IPO,
19:17
Ejaaz: you aren't just in it for the near term, you're in it for this next three to
19:21
Ejaaz: honestly 10 year window, where
19:23
Ejaaz: AI data centers in space become a much real thing, much more real thing.
19:27
Josh: Yeah, and this is a roundup. So we're going to wrap this up in a second,
19:29
Josh: get some more news. But the final point that I want to make is you mentioned
19:32
Josh: the cost of reusability for Starship.
19:35
Josh: If you look at the Falcon 9 rocket, there's two stages, and only one of them
19:38
Josh: is actually reusable, so there still is a higher cost.
19:41
Josh: With Starship, 100% of the rocket is reusable, and the actual cost to launch
19:45
Josh: it in the case that 100% is reusable is just the cost of fuel.
19:48
Josh: And the cost of fuel on a Starship rocket is actually cheaper than the cost
19:52
Josh: of fuel on something like a Boeing 747. Because...
19:56
Josh: 78% of it, I think, like almost 80% is just liquid oxygen. Liquid oxygen is
20:00
Josh: like very abundant. We breathe it all the time.
20:02
Josh: It is just a liquefied form of oxygen with the remaining 22% being just liquid methane.
20:08
Josh: And methane is like, it's the stuff you like cook on your grill with.
20:11
Josh: So this isn't very expensive. It's not crazy to think that they will need a
20:16
Josh: lot of money to launch these things. And the idea is that that will bring the
20:18
Josh: cost per kilogram down to sub $100.
20:20
Josh: And that means if you are, it does, if you want to go to space,
20:23
Josh: depending on how many kilograms you are, you just divide that by 100 and you're on your way.
20:27
Ejaaz: It's a cheap ticket weight so i'm meant to lose some weight now
20:31
Josh: So yeah go on a diet so you can get a cheaper ticket to space but that's i think
20:35
Josh: that's kind of like the bulkiest why i'm excited i have to ask you as we wrap
20:37
Josh: up this section before we get into the next one
20:39
Josh: which is going to be open ai building this like gigantic data center is are
20:43
Josh: you going to be participating in the spacex ipo are you excited about it are
20:46
Josh: you going to buy some what are your general thoughts.
20:48
Ejaaz: So if robin hood and my broker allows me to which i've signaled interest.
20:52
Ejaaz: Absolutely, I will be as this goes live today.
20:56
Ejaaz: My thoughts in general are I wouldn't bet against Elon when it comes to hardware
21:02
Ejaaz: scaling. He's less proven on the software side of things.
21:05
Ejaaz: This is just me trying to be balanced here, but he has proven himself time and
21:08
Ejaaz: again on hardware scaling.
21:10
Ejaaz: And SpaceX is solving a hardware issue, one of the biggest hardware issues or
21:15
Ejaaz: challenges ever with AI data centers in space.
21:18
Ejaaz: As everyone knows, I'm a big AI fan. So if he's able to pull this off.
21:21
Ejaaz: More compute equals better models, better inference.
21:24
Ejaaz: He'll be able to serve agnostically any AI company. It's an obvious buy for me.
21:28
Josh: Yeah, I think I'm pretty aligned. Me too. I have been so I got lucky in the
21:32
Josh: case that I had some private shares from about five years ago.
21:34
Josh: So I have been a SpaceX holder. I continue. I plan to continue that going forward.
21:38
Josh: I also requested some at IPO.
21:40
Josh: I'm also just planning to continue to buy like, dude, I can get a 10 plus trillion
21:43
Josh: dollar company for 2 trillion bucks.
21:45
Josh: Sign me up. I think the people who are a little more time sensitive,
21:48
Josh: who are looking to get a return sooner, that might be a bit iffy.
21:51
Josh: I mean, $2.1 trillion is high. There's still a lot of variables that need to be worked out.
21:55
Josh: There's a chance that this thing goes public, shoots up, and then falls for a little while.
21:59
Josh: I don't know. We're not traders. But I think as long-term investors,
22:03
Josh: as someone with a low time preference, SpaceX at $2 trillion, two thumbs up for me.
22:07
Josh: Let's keep the train going. Anyways, we probably should move on.
22:10
Josh: This is a roundup. This is Friday. We got to fill the people in on what's going on.
22:13
Josh: EJ, there's some news from OpenAI about a new gigawatt data center.
22:18
Josh: We're back on Earth, by the way. We're talking about data centers on Earth.
22:22
Josh: What's the deal with this new OpenAI news?
22:24
Ejaaz: Yeah, so we're out of the world of single-digit gigawatts now,
22:29
Ejaaz: Josh. We're in double digits.
22:31
Josh: Run it up expeditiously.
22:32
Ejaaz: OpenAI has basically signed a major deal for a 10-gigawatt data center.
22:38
Ejaaz: For context here, the largest gigawatt data center, which is what Elon's building
22:43
Ejaaz: with Colossus 3, is aiming to be around 1.5 to 2 gigawatts worth of data centers.
22:48
Ejaaz: For a single site to own 10 gigawatts of compute, you need roughly around $500 billion.
22:56
Ejaaz: Now, obviously, OpenAI, they haven't IPO'd yet. They might be pretty soon.
23:01
Ejaaz: They don't have that kind of money. So they're being backstopped by their best
23:04
Ejaaz: friend, SoftBank, and of course, NVIDIA.
23:07
Ejaaz: Jensen Huang is stepping up to backstop this entire data center campus in Ohio.
23:12
Ejaaz: It's going to be an absolute gargantuan beast of a site.
23:16
Ejaaz: And it kind of follows in the trend of
23:19
Ejaaz: Sam Altman and OpenAI specifically, who have been incredibly aggressive about
23:23
Ejaaz: acquiring as much compute as they can to serve their models.
23:27
Ejaaz: Now, if you compare them to Anthropic, you might notice sometimes when you're
23:30
Ejaaz: using clawed models, the performance gets a little degraded,
23:32
Ejaaz: or you might get limited, or you might hit your limits really quickly.
23:35
Ejaaz: That's because Anthropic has fairly been conservative with compute,
23:39
Ejaaz: and that's given them advantages in other areas.
23:41
Ejaaz: OpenAI has been the opposite, and that's why it's always kind of like relatively
23:44
Ejaaz: free and flexible to use as much of your rate limit at OpenAI as you can,
23:48
Ejaaz: which has led to arguably a better developer experience, a better coding experience,
23:51
Ejaaz: a better LLM experience in general.
23:53
Ejaaz: And Sam is doubling down on that bet. He believes in the future,
23:57
Ejaaz: the demand for AI is going to outstrip the amount of available compute supply
24:00
Ejaaz: that he has currently on Earth.
24:02
Ejaaz: He is kind of not on good terms of the alarm, you could say.
24:05
Ejaaz: They recently just got out of a court case.
24:06
Ejaaz: So he's probably not going to be using data centers in space anytime soon.
24:10
Ejaaz: So he's trying to secure as much ground space as he can on this new site.
24:14
Ejaaz: Now, what worries me a little about this is just the pure amount of money that
24:19
Ejaaz: you need. Now, obviously, this is going to be done in phases.
24:21
Ejaaz: 10 gigawatts of computers is going to take absolute years to spin up.
24:24
Ejaaz: But it's an important bit of news that I think kind of lays into this next headline.
24:28
Ejaaz: Josh, I don't know if you saw this, but OpenAI has confidentially filed for
24:32
Ejaaz: an S1. They are going to be IPO-ing relatively soon.
24:36
Josh: OpenAI is copying SpaceX with data centers. They're copying SpaceX with IPOs.
24:41
Josh: It doesn't work with Elon. All right, I'm getting the trend.
24:43
Josh: Yeah, but it seems like this is actually official now.
24:46
Josh: I hope they can build that data center. That would be amazing.
24:48
Josh: But we do have a official filing of a confidential S1.
24:51
Josh: And there was also a leaked all hands news report that I read that said Sam
24:56
Josh: is kind of using this as an option.
24:58
Josh: There is the option to go public soon. But depending on how their model progress
25:02
Josh: goes over the next few months, they will then make that determination.
25:05
Josh: His perspective is that the world will meaningfully change in the sense that
25:08
Josh: they do get kind of like this runaway AGI self-fulfilling loop type thing.
25:13
Josh: And he wants to be a private company in the case that that does happen sooner
25:16
Josh: rather than later. This is some pretty...
25:19
Josh: Big grand visions that they're dealing with so we'll see but now they do have
25:23
Josh: the option they have confidentially filed someone out there is reading through
25:26
Josh: this and probably having a field day understand the internals of how this company
25:29
Josh: actually works so that's pretty exciting news.
25:32
Ejaaz: Yeah and there was another bit of news that uh open ai reportedly is looking
25:35
Ejaaz: to slash the cost or prices to their subscriptions which is kind of a change
25:41
Ejaaz: in the directional trend that we've seen from anthropic um
25:43
Ejaaz: google actually this week cut their prices as well i think we're starting to see,
25:48
Ejaaz: we're starting to reach a point where all these AI models, as good as they are,
25:51
Ejaaz: are starting to become more commoditized.
25:53
Ejaaz: They kind of all do the same things. The leap between the prior version and
25:57
Ejaaz: the current version isn't as large as it used to be, like, say, a year back.
26:01
Ejaaz: And OpenAI is taking the first step to being competitive towards Anthropic and
26:05
Ejaaz: Google. So Google slashed their access to their AI studio to about $20 a month
26:10
Ejaaz: for their cheapest thing. I think OpenAI is planning to do the same.
26:13
Ejaaz: And I'm guessing, like, Anthropic and OpenAI are going to get into some kind of a bidding war.
26:16
Ejaaz: OpenAI kind of like taking its time to go public, as Sam suggested in that leaked
26:20
Ejaaz: memo, is very interesting to me because you see Anthropik and SpaceX kind of
26:24
Ejaaz: like racing to raise as much capital as they can.
26:26
Ejaaz: And if Sam wants to pay for a $500 billion data center, he kind of like,
26:29
Ejaaz: I'm confused where he's going to get the money from. Maybe he's just going to
26:32
Ejaaz: keep borrowing privately.
26:33
Ejaaz: But it's just a unique case where OpenAI is kind of breaking away from the pact.
26:38
Ejaaz: And I'm wondering if these financial decisions actually end up paying off.
26:41
Josh: Yeah, we'll see. We got to cover Fable next because it just released this week.
26:45
Josh: You've been using it for a few days. I've been using it for a little while.
26:48
Josh: It's a pretty amazing model. I've been using it for a lot of fun extracurricular
26:53
Josh: stuff I found. Like I created the Minecraft clone to play around with it.
26:57
Josh: I've been building kind of games in 3D spaces and I found it to be really fun.
27:00
Josh: It genuinely feels like this novel model architecture that I've been enjoying
27:04
Josh: playing with. Ejaz, over the first few days that you've been using this,
27:07
Josh: what are your initial thoughts, opinions, kind of like after the dust settles, how do you feel about,
27:12
Josh: the new model fable 5.
27:14
Ejaaz: Um i think it is less sycophantic than opus 4.8 i'm not gonna lie i
27:19
Ejaaz: when 4.8 came out i use it primarily as my main model but like it was way more agreeable than 4.7
27:26
Ejaaz: we're now back with fable 5 and that might sound like it's got nothing to do
27:30
Ejaaz: with intelligence to me it is it factors in the personality of the model and
27:34
Ejaaz: i'm able to like work with it longer over certain tasks it just understands what i want to do
27:39
Ejaaz: the second thing i've noticed is i've been working on a couple of personal projects
27:42
Ejaaz: that involve a lot of code and research.
27:46
Ejaaz: I can just one-prompt Fable 5, and it is able to just work continuously for up to an hour.
27:53
Ejaaz: Now, I know software engineers listening to this are like, I let my thing run
27:56
Ejaaz: for five to six hours or maybe even overnight.
27:58
Ejaaz: But for me, for the average person, it is becoming incrementally more valuable
28:04
Ejaaz: for a model that understands what you want, that doesn't come to you for every
28:07
Ejaaz: kind of challenge it hits, and just kind of figures it out for himself.
28:11
Ejaaz: And I think this kind of trend of models doing longer horizon tasks way better than the average human
28:16
Ejaaz: is a subtle trend that is going to be increasingly valuable for corporate workers
28:20
Ejaaz: at enterprises, as well as personal people that are like starting their own
28:23
Ejaaz: businesses or want to spin up multiple of these agents instead of hiring an employee.
28:27
Ejaaz: I just think it's a wonderfully intelligent model. And I'm excited to see further
28:30
Ejaaz: iterations of this mythos grade kind of type of model be released.
28:35
Ejaaz: But there was some pushback, right, Josh, like there were some policies that
28:39
Ejaaz: came out that were basically like, hey, like, You get access to this public
28:43
Ejaaz: version of Mythos 5, but it's going to be heavily safeguarded.
28:46
Ejaaz: Now, one of the biggest criticisms that they faced was if you ask it questions
28:50
Ejaaz: like, hey, I'm trying to train my agent to become more efficient and use less
28:54
Ejaaz: tokens, it'll take that as some kind of an inclination that you're trying to
28:58
Ejaaz: learn how Claude Fable 5 is built.
29:00
Ejaaz: And Anthropic installs these anti-distillation maneuvers, basically rediverting
29:06
Ejaaz: you back to the old model, Opus 4.8.
29:08
Ejaaz: But in this case, it would give you degraded quality intentionally because it
29:13
Ejaaz: thinks you're trying to like steal its blueprints.
29:15
Ejaaz: Anthropik has now walked that back because previously they weren't going to
29:18
Ejaaz: tell you when they were going to do it.
29:19
Ejaaz: And now they are telling you. They reversed that decision as of yesterday,
29:23
Ejaaz: basically saying, hey, we'll be up front with you when we're going to intentionally
29:26
Ejaaz: degrade the quality of your answer because we think you're trying to like copy
29:30
Ejaaz: our model weights, which is a very political maneuver.
29:33
Ejaaz: And it's nice to see that they backtracked on that.
29:36
Josh: There you have it. So there's the update on Fable and the final topic of the week.
29:39
Josh: We filmed an entire episode about Fable and also this final topic, WWDC.
29:43
Josh: So if you have not watched either of those, I highly recommend going to check
29:46
Josh: them out. Pretty amazing episodes. We had a really strong week of fun topics
29:49
Josh: to talk about this week, but I do want to trace over WWDC, particularly around
29:53
Josh: Siri AI, which seems to actually be as good as they said it would be.
29:57
Josh: And we have to cover this because that comes as a surprise. This is not the Apple of the past.
30:01
Josh: This is the Apple that promises things and doesn't always deliver so the people
30:04
Josh: have had to download the beta actually fact check try it for themselves and
30:08
Josh: the initial reviews have been spectacular
30:10
Josh: it seems like it actually works better than what they demoed in real time on
30:14
Josh: stage it actually works faster it understands context it works very well on
30:17
Josh: people's phones and this is available through the developer program that apple
30:21
Josh: offers through their ios developer membership
30:24
Josh: it seems like we're actually going to have a pretty powerful siri experience
30:27
Josh: we're going to have on-device ai that is actually.
30:31
Josh: Contextually aware of everything that lives on your phone and off of it.
30:34
Josh: And they actually addressed something that I found really interesting.
30:37
Josh: Maybe you saw, maybe you didn't about their removal of Siri from Europe and
30:42
Josh: from China in the sense that they will not budge on the privacy thing.
30:45
Josh: They will offload some of the compute to cloud in the sense that,
30:49
Josh: or in the case that it can't run locally on your device, it's a little too challenging for that.
30:53
Josh: And when they do, it's done so entirely privately.
30:56
Josh: Europe doesn't want this to happen. So therefore Apple's like,
30:58
Josh: no, no, no, we're not giving you the backdoor. We don't even have the backdoor.
31:01
Josh: We are just not going to push it there. And I really admire the like standing
31:05
Josh: strong on the stance of being private of running everything on device.
31:08
Josh: It seems like Apple has a grand slam and it comes maybe a couple of years later
31:12
Josh: than everyone would like.
31:13
Josh: But this is something I am stoked to get. And I think the downstream implications
31:16
Josh: of this are probably pretty big.
31:18
Josh: There's a lot of casual users of something like ChatGPT who pay $20 a month
31:22
Josh: to have an LLM that they basically use as Google Plus Plus, this now substitutes
31:27
Josh: and subsidizes a lot of that all for free,
31:30
Josh: running local inference on device. So very excited, very bullish about Apple in general.
31:34
Ejaaz: You know, it was funny. I was reading the release article for Siri AI,
31:38
Ejaaz: and typically at the footnote, you'll see, you know, these regions are excluded.
31:43
Ejaaz: It'll just be a minor bullet point in tiny, you know, writing.
31:46
Ejaaz: For Europe, they had an entire A4 page basically saying this is Europe's fault.
31:52
Ejaaz: Europe government is the reason that we can't release our product,
31:56
Ejaaz: and we are not budging because we just think, frankly, that this is ridiculous.
32:01
Ejaaz: It was a larger piece than the China bit, which just came in as like a paragraph.
32:05
Ejaaz: Obviously, China is like going to want access.
32:07
Josh: I think everyone expects the China one.
32:09
Ejaaz: Exactly. But for Europe, they've been sort of anti-AI and anti-privacy for quite a while now.
32:14
Ejaaz: And it has resulted in slowed down releases of the latest AI products and features
32:20
Ejaaz: in that region, which is just kind of like sad to see, especially like,
32:23
Ejaaz: you know, that's where I'm from.
32:24
Ejaaz: And it's annoying that like, I can't talk to my friends about the latest features
32:27
Ejaaz: because they have to wait like months to a year to get access to the exact same thing.
32:30
Ejaaz: That being said, a lot of the haters for Apple, which included me,
32:34
Ejaaz: which was me a year ago, was like, I don't think Apple's gonna be able to catch up.
32:38
Ejaaz: We have been proven absolutely wrong. And I wanna make that very clear on camera right now.
32:42
Ejaaz: Apple has a distinct advantage, which is they own the entire consumer hardware
32:46
Ejaaz: ecosystem, 3.5 billion devices.
32:48
Ejaaz: And now they have the hardware, the chips, to run some of the best frontier
32:52
Ejaaz: models on your phone. They may not be as large as Anthropics.
32:55
Ejaaz: They may not be as large as OpenAIs.
32:57
Ejaaz: But they're trained on your data. And you can do something that none of these
33:00
Ejaaz: other AI labs can do, which is ask Siri to basically find that text that your
33:04
Ejaaz: mom told you to do the thing and you forgot what it was, or,
33:06
Ejaaz: you know, bring back that memory of like when you were hanging out with your
33:09
Ejaaz: best friend, doing blah, blah, blah.
33:10
Ejaaz: It has full context, grounded context on your entire thing.
33:14
Ejaaz: And that is a feature we haven't seen and one that I'm personally super excited to use. So...
33:18
Ejaaz: That is all the news for today, folks. Josh, any final thoughts?
33:22
Josh: You're caught up. No, that's everything. The vibes have been crazy high this
33:26
Josh: week. The news has been insane.
33:27
Josh: Lots of things that I think we're probably personally excited about have been released. We got WWDC.
33:33
Josh: We got a new Frontier AI model. And we got the biggest IPO in history all in one week.
33:38
Ejaaz: It's crazy.
33:39
Josh: It's like there are weeks where- We also got.
33:41
Ejaaz: The Knicks dub yesterday. Please, let's not forget.
33:43
Josh: For the New Yorkers. Ejaz and I are both in New York, both big Knicks fans.
33:47
Josh: Huge. hopefully the day after you're watching this they win the championship
33:51
Josh: that'd be pretty amazing given the vibes this week i suspect that's that's likely
33:55
Josh: that feels probable just on a vibes based opinion
33:57
Josh: um but yeah very excited about all the things that are happening very excited
34:00
Josh: that you have made it here to the end of the episode and are still listening
34:03
Josh: so thank you so much for all the support as always
34:06
Josh: if you enjoy this as always also don't forget to share it with a friend anyone
34:09
Josh: who you think might be interested anyone who asks you should i buy the spacex
34:12
Josh: ipo maybe send in this video and say,
34:15
Josh: not only should you buy it, you should set, I'm not, I can't go that far.
34:18
Josh: Don't make your own. This is not financial advice.
34:21
Ejaaz: Actually, if you disagree with us, if you disagree with us, we actually want
34:25
Ejaaz: to hear from you. Like we have been combing through the numbers.
34:28
Ejaaz: We've been combing through the projections and we're like, this kind of makes a lot of sense to us.
34:31
Ejaaz: But if you're adamant that we're wrong, let us know in the comments.
34:34
Ejaaz: We're hearing, we're going to hear you. We're going to see you.
34:37
Ejaaz: We're going to respond to you. So we want to hear from you, whether you hate
34:40
Ejaaz: our opinion or you love it. Like tell us where we're wrong. Tell us where we're right.
34:44
Josh: Yeah, and I'm also going to disagree with you. So I'm ready to go to war in
34:46
Josh: the comment section. Bring it on, baby.
34:48
Josh: Tell me why SpaceX isn't the most valuable company in the world.
34:51
Josh: Come on. But yeah, that's it. Thank you guys always so much for watching.
34:53
Josh: Have an amazing weekend.
34:55
Josh: Enjoy the weekend. Hopefully, by the time you're watching this,
34:57
Josh: SpaceX is going up only. That'd be great.
34:59
Josh: If not, I'm sure we'll hear about it.
35:02
Ejaaz: Long term. The thesis is long term, Josh.
35:04
Josh: Yes, we're long term. This is a low time preference show, man.
35:07
Josh: I'm here for a long time, not a good time. Or a good time and a long time,
35:10
Josh: I guess. But it doesn't matter the timelines that it happens on.
35:12
Josh: But yes, thank you all so much for watching.
35:14
Josh: Have an amazing weekend and we'll see you next week for another set of episodes on Limitless.
35:18
Ejaaz: See you. See you guys.