Managing Partner and Chief Investment Officer, Atreides Management, LP
CNBC's interview with early SpaceX investor Gavin Baker
Most people think the future of SpaceX depends on one thing: Starship.
If Starship launches on time, the company wins. If reusability is delayed, the valuation falls apart.
Technology investor Gavin Baker thinks Wall Street may be watching the wrong machine.
In this full CNBC interview, Baker argues that the most important part of the SpaceX story over the next several years may not be happening in orbit at all. It may be happening inside enormous terrestrial data centers, where gigawatts of electricity are converted into AI models, tokens and revenue.
His claim is striking. If SpaceX can...
About Gavin Baker
Gavin Baker, managing partner and chief investment officer at Atreides Management, has been active in media appearances discussing SpaceX, AI infrastructure, and market dynamics. Following the SpaceX IPO, Baker praised Goldman Sachs and Morgan Stanley for their execution, describing the offering as "flawless." He argued that Wall Street may be underestimating SpaceX's potential revenue from terrestrial AI compute, stating that if the company monetizes three gigawatts of installed power at its announced rates, that could represent $150 billion in revenue not currently reflected in models. Baker also suggested SpaceX could become "one of the most important iconic companies of all time" and possibly "the greatest." In other appearances, Baker shared his view that open-source AI models shift margin from the model layer to the infrastructure layer, and that the market is "structurally short compute." He predicted the U.S. would print at least one year of greater than 5% real GDP growth over the next four years, driven by AI and deregulation. Baker described himself as a "double down late" investor rather than someone who panics early, and said he is increasingly looking at enterprise value to net property, plant, and equipment as a valuation metric, believing "installed atoms on Earth" will appreciate. He also stated that the public market has a greater tolerance for investment and a longer time horizon than many in venture capital recognize.
Transcript (53 segments)
H
Host0:00
AI competition and the direction of the market overall. Joining us here at Pace Nine, Post Nine, early space investor and Atreides Management managing partner and CIO Gavin Baker is back. Welcome back, G. Welcome.
G
Gavin Baker0:12
Thank you for having me.
H
Host0:13
I'm thinking of that day, the SpaceX IPO day and then a nice shot up to 220. How are you thinking about this drawdown? Is it sort of how IPOs go in the early days.
G
Gavin Baker0:25
I think.
H
Host0:26
It is.
G
Gavin Baker0:26
How IPOs go.
I mean the simple answer is I'm not really thinking about it. Just we'll see where we are in a year, two years, three years, my friend. Well done off once came on CNBC and he said, you know, he was on the show Fast Money. And he said he tried to be slow money, I think, long term. And I think there's a lot of wisdom in that.
H
Host0:45
Yes, there is still a process, though, to witness the street getting familiar with the story, getting its arms around the possibility. And now we're starting to see some pretty thick reports, analyst reports. And I mean, I think every underwriter has a buy on the stock.
G
Gavin Baker1:01
I will admit I do not know what where what the Wall Street ratings are, but that doesn't surprise me. And I think the big variable that maybe markets are missing with respect to SpaceX is they monetize a gigawatt of installed power at a meaningful premium to anyone else. They bring on those gigawatts of power faster and cheaper than anyone else. So they make the most profit per gig. And like, let's just say they bring on three gigs next year at their kind of announced monetization rates. That's $150 billion of revenue. That's not in the model. And you know bring on gigawatts of power. It's really hard. They've proven they're good at it. We'll see what they can do. But like I looked at the consensus, it was 73 billion. And that embeds very little for incremental terrestrial gigawatts that they bring on and monetize.
H
Host1:54
People are a little bit concerned, I guess, about the delay in the launch, right, of Starship. I mean, part of the bull thesis here is that they are able to do this much more frequently at lower costs. So any hiccup in that causes some concern.
G
Gavin Baker2:08
Starship is a really hard engineering problem. And you know, I think they've made incremental progress each launch. But if you're a long term investor, it doesn't really matter whether we get to reusability for Starship in three months, six months, nine months, 12 months. And I would focus respectfully people more on terrestrial compute in the next 2 to 3 years, how quickly they can bring that on and energize those GPUs and rent them out.
H
Host2:35
You mean what they're doing with Anthropic, for instance?
G
Gavin Baker2:37
Absolutely. I think you will. I think you're likely to see a lot more of that.
H
Host2:41
There are questions about the sort of value of that as the technology changes very quickly and as these new models come out and leapfrog the old models, and especially if they're coming from China, where they're open source and they're a lot cheaper about the entire economics of the system.
G
Gavin Baker2:57
You know, I mean, I guess there are questions I don't.
H
Host3:00
You don't know why. Why, what's your thesis here?
G
Gavin Baker3:02
So open source models fundamentally just shift margin percentage and margin dollars from the kind of model layer to the infrastructure layer. If you have cheaper tokens of an equivalent intelligence, we're going to consume more of those. And if you believe like I do, that we are structurally short compute, then you don't worry so much about margins at the model layer.
H
Host3:28
And that's why that feeds you to a bullish thesis on NIO clouds and other elements of the actual infrastructure build. Yes.
G
Gavin Baker3:36
I mean, I don't know if I would use the word bullish, but I do think it leads you to an interesting risk reward. You know, cloud Cerebrus, Nvidia, providers of compute, whether it be to frontier models or open source models.
H
Host3:54
But talk a little bit more about, you know, we were teasing out what you were just saying about. So we get these Chinese open source models like the Kimmy one, and everyone freaks out that they're better and they're cheaper and they're open source. You don't necessarily think it's bad for OpenAI and Anthropic. Why?
G
Gavin Baker4:11
Well, it may end up being bad for them, but I just think the fundamental principle, I think open source software like Red Hat, you know, back in the day and now part of IBM. That is much cheaper than software. It's free, but a token AI, when you consume it, it costs the same amount of compute of watts, of energy, of capex, of opex to generate an open source token as it does to generate an Anthropic token.
H
Host4:39
So then why do they why do corporates have to pay so much more for Anthropic?
G
Gavin Baker4:44
Well, because it turns out that there is. And this has been very surprising to me over the last year, that there is an enormous return to being at the frontier of intelligence. And I think the future is likely to be one in which maybe a majority of economic returns continue to accrue to these frontier tokens. But the majority of tokens processed are these very cheap open source tokens. And you can think of it, you use, you know, if you know every enterprise, they're going to have a symphony of AI models that they use. And you're going to have the, you know, the conductor, maybe that's Anthropic, maybe that's Grok. And then you're going to have a lot of people playing instruments. And those are these cheaper choices, these open source models that maybe have been fine tuned on an enterprise's own data. If they're worried about their data leaking to a closed source frontier model provider. And I think that is a likely future.
H
Host5:39
Where the US frontier models remain sort of in charge when it comes to the corporate customers.
G
Gavin Baker5:45
I think so, and there is also a lot of. Eric Fisher, a partner at Benchmark, posted something maybe 18 months ago that I thought was fascinating. Just there may be as much lock in at the product layer, at the harness layer as there is at the model layer and the products that Anthropic and some of the other frontier model companies have created cloud code Grok build around their models are, I think, an increasing source of competitive advantage in switching costs.
H
Host6:14
So I'm just trying to think ahead, like, will a corporate have a lot of vendor diversity in this symphony that you describe? Can it be one player delivering a suite?
G
Gavin Baker6:25
I think it is highly likely to be a multiple model future, and maybe some of these frontier models over time, incorporate the ability. They make it really easy. If you're Claude or your Grok to bring in a cheap open source model that maybe is a violinist.
H
Host6:42
Right.
G
Gavin Baker6:42
And the conductor orchestrates.
H
Host6:45
I mean, it sounds I mean, I'm just trying to figure out whether companies will want to minimize complexity, minimize vendor diversity, keep, you know.
G
Gavin Baker6:56
Well, I think they'll want to minimize complexity, but they will also want the most intelligence per dollar. Yeah. One thing that's really unique about AI, for the first time in my kind of history as a tech investor, being the low cost producer matters. Intelligence per dollar is what matters. And that is a function of what is your cost per token. And having the most efficient infrastructure is really going to matter. And then how much intelligence do you have per token? So to get to an equivalent answer B, K three, which is a great model and an achievement for China, it processes two. It needs 2 to 3 X more tokens. And each token costs compute, watts, energy, capex all of that. And so if you need to produce 2 to 3 X more tokens well you know, and so token efficiency and cost per token I think are going.
H
Host7:45
To matters.
G
Gavin Baker7:46
Yeah.
H
Host7:46
As long as, as long as American corporates are able to buy, I guess, to, to subscribe for there's all sorts of regulatory questions there with China.
G
Gavin Baker7:54
Well, I do think you will see we do have a really great American open source company and it's called Nvidia. Their model is actually really good. And I think that they are probably throttling it to a degree. You know, they don't want to scare and upset their customers, but I think they could very quickly bring an American source model, an American open source model, close to the frontier whenever they want.
H
Host8:20
Really? Do you have a position, Nvidia?
G
Gavin Baker8:22
I, I do.
H
Host8:24
Just just for and you do not in Anthropic and OpenAI.
G
Gavin Baker8:28
I do not. I was I was a SpaceX investor and there's been a détente lately between we've noticed Anthropic and SpaceX. But that came too late for me.
H
Host8:37
Well, I'm curious where you see Grok actually competing relative to the to the two of them.
G
Gavin Baker8:42
I think that the market really the market reacted in exactly the wrong way to Kimmy K three. It's wildly bullish for all of compute, wildly bullish for the AI and for trade. But I think it did really underreact to Grok 4.5 and Muse 1.1 for Meta. Grok 4.5 is on what is called the Pareto frontier, which just measures intelligence per dollar. And on some metrics, it's kind of outside the Pareto frontier. So you cannot get at its level of intelligence, cheaper intelligence.
H
Host9:15
And so it's not as intelligent as the Anthropic or OpenAI models, but it is more cost effective.
G
Gavin Baker9:20
It is more intelligence per dollar.
H
Host9:22
Yeah.
Someone was reminding me that. Speaking of the SpaceX shares that the xAI investment or acquisition came around 105. Do you see that as like a critical level for SpaceX?
G
Gavin Baker9:35
No, I think the stock's going to do what it's going to do. You know, I think there's a lot of people who maybe own stock privately and are shorting the stock to hedge that. Right. We will see. And then there's you know there's always for whatever reason a big kind of New York hedge fund short case and Elon companies. And, you know, there's a graveyard of funds that shut down because they're Tesla short.
H
Host9:55
But but a lot of people look at the valuation in this one.
G
Gavin Baker9:58
And then that comes down to how many gigawatts of power can they add. And the estimates on the screen might be really wrong. And then what happens. Grok 4.5 is a 1.5 trillion parameter model. They're trading a 2 trillion parameter model right now. I think the Cursor acquisition for 60 billion is looking very smart. That was us doing 3 billion in AR when they bought it. I think it's probably materially higher. So I think there's there's a lot of variables they have to execute on a lot of things. We'll see how this 2 trillion model comes out. And after that they'll probably be a 6 trillion and then maybe a 10 trillion. And we'll see how those trading runs go. And we'll see how those models benchmark when they come out. But if they are able to bring on a lot of power and sell that to other AI players at very high rates, and they are also able to bring on enough power where they can also serve their own models at a very low cost. I think that could be very interesting.
H
Host10:55
It does feel like Wall Street has really put a premium on like the model supremacy over the cost so far.
G
Gavin Baker11:01
I think that's a mistake.
H
Host11:03
Yeah, yeah. Because that's the edge that you think Elon has.
G
Gavin Baker11:07
I do think to the best of my knowledge he has. Oh. I mean, Jensen said the speed in which you bring on a gigawatt is literally cost, because every day that you're constructing a data center, you're employing. And by the way, we should talk about this like data center builds. They're amazing for blue collar America. You know, you're employing a lot of electricians, a lot of HVAC contractors, a lot of plumbers at like truly, you know, unimaginable rate.
H
Host11:32
Yeah. New York