Episode 13: AI's $500B Financing Boom, Satellite Competition & Sunlight From Space

Peter Jarich (00:00)
Hi everyone, and welcome to Intelligence and Noise, the GSMA Intelligence podcast, where we look at the week's news. We try to sift through the noise, dig out some intelligence, tell you about things that you're probably already talking about and why we think they're important, and maybe flag a few things that you didn't pay attention to but probably should have.

I mentioned this a few weeks ago. We often do this in the context of the GSMA CEO newsletter. Congrats if you're on that distro. It means you're important. But the CEO newsletter is on holiday, so while the CEO newsletter is on holiday, Tim and I are unleashed, which is my cue to say, "Hey, Tim, how are you doing?"

Tim Hatt (00:42)
Peter, how's it going?

Peter Jarich (00:43)
I am doing well.

For those of you tuned in last week, you'll recognize that was the first, maybe last, time where I actually tell you what Tim does, other than just a random who shows up and talks with me. It's on his business card. Ask for it if you see him in person.

I also let everyone know that we've moved, last week and this week, to what we're doing now, which is some shorter versions, just to see, mix it up, see if folks like it. Let Tim and I start doing some packing for our own holidays, which are coming up soon.

So we're gonna get into it in a second, but, as is my want, I'm gonna kick off with a bit of a personal anecdote that hopefully leads into the news.

So, a couple of days ago, I went into work and got my bike tuned up. So I brought my bike with me, went and picked it up. My wife was very happy because it meant she didn't have to pick me up from the train. I could bike home.

And she reminded me, on my bike home, I'd be biking home during the eclipse. So I think her main worry was, "Watch out for other folks. Watch out for other folks that are looking at the eclipse while they're supposed to be driving," and be careful to not get hit by eclipse watchers.

What she didn't tell me — and she probably should have, but she probably assumed, as a grown person and not a child, that as I was biking, I probably shouldn't be staring into the sun.

Because that's bad.

But how can you not?

Right. So I was a little bit blinded as I biked. Not too, too bad. But I got to a point right before I got to my house — wonderful place, with some nice tree cover. And you'll see these pictures every now and then, but an eclipse coming through tree cover, you get some really beautiful effects on the ground.

I was thinking, "This is really great," but what I could really use is — I think about this whole, you know, we've got weird things with the sun going on and being covered up and all these effects — what I could really use is, this would be nice if it looked this way at night.

And I only wish there was a solution that, if in the middle of the night I could just flip a switch and have this sort of full-sunlight solution.

And for those of you who think that I've started my holiday drinking a little bit early, hold on to the end of the podcast, because we're gonna get to just that sort of — it's almost like someone heard me say, "I would love for it to be bright in the middle of the night," and we have a solution.

What's not gonna get us there, though, Tim? Massive data centers, left, right and center, every corner. What is your news? What were you paying attention to in the news today?

Tim Hatt (03:31)
Well, there's actually a lot on for August, but the story I wanna highlight is the NVIDIA financing deal with a load of Wall Street banks and asset managers to fund and finance a lot of its compute and its partners' compute aspirations.

So the story here is that NVIDIA has announced a deal this week for a total sum of around $500 billion with a suite of banks and asset managers: Blackstone, Apollo, BlackRock, Goldman, Brookfield, KKR — so kind of a who's who of Wall Street finance.

And the deal is there to create a financing pool for NVIDIA partners to be able to lease GPUs and therefore compute capacity over time, which NVIDIA partially underwrites to the total of 25% of the GPU value over the term of the lease, but ultimately is there to allow capital access for unleashing the AI era.

And I think, Peter, what's interesting — I want to get your reaction on this — but I think there's sort of two points of interest.

One is we're losing perspective on how big these numbers are. Every week we see new hundreds of billions, you know, wands, whatever, of AI financing. I've lost track of how big we're talking.

But two, it goes to this point of: do GPUs become a new asset class, as Jensen Huang is talking about, and, you know, the Wall Street partners?

Interesting question there.

But three, I think there's still a lot of open questions and some risks. You know, we talk about the circular financing, there's some other ones, but that's my starter for ten.

What do you think?

Peter Jarich (05:20)
Yeah, I think it is — I think for me the key part that you hit is that first part about just losing track.

And, you know, I don't think it's just a function of being lazy and not having a tracker, right? Not having a big whiteboard where you write down how much money is being committed this week.

But you look at these numbers and it's not just losing track to say that, "Wow, there's a whole lot going on," as much as it is, what is the limit? What is the extent to which people can actually commit these amounts of money?

What happens if something goes wrong?

And by "goes wrong," I don't mean, you know, something breaks in the industry. I mean, like, we recognize that as we move to the next generation of GPUs, the older ones have become a lot less valuable, and thus the value of those investments goes south, right?

Or people recognize that they're sitting on a whole bunch of devalued assets.

I think it gets a little bit scary when you're talking about these sizes, the size of these numbers, but then also it gives it a bit of a different commitment to finance, which is the case here, versus a commitment to actually deploy.

Right? Because I think that's the bigger issue for me, at least.

I think a lot of the folks, when the reading I'm seeing is, it's one thing to say, "Let's have a commitment to finance this, and we're gonna go and help partners," versus, "I'm gonna commit to deploy $100 billion worth of data centers," and then maybe I do, maybe I don't, right?

Or maybe I actually — I think there's less of a risk here where it's not a commitment to actually deploy that as much as it is a commitment to raise that to support people who deploy it.

Because if, in the end, that $500 billion doesn't all get used, we're not going to get too, too worried.

Tim Hatt (07:18)
No, exactly. But you bring up an interesting point, and it's when you look at this, implicit in what you're saying, there's a question that's been asked again and again: Is AI due for some type of bubble or correction, or call it what you want?

Peter Jarich (07:32)
A reset.

Tim Hatt (07:33)
A reset.

But I almost think that's not the question. When we look at this, there's not very little doubt that AI will be the foundational compute layer for everyone and everything for the foreseeable future. We get that.

The question is how big that is.

And I think these financing deals are all rushing in to lay the infrastructure for what is projected to happen five, ten, fifteen, twenty years down the road.

But what if it doesn't happen to that extent? Or what if, for example, compute prices change because there isn't a long-term store of value?

That's an open question.

Peter Jarich (08:10)
Yeah, I think that's — all of these bets are being predicated on something.

I'm a little bit less worried. Like, I'm a little bit less worried when it's a commitment to self-finance things versus saying, "Hey, we're going to deploy this," or, "I'm going to buy X amount from you," which, if it doesn't happen, then becomes something that you've got to change the assumptions of someone's books.

But I do think it just goes back to the point — your point, you know, if I think about bubbles, right, and I think about bubbles bursting, two things, you know, I think about.

One, an inflated market set of expectations. But two, oftentimes a whole bunch of things going on behind the scenes that most regular people don't understand because they end up involving some sort of convoluted set of financing.

And I mean, I think a lot of us remember that going back to 2008, right? And yes, there were a lot of investments going on, but a lot of it was around financial jiggery-pokery, right, and how things were being financed.

And I think you're seeing a bit of both of those here, where you've got off-sheet investments taking place, where it doesn't actually factor into the finances, but they're commitments.

But what does that mean? And is there full visibility?

And then, on top of that, you've got sort of, on this hand, a commitment to finance, on this hand, a commitment to deploy. You're selling things to people to buy back from yourself.

I think you have the makings of a classic sort of, "This could go wrong," in part because it's building complex financial solutions to make this all work.

Always means you're — I mean, it goes back to the conversation that we've had around AI models going rogue. When you add a whole bunch of things in, something's gonna happen you didn't foresee.

When you add complexity on top of complexity on top of complexity, you're setting yourself up for something that goes, "Wow, didn't see this piece supporting this piece supporting this piece," and it crumbles down.

And it is fair to say, anyone out there who's like saying he's got, you know, crazy, crazy, crazy — to your point, it's not a questioning of whether AI is going to add value and are we going to be spending a lot more time figuring out how AI revolutionizes our digital lives.

I don't think there's a question around that.

It's a question of how are we financing it, how fast does it come, how much compute do we need?

And then — and this goes back to something that we mentioned a while ago, and I talked a bit when Pow was on — I mean, how much do we need the latest, greatest model for 90% of what we're doing?

Right. And not, you know, not on the cutting edge of business, right? If you're out there and you're in finance and you need to make sure your models are — you're paying for the early access to Truth Social because that extra millisecond of heads-up on whatever, whatever's being said on Truth Social is gonna help make you some money, you probably want the latest and greatest models for doing whatever you're doing.

But for so much more of how people are using AI, and they will use AI, it is that further question of the race and how you think about this.

This is the moment of FOMO, where everyone wants to make sure they're not going to miss out. And so we are playing these huge investments out there on hope.

And I think, to some extent, we're also ignoring — I mean, this is my test, right? Ninety — like, if you think about a lot of the different use cases for AI, can you do it with Copilot in your average day?

If the average person can, then it's probably good.

And I think we've seen that improve. I think it's not perfect. I think people's expectations have improved as they've used these tools.

But that is my other part: Are we getting ahead of ourselves and building the latest and greatest and expecting that we are going to need all of these AI factories when I think, in many cases, it may be overkill for much of what we're trying to do?

Tim Hatt (12:39)
And I think the last point I would make on this is, as much as there is this sort of financial-market exuberance to fund the AI boom, and I understand where that comes from, China is still the elephant in the room here.

And right now GPUs don't have a long-term store of value — not yet.

And we've already seen with some of the frontier models coming out of China that are able to do it much cheaper and more efficiently.

What happens if we see the same in AI data-center compute and you start to erode some of the value domestically?

So there's a lot of variables here, I think we've gotta watch, but it's the orders of magnitude.

Peter Jarich (13:12)
Hey, if this means, for our core folks and all the folks that we see day in and day out on the telco side of things, that they can take advantage of — I don't want to say secondhand GPUs, you know, you don't want a secondhand GPU — but lesser-value GPUs to do what they need to do to deliver the value to their customers in a transparent way, cool for them.

Right. That's all I have to say is that maybe, in the end, if the answer is cheaper services that shall deliver what people want, then ultimately maybe there's a good side of things.

So we're moving from AI data centers, though, to other news.

If we were doing the full, the full version, this is the "On the Road and On Our Mind" section, the Willie Nelson section, as I like to call it. Willie is going to be joining us later in the year as a guest. He promises me.

But Tim, what are things that you were looking at that was maybe a little bit of a — not as big news as $500 billion of AI factory commitment, but also worth thinking about?

Tim Hatt (14:24)
Well, switching gears to the satellite sector, which we often talk about, but it's not about Starlink, SpaceX. We talked about their results last week.

But it's AST's results, AST SpaceMobile, who reported their Q2 results to a lot less fanfare than you'd expect, but still some interesting things in there.

And just as a reminder for our listeners, AST SpaceMobile is a competitor to SpaceX in the US. They do operate internationally as well. They are building a low Earth orbit constellation to provide direct-to-device connectivity in partnership with mobile operators across the world.

They reported their Q2 results this week and the quarterly revenue just over $30 million. Hit below analyst expectations, but they reaffirmed the full-year guidance.

They did see a little widening in the loss.

And I think the reason I bring this up is that it reaffirms how important scale is for being competitive in direct-to-device — frankly, any satellite sector.

And two, the fact that when we look at the direct-to-device landscape, the reason we talk about this as being a five- to seven-, maybe even a ten-year story is that there's a lot of time lag.

When you put the space rockets and you get the satellites launched, that's great, and you can sign partnerships with operators, but the revenue streams still take time to come through. People are not paying for the service yet.

And I think what you see with AST now is that they're seeing growth in the number of take-up from the mobile operator partnerships they have. They actually have quite a lot. You can check out our tracker for them — well over thirty.

But the revenue lag is there. And so they're gonna be having to look at bringing in the satellites quicker than they would expect to try and close that gap.

Peter Jarich (16:23)
But this is also why — and it's interesting because you see this in lots of new markets, right? We started talking about this a while ago: clearly there's going to be consolidation, clearly there's going to be folks having to drop out of the market.

That always takes longer than folks expect because there is a recognition, right, that it's going to take some time and that, well, these are long-term bets.

And I think, particularly in a case like AST, you've got a whole bunch of folks going, "Okay, they're kind of the next-best alternative in many ways to a Starlink, to maybe an Amazon."

So that's why you see so many different telcos partnering with them. You see sovereign solutions, you see them being the partner.

We talked about this a while ago, right, for Rakuten doing whatever they're doing in Japan. They sort of become the, "Hey guys, can you help us out?"

And, to their credit, as much as we —

I think the early story, and this is, I think, the fun part for me, at least, of the focus on NTN: I think a lot of the early focus around AST was around the tech side of things.

"Hey, they've got these — I mean, here's the Starlink side of things: tiny satellites, and you can send a whole bunch of them. And AST is going the opposite direction with these giant, with these giant satellites."

Crazy.

We talked about this. If you look at the drawings for what Starlink has said, SpaceX has said, in terms of their AI satellites, they're huge as well.

But I think that kind of — I don't know if it's the Musk of it all or if it's something else — but that didn't get a lot of news.

And I think that's in part because we've gone past the "How does one tech work versus the other?" to the business side of things and the recognition, as you said, that you're going to have to have scale and people are gonna have to place their bets, which they seem to do.

I mean, I do think there's obviously a lot of other players out there. We will see consolidation, but it's good to see them plugging along.

Tim Hatt (18:33)
What's on your mind, Peter, in terms of the other stories?

Peter Jarich (18:36)
Yeah, we're gonna stay on the satellite side of things.

So, you know, going back to the scale, and this is probably more of — I think it's gotta be more of a GEO than a LEO story.

I did see, sort of randomly, this was a fun one: Northrop Grumman has launched a new set of their satellites. The vehicles that go up there need to repair or refuel existing satellites.

It's, I believe, an MRV — is the Mobile Robotic Vehicle. I'm not —

But so it's interesting because, as we think about all of the number of satellites that are gonna be up there, you called it out exactly right. The cost to launch is one thing. Once you get that up there, you'd rather repair if you could.

That does go back a little bit to the technology side of things, right? Are we talking about replaceable, disposable, throw these up there? They live a certain number of years and they come down.

I think it's a neat concept. It feels like we're living in the future of robots going up there to replace or to repair satellites.

I don't know how well that works for LEO, where these things are moving in all sorts of much faster, different directions. They need to self-correct, but we'll see.

But the bigger story, and this goes back to my anecdote from the beginning: Reflect Orbital, a company that probably most folks have never heard of, got FCC approval to launch their first demo satellite.

It's a LEO satellite. I believe I'm getting it right. Its name is Arendil, because if you're cool, you name yourself after some sort of Tolkien thing.

I assume that's Tolkien.

I don't know. It feels like I'm probably gonna hear from them that it's their grandmother's maiden name, and now I will feel bad.

Tim Hatt (20:14)
Could be Disney. We'll see.

Peter Jarich (20:19)
But why this is interesting — this goes back to my anecdote — is that this is a company that is basically putting mirrors in space.

Right.

So they label this — and they've been around for a while, right? If you look at how long they've been out there, I'm gonna see what was the earliest blog where I started seeing them. It's like back in 2022, right?

So we're talking about starting to talk about this four years ago. So it takes a while to get this.

But the idea is you control light.

Right.

And if that sounds like an evil Bond villain kind of thing, it — I think there are some worries there.

If you look at their taglines, right, it's "We're launching the future of sunlight."

So they've got these big aspirations, but the aspirations are literally kind of what I was talking about, right?

You've got a search-and-rescue mission. It needs some extra support. And so can we get light directed into a specific area?

We want to more cheaply light up city streets.

We want to do construction projects 24/7 without having to worry about lighting.

Or we want to direct light to solar farms so that we can get them taking advantage of 24/7 light for power generation.

It's an interesting — I think for me, it's fun because, one, it goes back to the satellite technology is still evolving.

Again, whether it's the Northrop side of things or this, we're seeing new models come out of just the ability to launch these things means you're going to see new models.

I do think there will be concerns about how this fits with light pollution.

What goes wrong if someone turns their telescope in the wrong direction?

Like, there's a lot of potential for things to go wrong here, which is obviously why they're launching a demo.

But I think it's a fun, it's a fun example.

Tim, are you looking for 24/7 light at your house?

Tim Hatt (22:20)
I had a trip planned to the Arctic Circle, but I might not need it now if I can bring it right down to our street.

Peter Jarich (22:27)
Is this how we turn Cancun — is how we turn, rather, Canada into the new Cancun?

Tim Hatt (22:31)
That's pretty — pretty central. Pretty central.

Peter Jarich (22:35)
Yeah, just 24/7. It'll be a little bit — I don't know how that works with LEO.

I mean, I get a little bit worried of, like, if the satellite's constantly — like, how do I ensure the mirror is shining on me 24/7 when the satellite is on the other side of the world?

I think —

So I think it'll be interesting to see, as opposed to one giant spot beam that just circles the world.

But it is fun in the wild and wacky world of satellite.

On that note, we said we're gonna cut this one a little bit short.

Please, everyone, leave us some comments. Let us know if you like the short version, if you want us to go back to longer. Let us know what you want to hear about.

In the meantime, thank you, Tim.

Thanks, as always.

Thank you all for joining us.

And please remember to embrace the — well, seek out the intelligence, but embrace the noise, and tune in next week.

Thanks, everyone.

Tim Hatt (23:33)
See you guys.