Lambda Labs CEO Stephen Balaban on raising $1.5B equity round and building long-term GPU infrastructure

Um, our next guest is Steven Balaban from

Lambda Labs. Or is it just Lambda now? I think it's just Lambda.

Did we drop the labs?

I think we dropped the labs. Stephen, did we drop the labs? How you doing?

We dropped the labs.

We dropped the labs. Lambda.

Okay, I'm dating myself. Well, I at least I feel like a day one. I don't feel like a bandwagon fan cuz I'm using the old name. There's a little bit of cool. I liked it back when it was labs. But welcome to the show. Thank you so much for taking the time to talk to us. Uh, congratulations. You look incredibly yellow. You're making us You're making us look we got to put on the couple casuals.

Give us the news. What happened? Let's break it down.

Yeah. Well, so one day I was training some compets on my workstation. Next thing you know, we're raising 1.5 gigab.

Gigab.

We say we say we say

gab gigawatts, giga chips, gigab.

Yes. Uh yeah. What what does that actually mean? I mean we we we we see we see 10 billion, 100 billion, 10 trillion, quadrillion every day. Uh is this cash? Is this debt? What are you are you buying GPUs? Are you buying land? What are you doing?

All equity. Okay,

let's give it up for it.

Extremely well. like our our capital structure is really nice in terms of we've been very conservative in terms of the amount of debt that we've taken on and that's kind of been one of our philosophies and we've we've aimed to have

you know a business that's just super robust to ups and downs in the market because we're swimming with our swim trunks on.

Yep. And then uh

and then you uh

that's it. [clears throat] in exchange for the money you gave them. You gave them equity that that there's no one hand washes the other type thing where like they pay you, you pay them. It's all one round trip.

No, this this round was led by by TWWG Global, which is

financial investor, [clears throat]

which is Thomas Tull and Mark Walter. You may know Mark owns the LA Dodgers and and also now the Lakers. Thomas started Legendary Entertainment, which makes great movies like the Batman series and Dune and Inception. And

so it these are business partners who I've gotten to know over a number of years now, and this is just they're they're making some uh some big investments in the space.

Okay. Uh

I'm so happy you guys have your trunks on because not everyone out not not every player out there has their trunks on right now. And it's hard to tell who does and who doesn't. Yes.

But at some point we're going to find out and it's not going to be it's not going to be pretty.

It won't be pretty for people who are overlevered. And we just have this philosophy that with exponential growth that we're seeing in the AI industry. All of the upside is in the last period, right? You know, if you're if you have a doubling function, right? The sort of the definitional thing of that is that the last period is more growth than all the sum of the previous periods combined. And so from my perspective, it's just like stay alive and build a rock solid business because we got to capture all this amazing upside in the long term.

Yeah. So uh talk about use of funds. Uh

well even even before that maybe maybe uh feels like and it potentially an an advantage right now just in terms of focus is like being private. There are other other companies in the category that are public and they're now having to contend with

uh you know what's been a pretty big correction in in Neo at least a local correction in NeoCloud uh over the last month.

Uh has that been helpful in terms of the team of just like staying focused and you're not getting you know marked every single day? Well, I think that certainly that level of a distraction isn't helpful and I always encourage the company to just focus on building a heavy business for the long term. You know, if if in the short term the market's a voting machine, in the long term it's a weighing machine. We just got to build a business with good cash flows, a good capitalization structure that's robust. And so I kind of try to focus the team on that. I mean, these days the the secondary markets, as you know, are actually, you know, pretty deep for for for for companies that are that are kind of at our size. And so, I think that some of that can start to creep in.

Yeah, that makes sense.

Uh, where are you seeing value spending some of this money? I imagine that there's hiring, R&D, all the traditional things, but you're at a scale where uh it's a lot of money. How do you actually think about allocating capital at this point in this in this phase of the journey? It's been uh over a decade now, right?

Yeah. Uh we started in 2012.

Wow.

And was doing we were doing face recognition software and the Alexet paper came out. Wow. I mean that's how early it was and I I downloaded the CUDA confent library off of Google code and that will tell everybody kind of

how old school Lambda is. And you know as far as use of funds obviously a lot of it goes towards the GPU infrastructure that goes into data centers. Yeah.

We are also starting to put that into investments into data centers themselves. M

um we I I think that what we're aiming to do longterm [clears throat]

is kind of build this almost like Tesla for AI infrastructure where we kind of look at this as like a similar buildout that you would expect from the like electrification of the United States or the railroad and like a degree of vertical integration we believe is going to be in the future for us and is like the right direction and that that that goes from everything from, you know, energy procurement and construction because I think a lot more of the stuff is going to have to be behind the meter power plants to actual construction and design of data centers that can sort of rapidly adapt to the changing chips that go in, right? because the the rack densities and the the the movement from air cooled to liquid cooling that we're we're really pioneering alongside Nvidia.

These are all examples of use of funds and

it's exciting because we get to kind of make good investment decisions that are really sort of IR based in an almost industrial way which I think is unique from a company building perspective and it's a it's an honor to be able to do that. Can you get me up to speed on some of the trade-offs between like one really big mega data center and a bunch of really small data centers? How because there was a moment when we were just doing bigger and bigger training runs then it became RL all over the place then you actually have to serve these things but actually if it's going to take me 10 minutes I don't mind if you do it across the world and take it back but if I do care that it's right now I need it like right colllocated. How are you thinking about the tradeoffs there? So, so the the mix and the main driver over the next five years we believe will likely be mostly on the inference side.

Mhm.

If you look at some of the financial models that have either leaked or otherwise been published around what OpenAI thinks they're going to be spending, it looks to be about 50% on training and then 50% on inference growing towards

75% inference. Yeah. and you know a smaller chunk of that on training. [clears throat] And as far as like what that means for the larger data centers, I I certainly don't think that this is like going to a world where there's a bunch of micro data centers. I think that that's a little bit hard to sort of manage and deal with. But one of the things I think that you're going to start hearing a lot more of is how adaptable and how quickly can you bring on the data center in an incremental fashion because that's going to be a lot of the main drivers for how successful infrastructure builders like us are is how quickly and we're just focused on optimizing that time to first token for our customer.

How do how do you think about revenue quality and customer selection? because we've we've seen some some deals go down that's that look big and cool and good on the surface and then you dig into them and maybe the maybe the underlying uh infrastructure provider is not actually getting that great of a deal at the end of it. Well, we see we certainly see a lot of people with very high levels of customer concentration because Lambda started off as this developer cloud that evolved and morphed into a cloud that's providing for the biggest companies in the world. We have a really really strong user base. you know, um, if you if you look at our breakdown from our revenue mix in terms of you looked at like let's say our Q3 stuff, and I I don't want to go into exact specifics, but it's sort of like one or two big customers, a bunch of sort of the bigger, smaller customers, and then it's something, you know, it's a nice really big chunk of this long tale of customers that we have. And we have a very very you know I've seen some other people's customer books and I I can just say that we've got a very diversified customer base and that's kind of all part of the strategy of how do you build a great long-term business? Of course customer diversification is one of those parameters.

How do you think about diversity of of product offerings? Are you seeing customers ask for uh API endpoints for particular models or do they want access to bare metal? Um or have you gotten any customers that are like, "Hey, we just want, you know, you seem to know about this data center business. Can you just build a data center for us and hand it over to us when you're done and we'll just pay you as a consultant?"

We have no interest in doing that that one. that's, you know, we we want to do something that's really vertically integrated and,

you know, kind of going back to that like larger smaller data centers. I think the most important thing is just being able to deliver this incremental um live deployment for a customer. We have an entire full stack cloud product that, you know, it's got things like single sign on. It's it's got things like uh long-term high-speed AI file systems. It's got instances that go down from one GPU to an entire cluster with one-click clusters that we've that we've got. And so we've built an entire cloud platform. We have previously been in the inferencing space where we're actually giving an API for inferencing. And we've actually exited that business to just focus. I I think that that's like one of the things that we really try to do at Lambda is just say where are we making money? What are good investments? and where are we going to really dominate the market and focus there and so we've actually exited for example the inference market we we had a $200 million plus a year hardware business that we've exited right you know [laughter] I mean it it actually like kind of crushes me because like that was the business that got off the ground but

can you imagine just like winding down like well we're just gonna take this business and not do a $200 million a year business anymore because we're trying to focus

that is Crazy. That is crazy.

Thanks, Scots.

Um, I have a I have a crackpot theory that I'd love to run by you. What do you think the odds are that uh the I like I noticed I was traveling in I was traveling in Mexico and I noticed that Carlos Slim is the richest man there. Uh, and he's a telecom magnate. He he owns a lot of the telecom infrastructure. Um, and that's true for a lot of a lot of countries. the the the richest person in that country is a telecom person or a mining magnate uh in the sense that they've been able to corner a resource a physical resource infrastructure and that's generated a lot of wealth for them and I was wondering if you had a thought on do you think that in the future we'll see uh the some of the wealthiest most powerful people from other countries non-American countries um be uh you know GPU cloud hosters or data center develop velopers like is this going to be a new boom uh across the globe? It's kind of a different twist on the sovereign AI project. I was just I was just wondering if there's if there's going to be some some way that this plays out where there's this sort of like one-time opportunity to kind of get a cornered resource or is the nature of the internet such that the compute is actually much more funible than um than say you know

telecom or you know like copper in the ground. localization. There's such a physical localization. I think if you look at telecom, you look at cable as well as regulated utilities from an energy utility perspective. You know, these are all things that benefit from a physical geographic monopoly, right? And and AI data centers don't have that same thing. Now, I just want to step back for a second, guys. The United States is basically the only country in the world. We have the most unbelievably good economy. This is the the the idea that there's going to be these sort of like massive AI infrastructure projects that I think are going to be like super super successful outside of let's say China and the United States right now is really increasingly big question mark and I I just am so bullish about where we're going in America that I I don't really pay a lot of attention to and that our focus is just in you know in in North America generally And I I I just that's kind of my perspective on it to be honest.

Yeah. Yeah. Know that that's really helpful. I agree. Um it's uh it's interesting a toy. I mean there's a lot of money being thrown around with some of these projects and uh I'm always interested in you know how they all shape out. Uh last

go.

Yeah. May maybe go for it. I was going to ask uh

like how you guys are navigating energy constraints with with new developments. Are you seeing uh we've heard you know anytime obviously there's like massive demand for something new sources kind of come out of the woodwork. We've seen back and forth some people that are building AI infrastructure say like energy is our primary constraint. Others are saying actually that's not my you know it's u so where where do you sit? We are aiming to reimagine the sort of step process from whether it's photons or molecules of natural gas to tokens.

And we strongly believe that a lot of this is going to have to come in reimagining like well how do you inter interact with the grid? How much power generation do you bring to the grid yourselves? And I think that that's the the the successful AI infrastructure companies in the future. Again, this is like why I kind of said like I look at this like Tesla for AI factories, which is you got to reimag how the world has worked previously and you have to kind of bring together this level of vertical integration because that's how you move fast, right? you know, when you can control every step of that way from uh the power generation and not having to necessarily deal with a um regulated utility and you can go and do behind the meter generation with a natural gas power plant. If if that can speed your time to market up, this is just so important. And [clears throat] that's kind of how I approach it, which is there's certain barriers like regulatory barriers which look you try not to run through those like a brick wall because it's kind of like an immovable object. But if you can if you can just bring your own if you can just sort of get around that sort of regulatory constraint of having to interact with a regulated utility by bringing your own power to the grid, then that's that's what I think is going to be successful.

Yeah, makes a lot of sense. Uh, thank you so much for taking the time out of your busy day to come and hang out with us and answer us questions about

Jordy. John, thanks for having me, guys.

It's always a great time. Congratulations.

Have you seen the new Gemini 3? This is like

Yeah. Can you give us your review and actually explain how it interfaces with your business? I'd love to know.

So, so I haven't I haven't used I haven't uh I haven't used the Gemini 3 yet. I've seen the uh the updates. I I'm still, you know, hey, Synindar or whatever, give give Lambda's enterprise account access. We're on Google uh Google Suite or Google Enterprise or whatever it's called now. So, we'd love that upgrade. But I'll tell you what, this is the cool thing.

I use things like chat GPT and Grock to learn more about topics like regulated energy markets and how to build power plants and data centers. And that makes Lambda faster at standing up AI data centers.

Mhm.

And I I pay attention. I actually just like kind of do what the AI tells me to do.

And that gives more compute to the AI to train bigger models which makes [laughter] faster

AI is working through you to make more AI. the

the beginning of these types of positive feedback loops and and I think that if you privately talk to a lot of executives,

you'd be surprised by the amount of,

you know, the strategic conversations I have with these AI models has gotten more and more advanced with with with the the level and quality of the model. The first versions were not great and I didn't really take a lot of its advice, but now I am. I mean, next thing you know, it's sort of like, well, you know, maybe AI is the one making the the running the show [laughter]

into sessions. Yeah. [gasps]

Next thing you know, we'll be hanging out on TVPN discovering novel physics with with Gemini 4. You know, we'll we'll see how far we get. [laughter]

Yeah. Yeah. It's it's a good time. Well, thank you so much for coming by the show. We'll talk.

I have I have a bunch more I have a bunch more questions, but but come back. Let's get you back on in uh before before the end of the year and we'll continue the conversation. Congrats to the whole team.

Yeah, we'll talk to you soon. Take care. Have a good one.

Bye.

Uh quickly, let me tell you about Privy. Privy makes it easy to build on crypto rail, securely spin up white label wallet, sign transactions, and integrate onchain infrastructure all through one