Latest / In the Loop AI / AI In Space: Google and SpaceX's Wild Plan To Fix Our Energy Crisis

Transcript

AI In Space: Google and SpaceX's Wild Plan To Fix Our Energy Crisis · In the Loop AI

0:00
Ejaaz: Everyone thinks that the most valuable resource in AI is the GPU.
0:04
Ejaaz: It's why NVIDIA is worth $5 trillion, right? But what if I told you you're wrong?
0:08
Ejaaz: What if I told you that the most valuable resource in AI is actually energy?
0:12
Ejaaz: Without energy, you can't power these things. Without energy,
0:14
Ejaaz: the GPU is just a worthless piece of metal junk.
0:18
Ejaaz: And in fact, a lot of AI labs are facing this challenge where they just don't
0:22
Ejaaz: have enough energy to power the tens of billions of dollars of GPUs that they've spent so far.
0:27
Ejaaz: In fact, Satya Nadella of Microsoft recently said that he has hundreds of millions
0:31
Ejaaz: of dollars worth of NVIDIA GPUs in his data centers collecting dust.
0:36
Ejaaz: Now, there are a bunch of efforts focused on solving this energy crisis.
0:39
Ejaaz: You've got nuclear energy, you've got renewable energy, but one particular solution
0:43
Ejaaz: that's taking up a lot of mind space with Elon Musk and Google is literally out of this world.
0:49
Ejaaz: I've got six words for you. AI
0:50
Ejaaz: data centers in outer space. That is a real sentence that I've just said.
0:56
Ejaaz: It's the concept of putting GPUs in outer space to harness the energy of the
1:00
Ejaaz: sun for AI training and for AI inference.
1:02
Ejaaz: And whilst it might sound like a far-fetched idea, it's not when you consider
1:07
Ejaaz: that the resources needed on Earth to power up the data centers,
1:10
Ejaaz: which by the way, are requiring as much energy as a mid to large city in the US.
1:16
Ejaaz: You might need to look at alternative resources for this, the one being the
1:20
Ejaaz: sun. And the economics kind of work out in theory.
1:23
Ejaaz: You have up to three to five times more energy efficiency if you operate GPUs
1:29
Ejaaz: above the Earth's atmosphere than below it.
1:31
Ejaaz: So in this episode, we're going to dig into all things AI space data centers.
1:35
Ejaaz: Is this real? Is this achievable?
1:36
Ejaaz: Or is this kind of like a myth that we're going to laugh about and have pie
1:39
Ejaaz: on our face six months from now?
1:41
Ejaaz: Josh, you and I have a tumultuous relationship around this topic.
1:46
Ejaaz: 50% of us were optimistic. 50% of us, you, were pessimistic about this.
1:53
Ejaaz: What's your take on all of this?
1:54
Josh: I was the biggest hater in the world. Space was so stupid. Why are we sending
1:58
Josh: stuff to space when we have it all here on Earth?
2:00
Josh: Well, my mind has changed. I have flipped 180 degrees. I'm pro space.
2:04
Josh: I'm ready to send these things to space tomorrow. I want all the data centers
2:07
Josh: in outer space immediately.
2:09
Josh: And when I was considering why, I did a lot of research, which we're going to
2:12
Josh: get into in this episode.
2:13
Josh: But there's this combination of things where there is a power grid straining on the U.S.
2:18
Josh: Where data centers now are one of the main drivers rising electricity demand.
2:21
Josh: We can't fulfill all the electricity demand.
2:24
Josh: There are limits on land, water, permitting, cooling.
2:27
Josh: And I was thinking, I was like, well, if your house is full of all these gaming
2:30
Josh: PCs and the power keeps going out, well, you start to ask yourself,
2:33
Josh: well, where else can we put them? And you're like, on the roof? No.
2:36
Josh: Well, why don't we just send them off into outer space? And thankfully,
2:39
Josh: because of new technologies like SpaceX Starship, we can do this.
2:42
Josh: So there are some benefits to space that I want to go over very quickly before
2:45
Josh: we get into who's talking about this and who's doing this. One of them being...
2:50
Josh: The most important one, unlimited solar energy. A lot of people don't know this,
2:53
Josh: but 99.98% of the entire solar system is the sun.
2:58
Josh: All of the mass of the solar system, it exists in the sun.
3:01
Josh: And the sun is just this huge fusion reactor. And in fact, if you can get access
3:06
Josh: to the sun from outside of our atmosphere, if you leave Earth's atmosphere and
3:10
Josh: you go into just normal space, it's actually eight times more powerful than on Earth.
3:15
Josh: So if you remove the constraint of Earth's atmosphere, immediately the
3:18
Josh: solar panels get access to eight times more energy than they would
3:21
Josh: have in the case of the around space space is also
3:23
Josh: a vacuum which means you can radiate heat directly into this
3:26
Josh: empty space where there's nothing to worry about there's no
3:29
Josh: land there's security and sovereignty because let me tell you it's a lot harder
3:33
Josh: to hack a data center in outer space than it is to do so here on earth so there's
3:37
Josh: a lot of perks we're going to go through but i want to hand it back to you to
3:39
Josh: talk about who's been pro space yes because there's some notable names that
3:45
Josh: are very interested in the space race here yeah.
3:47
Ejaaz: There's a lot of people um talking about this josh um before
3:51
Ejaaz: we get into that i kind of want to like lay out that this
3:54
Ejaaz: is a very real problem and challenge because i'm sure a
3:56
Ejaaz: lot of people listen to this and thinking well like i haven't heard any news
3:59
Ejaaz: about this let me introduce you to some um there are two data centers that got
4:04
Ejaaz: permitted to build these data centers from 2019 by the way josh they're worth
4:09
Ejaaz: hundreds of millions i think like book value is like 500 million dollars between
4:13
Ejaaz: both of them, they're collecting dust.
4:15
Ejaaz: There's no lights on in the data centers because they just can't get the energy
4:19
Ejaaz: supply to power these, right?
4:22
Ejaaz: And I think there's none other than Satya Nadella of Microsoft,
4:25
Ejaaz: which can explain succinctly what exactly this problem is. I'm gonna play a little clip right here.
4:30
Ejaaz: The biggest issue we are now having is not a compute glut, but it's a power.
4:34
Ejaaz: And it's sort of the ability to get the builds done fast enough close to power.
4:39
Ejaaz: So if you can't do that, you may
4:41
Ejaaz: actually have a bunch of chips sitting in inventory that I can't plug in.
4:45
Ejaaz: In fact, that is my problem today, right? It's not a supply issue of chips.
4:50
Ejaaz: It's actually the fact that I don't have warm shells to plug into.
4:54
Ejaaz: And so how some supply chain constraints emerge, tough to predict,
4:58
Ejaaz: because the demand is just going, you know, is tough to predict.
5:02
Ejaaz: So he's not also the only one that's been talking about this,
5:05
Ejaaz: Josh. We've got the likes of SpaceX employees saying by 2045,
5:11
Ejaaz: We're going to need enough compute to the base power of the entire planet to
5:15
Ejaaz: power all kind of like data center efforts.
5:17
Ejaaz: Elon Musk says that there'll probably be electricity shortages in about two years time.
5:21
Ejaaz: Sam Altman says we need to make a breakthrough, whether it's through fusion or something else.
5:25
Ejaaz: Mark Zuckerberg says, listen, we would build a bigger data center,
5:28
Ejaaz: but we just don't have the energy to supply.
5:30
Ejaaz: So this is obviously something that is captivating some of the brightest entrepreneurial
5:34
Ejaaz: minds in the world right now when it comes to scaling AI.
5:40
Ejaaz: And, you know, we see this at the infrastructure level as well.
5:42
Ejaaz: You just heard Satya talk about something called the PowerShell.
5:45
Ejaaz: What he's referring to here is not even just the energy supply lines that you
5:50
Ejaaz: need to feed electricity into these data centers.
5:52
Ejaaz: There's also like the manual labor, Josh.
5:55
Ejaaz: There is the technicians, the specialists that you need to kind of like boost
5:59
Ejaaz: this entire thing. There's so many missing gaps here that we just don't meet it.
6:03
Josh: It's funny. When I was in school, BOCES and like specialty job training was
6:07
Josh: very popular. And that's kind of where you learn to become an electrician or a plumber.
6:11
Josh: And a lot of these people, they didn't accrue a lot of debt.
6:13
Josh: They left school early and they have like well-paying jobs now.
6:16
Josh: And it turns out in a world of post-AI where we have AGI and the robots haven't quite made it here.
6:20
Josh: The people who we need the most are the electricians, are the plumbers,
6:23
Josh: are the people who can do the cooling, the heating, the electricity wiring.
6:26
Josh: Those are the most valuable jobs in the world. and the people are getting paid
6:28
Josh: a tremendous amount of money. Now,
6:31
Josh: Before we get into this next section, I do want to frame a new question that
6:34
Josh: we're going to answer a little bit later, but it's the actual narrative of space
6:38
Josh: because a lot of people are trying to raise a tremendous amount of money,
6:42
Josh: which requires a tremendous amount of optimism from the investors and the public.
6:45
Josh: And you have to ask the question, is this a real narrative or is this kind of
6:51
Josh: an exponential kind of like moonshot guess, moonshot like kind of play that
6:56
Josh: they're targeting to get people interested to raise more money to build their AI?
7:00
Josh: I don't want to answer that yet. We're going to answer that as we go.
7:02
Josh: But by answering it, we're going to start with the first company who's actively
7:06
Josh: doing this, which is a project called StarCloud.
7:09
Josh: Now, StarCloud is the first prototype, as far as I'm aware, EJS,
7:12
Josh: right, that's doing this. Can you tell me a little bit more about it?
7:14
Ejaaz: For sure, yeah. So StarCloud is kind of what triggered this entire debate online, Josh.
7:20
Ejaaz: Prior to StarCloud announcing their prototype of launching a GPU in outer space,
7:25
Ejaaz: everyone was kind of like oblivious to this entire thing.
7:27
Ejaaz: And then Google and Elon Musk chimed in after it. So these are the pioneers.
7:31
Ejaaz: And they kind of pioneered it with
7:32
Ejaaz: this prototype, which should already be launched in November, actually,
7:36
Ejaaz: which is they launched an NVIDIA H100 GPU, which is technically the most powerful
7:40
Ejaaz: computer that's ever existed in outer space, into outer space using Elon Musk's SpaceX shuttle.
7:47
Ejaaz: And they ran a Google open source model called Gemma. So it's currently live and up there running.
7:54
Ejaaz: And the reason why this is such a big deal is it's mainly a testing phase to
7:58
Ejaaz: see if GPUs can survive in outer space. It's running on its own.
8:03
Ejaaz: It's training on its own. It's being iterated to on its own outside in space
8:06
Ejaaz: whilst communicating to their servers on Earth.
8:09
Ejaaz: And what they're really trying to figure out here is, number one,
8:12
Ejaaz: is it feasible or will this GPU just die from radiation? We're going to get
8:15
Ejaaz: into some of the challenges in a second.
8:17
Ejaaz: And number two, does it make economical sense?
8:20
Ejaaz: Like if we can do this with one GPU, can we scale this to hundreds of thousands
8:24
Ejaaz: of GPUs? So they're the real ones, Josh, that started this trend.
8:28
Josh: The initial thing that I found funny about StarCloud was the 16 square kilometer heat diffuser.
8:33
Josh: Oh, yeah. That was kind of where I got caught up on, where five gigawatts of
8:35
Josh: energy means 16 square kilometers of metal that it took to diffuse this.
8:40
Josh: Because again, in space, there is no atmosphere.
8:43
Josh: There is no place to cool it besides shooting the radiation off into deep space.
8:47
Josh: It's not like you can run like cold air over this because there is no air.
8:51
Josh: So I think as I've kind of learned more about the StarCloud project,
8:54
Josh: I've started to realize like, OK, that's that's that's kind of within reason.
8:58
Josh: And that's an early prototype in how they're going to go about actually dissipating heat.
9:02
Josh: So six square miles of solar panels, they fold it into a rocket.
9:06
Josh: They could send it out on a starship. It's like it's pretty cool.
9:09
Josh: It seems like it makes sense.
9:10
Josh: Ejaz, is there anything that I'm missing here that technically doesn't make
9:13
Josh: a whole lot of sense based on these posts?
9:15
Ejaaz: Yeah i mean you nailed a technical challenge number
9:18
Ejaaz: one which is like these deployable structures are not only so large for the
9:23
Ejaaz: amount of compute that it's supposedly meant to be supporting like these things
9:26
Ejaaz: like you said are 16 square kilometers that is huge to fuel do you know how
9:31
Ejaaz: much power it fuels actually josh it's like what's it one gigawatt or is it
9:35
Ejaaz: less than that like maybe 100 kilowatts well
9:37
Josh: It's showing five gigawatts now for the data center yeah.
9:40
Ejaaz: Yeah, okay, right. Four by four. Yeah. So that's technical challenge number one.
9:44
Ejaaz: And then technical challenge number two down here is the radiation that you
9:49
Ejaaz: mentioned, which is basically, you know, you have all these kind of like nuclear
9:52
Ejaaz: and atoms flying around, which
9:53
Ejaaz: can really mess with the composure and structure of these data centers.
9:59
Ejaaz: Radiation is a very real issue in space. That's why you have all these different
10:01
Ejaaz: space suits and insulation.
10:03
Ejaaz: So if you expose hardware to this and it kind of messes up with any kind of
10:07
Ejaaz: the machinery of it's doing its job, then it becomes more and more implausible.
10:10
Ejaaz: But these are things that they were kind of at like the frontier testing in
10:14
Ejaaz: terms of research. We haven't really found a resolution just yet.
10:17
Ejaaz: There are solutions like kind of like cooling that you mentioned,
10:20
Ejaaz: but otherwise there's no clear cut solution.
10:22
Josh: Yeah. Okay. But there is another company that we're all familiar with that does
10:25
Josh: have a seemingly more clear cut solution and a seemingly more clear roadmap
10:29
Josh: which um i believe the ceo goes by the name of elon musk if i'm not mistaken
10:34
Josh: and yeah elon and spacex have a strategy uh we have a clip here that i think
10:38
Josh: you wanted to share maybe we could start there and we'll get into how exactly
10:41
Josh: this is all going to work in.
10:42
Ejaaz: Order in order to harness a non-trivial amount of the energy of the sun you
10:46
Ejaaz: have to move to solar-powered ai satellites in deep space um
10:53
Ejaaz: Somewhat is a confluence of Tesla expertise and SpaceX expertise and XAI on the AI front.
11:02
Josh: So there's three critical pillars that Elon mentioned in that little summary.
11:05
Josh: There's Tesla, SpaceX, and XAI.
11:08
Josh: SpaceX is the single most important pillar in that because without the ability
11:12
Josh: to get mass to orbit, it is impossible to build these data centers in the first place.
11:16
Josh: And the reason why the window is just barely starting to open is because we're
11:20
Josh: able to get the cost per kilogram down low enough to make it economically viable.
11:24
Josh: Because when we started with the space shuttle, Ejaz, it cost $60,000 per kilogram
11:30
Josh: to get something into orbit.
11:32
Josh: Nothing is going to make that viable if it costs $60,000 per kilogram because
11:37
Josh: these data centers are very, very heavy and so are the solar panels,
11:40
Josh: so are the batteries, so all of that.
11:42
Josh: Falcon Heavy, which is SpaceX's newer rocket, got that cost down to $1,500 per kilogram.
11:49
Josh: So we've already seen like a few orders of magnitude decrease and Starship,
11:53
Josh: which we've covered in previous episodes, that is planning to go to Mars and
11:57
Josh: beyond, it's still not working, but they're planning to get it fully reusable by next year.
12:01
Josh: That brings the cost per kilogram down to $10 per kilogram, which is the window
12:06
Josh: that enables this conversation to happen.
12:08
Josh: Because in the absence of Starship, we cannot send mass to orbit at scale.
12:12
Josh: With Starship working, we can actually talk about sending these data centers into space.
12:16
Josh: And that's a hugely powerful unlock that only exists
12:20
Josh: because of spacex which is why i think it's important to pay attention to elon when
12:23
Josh: he talks about this stuff because he's the one who's enabled this to
12:25
Josh: even be a possible conversation um earth gets
12:29
Josh: only one to two billionth of
12:32
Josh: the sun's energy so if you go a million times earth's
12:35
Josh: capacity you must go into space is kind of thinking where in order to scale
12:39
Josh: ai to the size that we want in order for true agi for for ai that scales infinitely
12:44
Josh: with us we need to go past earth's ability to generate energy and we need to
12:49
Josh: go out into the stars and that's kind of the the thesis with elon and the xai thing.
12:53
Ejaaz: Got it. I mean, I was just looking at this other post, Josh,
12:56
Ejaaz: where, you know, StarCloud is kind of laying out the economics for their kind of projections.
13:01
Ejaaz: And they say something pretty similar, which is like, on Earth,
13:04
Ejaaz: it costs near like $200 million to kind of run a 40 megawatt data center for 10 years.
13:10
Ejaaz: Technically, or theoretically, in orbit, it would be worth $8.2 million,
13:14
Ejaaz: which is 20 times cheaper.
13:15
Ejaaz: So what I'm hearing from you and the Elon case is that, well,
13:20
Ejaaz: number one, he's going to be the toll master.
13:22
Ejaaz: He's going to basically own the highway into space.
13:24
Josh: He's the Jensen Huang of space.
13:25
Ejaaz: He's the Jensen Huang of space. That's perfect. We have to trademark that term.
13:30
Ejaaz: He's the Jensen Huang of space.
13:31
Ejaaz: That's crazy good. So he's going to not only own that toll, but the other thing, Josh,
13:37
Ejaaz: which I know we both love the idea of and which he actually commented on in this tweet,
13:42
Ejaaz: he says, simply scaling up Starlink V3 satellites, which are the main satellites
13:47
Ejaaz: that he deploys from his SpaceX launchers will have high speed laser links, which would work.
13:53
Ejaaz: And what he's referring to here is work in terms of like a interconnected network.
13:58
Ejaaz: So these satellites would beam lasers to each other one terabit per second speeds
14:03
Ejaaz: to be able to train or inference AI models up at that level.
14:08
Ejaaz: So imagine like a satellite network kind of in space orbiting Earth that you
14:13
Ejaaz: can use to beam terabits worth of data at super rapid speeds between each of
14:19
Ejaaz: the satellites, train your AI model, and then beam it back down to Earth.
14:23
Ejaaz: Just sounds so nuts at this point, Josh, that I can't quite wrap my head around.
14:28
Ejaaz: And then the third thing that he's talked about, Josh, I don't know if you saw
14:32
Ejaaz: this, he's talking about quantum computing on the dark craters of the moon.
14:36
Ejaaz: Can you, sorry, can you help me wrap my head around this? What the hell is this? What's this concept?
14:41
Josh: Yeah, I'm pulling a number out of thin air. I believe it's right,
14:43
Josh: though, where the craters, there are craters in the moon.
14:46
Josh: So there is no dark side of the moon, but there are craters that never see the sun.
14:49
Josh: Those craters exist at a perpetual negative 200 degrees Celsius.
14:53
Josh: They are literally frozen solid.
14:56
Josh: There is no heat that gets in or out.
14:58
Josh: And if you can place a gigantic data center in one of those craters at negative
15:04
Josh: 200 degrees Celsius or whatever the exact amount is, you have a pretty good cooling system.
15:09
Josh: Like it's not that bad. And each as you'll often see with quantum computers,
15:13
Josh: there's this huge mechanism on top of them. And then there's this tiny little chip at the bottom.
15:18
Josh: The entire mechanism on top of it is just to keep it cool.
15:22
Josh: The actual chip is is the size of a normal chip but because
15:25
Josh: it's so important and so necessary to get
15:28
Josh: quantum computing chips at solid zero so that the qubits don't move around too
15:32
Josh: much that's why they need it so there is a world in which quantum computing
15:36
Josh: makes sense if we can get it into outer space there is some data and energy
15:40
Josh: notes that elon has that i wrote down that i want to share briefly because this
15:44
Josh: is also fascinating and he said,
15:46
Josh: Starship should be able to deliver around 300 gigawatts per year of solar-powered
15:50
Josh: AI satellites to orbit, maybe 500 gigawatts.
15:54
Josh: So if this is true, within a few years of full-rate Starship launches,
15:59
Josh: orbital AI could rival or exceed entire country's electrical consumption.
16:06
Josh: So the scale of this is outrageous. And granted, this is all very optimistic.
16:09
Josh: This is all very forward-looking, very sci-fi-esque.
16:13
Josh: But in the case that is even directionally correct the
16:16
Josh: next 10 to 20 years of these launches the next 10 to
16:19
Josh: 20 years of compute can scale in
16:21
Josh: a really exciting way and we were talking to to luke who helps produce the show
16:25
Josh: right before this and he made a great point that said in order to get exponential
16:29
Josh: gains in progress you need exponential gains in technology or something along
16:33
Josh: those lines and it's so true where if we really want to build ai that changes
16:37
Josh: the trajectory of the universe you need to do something novel, something new.
16:41
Josh: And doing this all in space and unlocking this new energy source feels like
16:44
Josh: the natural progression given
16:45
Josh: where we are in terms of technological capability back here on Earth.
16:48
Ejaaz: It's also equally fair to say that Elon sometimes gets ahead of himself.
16:52
Ejaaz: Like he can see the vision, right? But his projections are way too near term
16:57
Ejaaz: to what it actually requires to get to that point.
16:59
Ejaaz: And I wish I could critique him, Josh, except that there's also another person,
17:03
Ejaaz: which is bull posting about AI data centers in space.
17:08
Ejaaz: And it is Sundar from Google, the CEO of Google, who in pretty much every single
17:13
Ejaaz: interview, this clip that I'm showing you on the screen right now is from an
17:16
Ejaaz: interview he did literally yesterday,
17:17
Ejaaz: which has got nothing to do with Google's AI space efforts specifically,
17:21
Ejaaz: but he had to shell Google's project called Project Suncatcher,
17:25
Ejaaz: where they are also aiming to put Google's TPUs, which is their GPU equivalent,
17:30
Ejaaz: into space so that they can harness the power of the sun.
17:33
Ejaaz: There are too many important people talking about these things,
17:36
Ejaaz: Josh, where like my inner, kind of like half my brain is kind of fighting against
17:40
Ejaaz: this, saying this is absolutely ridiculous.
17:42
Ejaaz: But the other half, I think, is overwhelmingly now in favor of this thing might actually work.
17:48
Josh: Yeah, it's their moonshot. It's another Google moonshot.
17:51
Josh: And again, I'm going to tease the question, is this just a narrative thing to
17:55
Josh: get people excited to invest more in their companies?
17:57
Josh: Or is this real technology? Maybe we'll see.
18:01
Josh: He is partnering with Planet, a company named Planet, and their plan is to launch
18:05
Josh: two prototype satellites by
18:06
Josh: early 2027 to test TPUs and optical links to orbit and see how that works.
18:10
Josh: In the previous example with SpaceX, you mentioned that there is these like
18:13
Josh: space lasers that can talk to each other.
18:15
Josh: A cool thing is that you can communicate at the
18:18
Josh: speed of light there is no kind of interference in space
18:21
Josh: so the time that it takes for a photon to get from one laser
18:24
Josh: to or one satellite to the next is the latency between the two which is very
18:28
Josh: very fast so there is a world in which theoretically if you're looking at this
18:32
Josh: from first principles based on there are no laws of physics that are being violated
18:36
Josh: in order to do this so therefore it's a matter of technical capability technical called prowess,
18:41
Josh: and our intellectual ability to actually figure out and solve these problems.
18:45
Josh: It's funny seeing Google and Elon kind of converge on this. I assume,
18:50
Josh: like you mentioned earlier, a lot of other companies are not far behind.
18:53
Josh: They just don't quite have the infrastructure or the ability to tackle a problem
18:57
Josh: like this. But it seems very real.
18:58
Josh: And they're planning to do this as soon as the next 24 months with the new TPEs
19:03
Josh: that we've been talking about so often. Sundar.
19:05
Ejaaz: Said that they're already prototyping a bunch of radiation measures on ground
19:10
Ejaaz: here at earth to make sure that their tpus can survive so the next natural step
19:13
Ejaaz: there is to launch these babies into space probably using elon musk's spacex
19:17
Ejaaz: because he is the thomas he's the jensen and
19:19
Josh: To that point actually the uh the economics and the launch cost they actually
19:22
Josh: we have a quote here that says uh their modeling says if launch costs get under
19:25
Josh: 200 per kilogram to orbit by the mid-2030s then space-based compute could be
19:30
Josh: comparable to earth data centers on a per kilowatt per year basis so
19:34
Josh: If we can get 20 times more than what Elon's projecting.
19:38
Josh: So if the cost is pretty high. So if it's 20 times more expensive than what
19:41
Josh: Elon's projecting in a decade from now, then we'll have parity with Earth.
19:45
Josh: So it gives you a testament to like the timelines of this. And maybe this is
19:49
Josh: a good time to go into the pros and cons of what.
19:51
Ejaaz: Is good about this,
19:53
Josh: What is bad about this. And maybe humble people a little bit,
19:56
Josh: Ejaz, if you want to start with the cons as to why this isn't as rosy as we
20:00
Josh: think it might be. like maybe maybe bring us back down to earth in fact.
20:04
Ejaaz: I i that's that's where i did did not expect that part josh um okay right so
20:10
Ejaaz: i'm going to start off with the bear case here which you've mentioned a few
20:13
Ejaaz: times which is this is just a narrative play by some of the biggest companies
20:18
Ejaaz: to boost their valuations and let me explain the problem to you right
20:23
Ejaaz: Up until today, speaking about these data centers in space, the focus has been on GPUs.
20:29
Ejaaz: GPUs is the new gold, right? It's what you need to kind of train your AI.
20:33
Ejaaz: And then we found out that they don't have enough energy to kind of like fuel these things.
20:37
Ejaaz: Now, the reason why this is so important, as this post explains,
20:40
Ejaaz: is a lot of these companies, Sundar's company, Google, Elon Musk's companies,
20:45
Ejaaz: are valued based on the amount of GPUs they've been purchasing.
20:49
Ejaaz: And the Wall Street analysts have basically assumed, okay, the amount of GPUs
20:53
Ejaaz: that this company buys is equal to the economic output. I can do a simple formula,
20:59
Ejaaz: right? If they're buying these GPUs, they're going live.
21:01
Ejaaz: What they haven't factored in is that these GPUs aren't going live.
21:05
Ejaaz: They're just sitting collecting dust.
21:06
Ejaaz: And so that's less ROI per quarter than they initially projected,
21:10
Ejaaz: which means that the valuations of these companies should probably be lowered.
21:13
Ejaaz: So what's the countermeasure to this, Josh?
21:16
Ejaaz: Well, obviously it's infinite energy. obviously it's what was
21:19
Ejaaz: that 99.98% of the solar system is the
21:22
Ejaaz: sun it's all that energy so if you can miraculously wow
21:26
Ejaaz: now figure out a way to harness that energy you can
21:29
Ejaaz: now maintain that valuation dare I say boost the
21:32
Ejaaz: valuations right share prices go up after you you hear Sundar talking about
21:36
Ejaaz: this uh google's on an absolute run so that's my skeptics take I think that
21:40
Ejaaz: they might be kind of like feeling this narrative which they know is a decade
21:43
Ejaaz: out from any kind of real uh prototype being out there that scales um so that's
21:48
Ejaaz: my back is do you agree do you disagree with this yeah
21:51
Josh: I i think setting expectations on timing is probably
21:54
Josh: going to be fairly important because again like satya's saying mid 2030s i mean
21:59
Josh: elon's probably saying tomorrow but the reality is it's probably mid 2030s it's
22:03
Josh: going to take a while to get all this stuff to orbit and do it at scale that's
22:06
Josh: comparable to the united states now when i think about that i think about what
22:11
Josh: does the world of ai look like,
22:13
Josh: In the mid 2030s. Like, what are we? What's it going to look like here?
22:18
Josh: Like, we are moving so fast.
22:20
Josh: And to project 10 years into the future, my God, I don't know what the world looks like.
22:25
Ejaaz: So our energy bill is going to be 5x, Josh.
22:28
Josh: Or maybe we have people like Isaiah Taylor, who was on the show last week,
22:32
Josh: solving nuclear reactors right here on home.
22:34
Josh: And we actually have energy abundance. That's enough for our midterm goals prior
22:40
Josh: to extending out to the long term goals. Maybe this is a fun time to talk briefly
22:44
Josh: about the Kardashev scale, which is a sci-fi thing, and there's three tiers to it.
22:48
Josh: The first one being you can harness the entire energy of your planet,
22:51
Josh: and then it's the energy of your sun, and then the energy of your solar system.
22:56
Josh: We have been able to collect perhaps a very small fraction of Kardashev level
23:01
Josh: one, which is of the planet, the energy that hits our planet.
23:04
Josh: So there's a very long way to go, but I think this is kind of like the grand
23:07
Josh: sci-fi vision that gets people excited, gets people motivated.
23:10
Josh: If I'm going to work at a company like SpaceX, I'm stoked at the idea of building
23:15
Josh: super intelligence in outer space. That's a cool mission to have.
23:18
Josh: And like so many of the other missions, it's not physics constrained.
23:21
Josh: It is possible. You can do this thing if you can figure out how to create the
23:25
Josh: technology to do it. So I love the ambition.
23:28
Josh: Is it a narrative play to get more money?
23:32
Josh: Maybe, probably. But in the case of Google and Tesla and SpaceX,
23:37
Josh: it doesn't seem that way. hey, if this was coming from someone like OpenAI or
23:40
Josh: Perplexity or Anthropic, I would feel a little bit differently about it.
23:44
Josh: But Google's always had a moonshot division since the beginning of time.
23:47
Josh: I mean, that's the reason why AI exists in the current state it does today is
23:50
Josh: because they spend so many resources on research and development.
23:53
Josh: If you remember from our episode yesterday, Ejaz, they developed the first neural network 15 years ago.
23:58
Josh: They did. So it's been under development for a long time. And this very much
24:01
Josh: feels like a similar project where these timelines are going to be 10 to 15
24:05
Josh: years. But the research starts now.
24:06
Josh: And it's very important research because directionally it feels like this is correct.
24:10
Josh: Directionally, we're going to need more energy than we can have on Earth,
24:13
Josh: even if we solve nuclear power, because of the hope that it's going to require
24:17
Josh: so much energy to power superintelligence that will be smarter than the smartest thing on Earth.
24:22
Josh: And it just it's a fun science experiment maybe it's a narrative violation because
24:26
Josh: they're getting a little ahead of the gun but you know it could it could be and i like this future.
24:31
Ejaaz: So do i i think my optimistic take on
24:34
Ejaaz: this is all of these things um nuclear fission nuclear power coming into like
24:39
Ejaaz: a scalable means renewable energy and you know harnessing energy from the sun
24:44
Ejaaz: all comes collectively at a time in the mid 2030s where our energy bills are
24:50
Ejaaz: going through the roof and we cannot afford anything else.
24:54
Ejaaz: We need to rely on another energy source. And I think it's all going to neatly
24:57
Ejaaz: come at the right time, Josh.
24:58
Ejaaz: I don't know what makes me think that this might actually happen,
25:01
Ejaaz: but I think we are well on our way to do that.
25:03
Ejaaz: But to your point, there's a bunch of challenges, right?
25:06
Ejaaz: So we've mentioned kind of like some engineering challenges around protecting
25:09
Ejaaz: against radiation and building and deploying these infrastructures out in space.
25:12
Ejaaz: Right now it's too expensive, right?
25:13
Ejaaz: But the idea is presumably using SpaceX's, I'll call it the tunnel to space
25:19
Ejaaz: or the highway to space, it'll eventually reduce the cost down by,
25:23
Ejaaz: you know, maximum weight. So we're able to kind of achieve this.
25:26
Ejaaz: The other thing is kind of like maintenance, right? So it's like a lot of these
25:29
Ejaaz: things require a lot of attention.
25:31
Ejaaz: Like we mentioned earlier that, you know, supplying energy is just one thing.
25:34
Ejaaz: You need to have technicians and experts on there that can like kind of connect
25:38
Ejaaz: the right plugs between satellites and make sure that they speak to each other
25:41
Ejaaz: or GPUs between each other.
25:43
Ejaaz: Are we going to have astronauts roaming around these data centers on space?
25:45
Ejaaz: That is kind of harder to wrap my head around because it just seems infeasible.
25:49
Ejaaz: We don't do that right now. why would it be in 10 years time?
25:51
Ejaaz: Maybe it will be. I'm being optimistic. I have to maintain optimism here, Josh.
25:54
Josh: And hey, maybe robots. Robots in 10 years, probably gonna be pretty good.
25:57
Ejaaz: Oh, Josh, I need to buy more Tesla. I need to buy more Tesla.
26:01
Ejaaz: You've just, the humanoids.
26:02
Ejaaz: Oh God, we are getting way too bullish on the con section.
26:06
Ejaaz: Okay, hang on, wait, hang on. We got to remain tethered to Earth.
26:09
Ejaaz: Speaking of tethering to Earth, actually, how do we get this data and compute
26:14
Ejaaz: that we generate in space down back to Earth?
26:17
Ejaaz: Presumably the majority of the humans are still going to be on Earth,
26:19
Ejaaz: right, Josh? So a question that's in my mind is like, what is the receiving
26:23
Ejaaz: infrastructure of this look like? And is that economically feasible?
26:27
Ejaaz: Elon seems to claim so in that thread that we showed earlier.
26:30
Ejaaz: And Sundar has said multiple times that, yeah, we can just beam data down.
26:33
Ejaaz: We can just beam some of the learnings, the compute down.
26:35
Ejaaz: I don't know if anyone's actually proven that. And it's because we need to do
26:39
Ejaaz: step one first, which is launch the data center.
26:41
Ejaaz: So there's so much that this relies on that needs to be kind of solved via frontier
26:46
Ejaaz: research that we're kind of
26:47
Ejaaz: hoping for, but we haven't got any definitive proof, if that makes sense.
26:51
Ejaaz: So if you're investing in this stuff, it's definitely a longer horizon type thing.
26:56
Josh: Yeah, conceptually, it all makes sense. We're learning a lot from Starlink,
26:59
Josh: actually, in terms of latency and sending things back to Earth,
27:02
Josh: where if you're sending data that's timely, it's a little difficult because
27:06
Josh: there are tens to hundreds of milliseconds that are added in the latency between the two.
27:09
Josh: But if you're batch training, if you're just training mega amounts of AI and
27:13
Josh: then you're beaming them in batches down to Earth, there's enough bandwidth,
27:16
Josh: if you disregard latency, to send quite a bit of data pretty quickly.
27:20
Josh: So I suspect there's some healthy middle ground there where they'll talk to
27:23
Josh: each other very fast and they'll beam data down kind of slow.
27:25
Josh: And that will be okay um there are pros there's lots of pros we talked about
27:30
Josh: a few energy abundance being one cooling being one where you can just create
27:33
Josh: mega radiators you could send things on the dark side of the moon put them in
27:36
Josh: craters where it's really cold um the environmental thing there's you don't
27:40
Josh: need permits or you don't use water or you don't need energy we're just,
27:45
Josh: unobstructed by ai which is kind of an interesting thing because currently ai
27:49
Josh: is very intrusive um it uses a tremendous amount of data and resources and energy and,
27:55
Josh: human labor, which maybe is a good thing to some extent, but that's another thing.
28:00
Josh: And then we have just like the sovereignty and resilience point where when stuff
28:03
Josh: is in space, it is pretty anti-fragile. No one's really going up there to mess
28:06
Josh: with it. It is very secure.
28:08
Josh: You kind of know what you're getting when it's up there. You set it,
28:10
Josh: you forget it. It's a nice thing. So that's kind of how I think we would weigh the pros and cons.
28:14
Ejaaz: Okay. So if those are the pros and those are the cons, the next natural question for me, Josh,
28:19
Ejaaz: is like, if I was an analyst looking at this over the next say like five to
28:22
Ejaaz: ten years yeah what kind of milestones should i be looking for like do you have
28:26
Ejaaz: any idea of what or like what's like what's sensor i guess is it like an entire
28:31
Ejaaz: data center out there or is it a couple of gpus what is it yeah
28:34
Josh: So because there are a few companies that are actively working on this problem
28:37
Josh: we can just kind of evaluate their progress as i think a way of tracking things
28:41
Josh: so we could set a few milestones starting with the suncatcher prototypes who
28:46
Josh: are planned to go up in 2027.
28:48
Josh: So if you're trying to map out trajectories, we have Project Suncatcher.
28:51
Josh: We could ask the question, did the 2027 Suncatcher prototypes launch on time?
28:57
Josh: Did they actually even make it to outer space? Is Starship working?
29:00
Josh: Are they able to carry the correct payloads? Is the payload even able to get into outer space?
29:05
Josh: Then once it's in outer space, we can ask the question, well,
29:08
Josh: does it actually work? Is it able to shield itself from radiation?
29:11
Josh: Is it able to power itself on? Is it able to cool itself down?
29:14
Josh: Do they show working optical links and TPUs in orbit? Can they actually make that work?
29:19
Josh: So we have the prototypes of Suncatcher. Then we want to look at Starship cadence.
29:23
Josh: Is Starship actually getting the cost to orbit low enough to make this economically
29:27
Josh: viable? That, I think, actually is the single most important thing to look out
29:29
Josh: for, is the Starship program.
29:31
Josh: If you are not watching every Starship launch with us on this show,
29:34
Josh: you are missing out, because we're going to cover every single one.
29:37
Josh: It's the most important thing to the progress of this project.
29:39
Josh: And we are going to closely monitor how the cost per kilogram to orbit is going.
29:44
Josh: As long as that number is going down, as long as Starship's
29:47
Josh: start blowing up less, that's going to be a huge win. And that's going to make it at least plausible.
29:52
Josh: And then the third thing I guess to monitor is the power crisis on Earth.
29:55
Josh: We want to see how much power we're able to generate outside of these moonshot ideas.
30:00
Josh: So are people like Isaiah Taylor of Valor able to create modular nuclear reactors
30:04
Josh: to power gigawatts of these data centers and eventually terawatts?
30:07
Josh: Are we able to use natural gas turbines and other forms of renewable energy
30:10
Josh: to not only power data centers, but our day to day lives?
30:13
Josh: Can we do so without making the electricity bills run too high.
30:16
Josh: So I think the convergence of these three things are things to look out for
30:20
Josh: that we can monitor, that will probably give us a better idea of the roadmap,
30:24
Josh: how we're going in the trajectory.
30:25
Josh: And again, 10 years from now, who knows what the world looks like.
30:28
Josh: But that's what I'm going to be looking out for, at least.
30:30
Ejaaz: Yeah, I think that last point is actually one of the most important things,
30:34
Ejaaz: because maybe this is a hot take.
30:35
Ejaaz: But I think the AI data centers in space hype dies down if things like nuclear
30:42
Ejaaz: and renewables can scale on Earth. because people aren't going to focus on things
30:47
Ejaaz: in outer space if they can just kind of fix the thing on problem on Earth.
30:50
Ejaaz: I think this is more of a philosophical semantic thing due to human nature versus
30:55
Ejaaz: kind of doing something in the sky where they haven't spent that much time out there, right?
30:58
Ejaaz: And then the other thing is like, to your point, we've got to reduce the cost
31:01
Ejaaz: of space travel, which I fully believe Elon's going to do, but he has to prove
31:05
Ejaaz: it before people get really bullish and confident that this can be a thing.
31:09
Ejaaz: But that's it. Those are the pros and cons. And that is the thesis and the pick,
31:13
Ejaaz: the for and against for AI data centers in outer space
31:18
Ejaaz: It's so funny seeing our evolution and reaction to this, Josh,
31:22
Ejaaz: given that six weeks ago, we were absolutely laughing at this when StarCloud
31:26
Ejaaz: announced that they were launching an H100 GPU in space.
31:29
Ejaaz: And now here we are opining about, you know, Elon being the toll master,
31:34
Ejaaz: it being cheap enough to launch data centers in space, and then figuring out
31:38
Ejaaz: how to beam down terabits of data to Earth.
31:40
Ejaaz: Just an insane thing that I can't quite wrap my head around.
31:43
Ejaaz: But it's been a super fun episode.
31:45
Ejaaz: And just like the trajectory of putting data centers in outer orbit,
31:51
Ejaaz: there's something else that is reaching the levels and that is exiting the stratosphere.
31:55
Ejaaz: Josh, it is the limitless follow-up page. It is our social pages. It is our YouTube pages.
32:01
Ejaaz: We have gained, I think, what was it? 3,000 subscribers over the last couple of... More? Tell me.
32:08
Josh: Over 4K. And that's not enough because 80% of those people are still not subscribed
32:13
Josh: who are listening to the episode.
32:15
Josh: So if you're listening on YouTube, please make sure not only click the subscribe
32:18
Josh: button, but there's a little bell next to it, which means you could turn on
32:21
Josh: notifications to get notified whenever a new episode comes out.
32:24
Josh: Happens three times a week. They're all pretty good. They normally range from 25 to 35 minutes.
32:27
Josh: And if you enjoy them with your friends, you should share them along because
32:30
Josh: it really makes a big difference to the growth of this podcast.
32:33
Josh: And we're trying to grow.
32:34
Josh: We're trying to make this the biggest thing ever.
32:36
Josh: We've been stalling out a little on the audio front. So if you do prefer to
32:40
Josh: listen to it as a podcast,
32:41
Josh: perhaps go find it on your favorite podcast player, like apple podcast or pocket
32:45
Josh: cast is my preferred one those are good places to to find out and hang out with
32:49
Josh: us but yeah i think that's mostly it jazz do you have any parting words because
32:52
Josh: i actually i have one parting story so you go first.
32:56
Ejaaz: Okay. One final thing is we also have a newsletter and we have about 70,000
33:01
Ejaaz: of you tuning in every single week.
33:03
Ejaaz: Either Josh and I write an essay or a thesis on what we think the future of AI is going to look like.
33:08
Ejaaz: Our last one is a juicy bull case on Anthropik. You'll go check it out.
33:11
Ejaaz: And we give you the five highlights of the week. So we are anywhere and everywhere
33:14
Ejaaz: that you could possibly ingest information about AI and frontier tech.
33:18
Ejaaz: Follow, subscribe, give us your email list. Josh, what is your parting words?
33:22
Josh: Yeah, there's actually a lot of alpha in there. Google, Ejaz,
33:25
Josh: you wrote the Google Bowl case like four or five weeks ago, and the price has
33:28
Josh: gone off like 20% since then.
33:29
Josh: So if you're not listening to or reading the newsletter, I would advise.
33:33
Josh: I wanted to leave with one little bit of sci-fi because when I do episodes like
33:37
Josh: this, I'm always reminded...
33:38
Josh: Of sci-fi. Like when I look at the future, I'm always reminded of sci-fi books
33:41
Josh: and a lot of them get it wrong.
33:42
Josh: But I think you could almost always find a sci-fi story that will map to the
33:46
Josh: reality that we're having currently. And you kind of see this with Black Mirror.
33:49
Josh: One of them that I would encourage everyone to read, it's called The Last Question by Isaac Asimov.
33:52
Josh: And it's where I see this going as we move AI into space, where he kind of like
33:56
Josh: tracks the progress of humanity over millions of years.
33:59
Josh: And each time there is a smarter and smarter and smarter AI.
34:02
Josh: And it is an all knowing form of intelligence. And every time it gets smarter,
34:06
Josh: because they've been able to capture more and more energy and eventually it leads
34:08
Josh: to the last question which is the last question humanity will
34:11
Josh: ever need to answer um which i will leave that for you guys
34:14
Josh: to explore and to uncover for yourself but it
34:18
Josh: is like it's so exciting that we get to live
34:20
Josh: in this world that feels like a sci-fi novel where
34:23
Josh: we're actually talking about sending ai into outer
34:27
Josh: space building artificial general intelligence doing so using space
34:30
Josh: rockets and satellites it's like it's really cool and it's
34:33
Josh: not here today but there is a very clear trajectory to getting there in the
34:36
Josh: future so i think that's mostly what we are excited
34:39
Josh: about covering and i think the people who have been here with us
34:42
Josh: long enough understand that and they're excited to be on the journey with us so for
34:45
Josh: all of you who have made it this far through our fairly long episode today thank
34:48
Josh: you i hope you enjoyed the space race um our agi in space episode i hope you
34:52
Josh: left informed or excited optimistic or just downright frustrated that we are
34:57
Josh: missing the mark so much and ai in space is stupid but whichever one it is let
35:01
Josh: us know in the comments down below and we will see you guys in the next one thank you so much for.