Lambda Co-Founder: How to Pivot and Scale to $3 Million in 1 Year

The company's gone through at least about 6 pivots. We had a $40,000-a-month Amazon Web Services bill, and that bill nearly put the company out of business. One of our early investors said, "Hey, you know, why don't you try building your own infrastructure?"

And so that was, like, this huge pivot that we did, and it completely wiped out this $40,000-a-month opex. And it just—exponentially growing to, like, $3 million of revenue that first year in 2017. Just to talk about pivots, if you look at any of the really big companies, all of them started with—

This is Stephen Balaban, co-founder and CTO of Lambda. He helped scale the company into a multi-billion dollar GPU powerhouse powering the biggest tech giants today, and now he's breaking down how pivoting saved the business and why stepping down as CEO might be the key to scaling yours.

It's just really amazing to see if you actually look at the history of all these companies, how much pivoting is involved, and how much the company completely changes over time. At some point that becomes just a normal way of life for you as a founder.

You try something, does it work, then keep doing it. If it doesn't, then move on to the next thing. But when you're doing a pivot, how do you decide whether something is a distraction versus the next thing for your company?

We did this very high-risk $60,000 capex at the time. People didn't like investing in hardware companies, people didn't like investing in capital-intensive businesses, and the result of that is that they missed out. I have raised a couple billion dollars across the credit equity spectrums, but it was never easy for this company.

You made the personal decision to go from CEO to CTO. Why did you actually want to do it?

There's so many very, very hyper-stressful things about being the CEO, and things I really disliked about the role. First of all—

Welcome to NEW ECONOMIES.

Hey Stephen, welcome to NEW ECONOMIES.

Hey Ollie, how you doing? Thanks for having me.

Amazing to have you here. You know, everyone is freaking out about GPUs now. Lambda, one of the biggest GPU companies that everybody knows about, that you co-founded. What the hell is happening in the world of GPUsright now?

So I'd say that what we're seeing is kind of continued surprise on the upside in terms of the performance of foundational models. So, you know, the frontier labs continue to produce models that have, I'd say, outsized value that's being created.

And the result of that is when a new model is released, sort of every single time, it's able to address a new set of tasks. And the sort of the more surprise there is on the upside, the more there is, like, sort of an increase in the intrinsic amount of demand for compute in the world.

And I think that's fundamentally what we're observing here. And really, we're just, I think, related to that, it's a lot of people are starting to come to the realization that every single one of these new models is just continuing to get better.

We're expanding the capacity of the model, we're expanding the amount of compute that's being used to train those models, and that cycle does not seem to be slowing down.

These updates are happening also just so quickly. All of these AI models are releasing new updates. You know, I kind of used to compare it to the iPhone updates. Maybe every 6 to 12 months. These updates are happening, I mean, every couple of months now,right?

In terms of, like, the fundamental, I'd say, class of the model, we're not necessarily seeing it every couple of months in terms of, you know, going from, let's say, a hopper-class model to a black-wall-class model to a, you know, VR-class model.

But there's certainly a lot of, you know, kind of constant updates in the market of releasing of models that have been, you know, maybe previously need a little bit more AI safety work to be released or stuff like that.

And so, yeah, the pace has been dizzying in 2026. I'd say that having been in this industry and following it for a while, 2026 is definitely one of the highest velocity years that I can really remember. Even more so in a sense than even when ChatGPT originally came out.

I'd love to get your take here. It feels like 2026 has been a massive breakthrough, where actually now the speed of these AI models is so good, the accuracy is getting better, but also just the user experience. 2026 feels like that inflection point.

Do you agree with that?

I do agree with that. I think that the combination of, let's say, some of the coding harnesses, the quality of the frontier models, their ability to autonomously software engineer, that is kind of unlike any other year we've seen before.

What's been most surprising to you about the updates since the beginning of the year?

Well, I mean, the big one is, you know, going back to kind of February of 2026 when Opus 4.5 came out and there was that really wonderful blog post that was published by Anthropic on creating a C compiler from scratch and how they kind of ran it 24/7 in a loop.

And I was able to, for example, take that blog post and kind of paste it back into Opus, you know, using Claude code to generate a harness that would be similar to that and use that to do autonomous software engineering.

And so I remember I was about two weeks late. You know, I was very busy and I hadn't had time to, like, sit down and, you know, relearn all the new things that had come out, which is, like, definitely a trend that I think I've seen a lot more in 2026 is that people, "Oh, you're like two weeks behind," which means that you're kind of feels like you're a light year, you know, you're a lifetime behind.

And I think that that was the huge, this huge breakthrough in terms of the duration and the quality of autonomous engineering that can happen with these models.

You know, some people might not know this, but Lambda actually is a 13-year-old company, 14-year-old company,right? Started in 2012. But back when you guys were starting with your brother, I'd love to get into the story of how you, you know, built the company with your brother.

But, you know, 2012, very different times. It's almost back then, if I seem to remember correctly, we were encouraged to buy GPUs. It wasn't kind of like necessity. Now it's essential.

Yeah, well, so back then, Lambda started off as a facial recognition and image recognition company. And 2012 was a very sort of fortuitous year to start that kind of a company because that was the year that the AlexNet paper came out, which was, you know, out of Jeff Hinton's lab at the University of Toronto.

It was Alex Krushevsky and Ilya Sutskever. And they really kind of cracked what was at the time a really impossible benchmark, which was called ImageNet. And ImageNet was this internet-driven image dataset. And people have been throwing all kinds of traditional computer vision methods at it, whether it's SIFT features or, you know, hand-created features with support vector machines.

And then come along the ConvNet, they threw the ConvNet at it. And I think the really big twist on it, you know, architecturally is quite similar to Yann LeCun's sort of papers in the 1980s. But the big breakthrough was they trained it on a couple of NVIDIA GPUs.

And so they were able to really scale up the amount of compute that was applied to the same model. And that kind of got me very interested in neural networks. And so that's how Lambda kind of got started.

And that was, like, I guess you'd say the sort of the environment in which it was started.

When did you realize that perhaps, you know, the facial recognition, the facial-like side of things, was it going to, you know, be the entire focus,right, of the company? You guys had to pivot.

There are at least, I'd say, got to be about half a dozen pivots that the company's gone through since its founding. At some point, that becomes just a normal way of life for you as a founder, is that you try something.

If it makes money or does it work, then, you know, keep doing it. If it doesn't, then move on to the next thing. So that became a little bit of a way of life, especially in the early days when you're just getting started.

I ran the face recognition API. So we actually even started off doing sort of augmented reality software. You know, I really wanted to make software for augmented reality glasses and things. And that's kind of how I even got into face recognition.

We launched this API. I think I probably ran the API for about four to five years. It kind of petered along, made a little bit of money, a couple thousand dollars a year maybe. So not really enough to pay the bills.

But, you know, was able to sort of survive by doing consulting and sort of whether it's computer vision consulting, software engineering, as a way to kind of keep money in the bank and keep the company going. Shortly after that, you know, I was working on this thing called Lambda Hat, which was one of the problems you have with image recognition datasets is that there's not enough data out there in the world to gather for just raw images, faces, all kinds of situations.

And I wanted to make a platform that allowed people to kind of gather that dataset for training these neural networks. And this is in 2013. It was a baseball cap with a camera embedded in the tip of the brim and the hat that I wear and sort of the Lambda Hat thing is a reference back to that original, one of those original first products in the hardware side of things.

My brother joined in 2015, actually. And his only rule, the only rule that he said when he was joining is, "We can't work on Lambda Hat," because he didn't really believe in that product. And so I sort of begrudgingly agreed, but it was a great thing to have him come on.

He actually joined from Nextor, and he was one of the early engineers at Nextor, then joined Lambda.

Very cool. We just had the founder of Nextor on the show, actually. So full circle moment.

And he would always compare me to Near of 2. And he's like, "Well, why can't you raise money like Near of can raise money?"

Near is a good storyteller.

And I was like, "Well, it's a high bar, Michael," you know?

And completely different industries as well, let's be honest. But 2015 then, okay, so you're working with your brother. You go through a bunch of pivots, six, seven pivots at this time. You're now on the train track on actually GPUs, Neo, gigabatteries.

This is going to be the way forward. What did you see so early that other people didn't quite see?

Well, I mean, the first thing I saw was really this AlexNet paper. And then after that, I think the second really inspiring research paper that I saw was the Alex Graves handwriting generation demo. He's got this amazing demo.

It's a recursive neural network, an RNN, that essentially can generate human handwriting, as well as he also did part of that paper, a demo of generating Wikipedia articles and, you know, generating the raw sort of Wiki markdown sort of formatting.

And I thought, "Wow, okay, you know, maybe in the near future, this thing is going to be able to generate images or, you know," so that was, like, for me, my earliest exposure to what people today would call generative AI, which, you know, generative AI, I mean, I can point to research papers going back to, like, you know, if you look at the wake-sleep algorithm papers, which is DIAN-94, that was a generative neural network.

They had a recognition pass and a generation pass that could actually generate and sample from the sort of learned distribution of the network. And so I think that there's a lot of people who saw this stuff very early.

It just didn't really start taking off until I think 2017 was a really special year in terms of when deep learning specifically started to get sort of this runaway traction within the academic community. Because even between 2012 through, let's say, 2016, there was still a lot of skepticism within the academic community of these neural networks.

They're not understandable. A lot of people, very credible academicians, would say, "Well, maybe there's going to be a breakthrough in algorithms that's going to, you know, we're going to be doing sort of random forest-like models that can be trained on your MacBook or something."

And I think seeing the improvement of the models with the amount of compute that was put in and sort of like the bitter lesson, you know, Rich Sudden's bitter lesson essay, that kind of getting bitter lesson pilled was like that turning point.

When you were speaking to all these academic communities back in 2017, you know, what you were speaking about and eventually you have to get this company going,right, and have to start making money.

We only stopped at the Lambda Hat pivot to, you know, at that point, after the Lambda Hat thing, we were like, "Okay, well, let's do AI consulting." And it'll be like, "We're going to build the Accenture for AI."

And of course, no VC liked this idea. This is a horrible idea in many ways for a venture-scalable business. It's a great idea for making money, I think, and a great idea for building a company, but not necessarily for a venture-scalable business.

First DeepDream came out, which was Chris Olah and then some others out of Google. And that was like really cool. He's like very chippy images. Do you remember the DeepDream stuff?

Yeah.

Lambda, through DreamScope, this app we were running, was probably the largest generator of DeepDream images in the world. And we got to like a million users on this app, processed tens of millions of images. The Gattes-style transfer algorithm came out where you can turn a photo into a painting.

And then it sort of like, you know, it crescendoed and then the FAD peaked and then there was this exponential decline down the other side. And at the height of that peak, we had a $40,000 a month Amazon Web Services bill.

And that bill, you know, nearly put the company out of business. And one of our early investors, this guy George, who was the eighth employee at Google, said, "Hey, you know, we always built our own infrastructure. Why don't you try building your own infrastructure and see if you can stand it up?"

And so we did this very high-risk $60,000 capex at the time to stand up a bunch of servers. And we were so like worried about it. We were like, "Okay, well, let's not just do servers, let's do workstations.

So worst case scenario, we can part this thing out and sell the workstations." And we turned that on and it completely wiped out this $40,000 a month opex. And so that was like this huge pivot that we did.

We're like, "Oh, wow, we're saving more money than we're making. Why don't we consider selling these workstations and providing a cloud service to other people who are doing AI research?" And that's how our compute business was born. And that pivot really was in sort of 2016, 2017.

And that business just really took off. I mean, I think, you know, like many other entrepreneurial journeys, it's like one of these things where I forget whether it was Patrick or John Collison was saying kind of, you're pushing a boulder up a hill and that's like the early part of the company.

And at some point, you go over the hill and then you're running away from the boulder, like Indiana Jones style. And that 2017 year was definitely when we started running away from the boulder because I think in March, I did $35,000 in sales of the workstation, $70,000 in April, $140,000 in May, and it just is exponentially growing to like $3 million of revenue that first year in 2017.

And that's when the business was like, "Okay, we're in the compute business now." And that was the real pivot.

I find it so fascinating. You know, if you compare this to the AI companies today, they are growing at 30, 40, 50% every single month,right? How did you stay focused back then, but also how did you make sure you're not running away from the wider picture?

You know, you've got to think about hiring, you've got to think about actually just keeping this thing going.

I remember, first of all, there being some pretty critical points of like, well, how do you decide whether something is a distraction versus the next thing for your company? When you're doing a pivot and when you're really doing something like that, because remember, we had raised about $600,000 for DreamScope, which was an image generation application that was powered by neural networks.

How do you decide whether or not selling workstations is a huge distraction from this thing you've signed up to do? Right? And certainly anybody who's in business will tell you that giving up is like the last thing you can ever do in a business.

Right? You just like just never give up and you'll eventually be successful. And so how do you determine what is a pivot versus what's giving up? And that's actually how I, you know, really thought about that kind of a pivot.

And I was worried that it was a huge distraction until it started sort of making money. Now, once you start making money and once it starts like really having a level of commercial success and you're running away from the boulder, you know, you just start hiring where you need it,right?

Well, what was the first, one of the first things that we needed to hire was somebody to help assemble the workstations,right? So, because I was assembling, I, you know, Michael, my brother and co-founder, Jackson, one of our early employees, and, you know, we were just assembling the workstations ourselves.

And then we eventually hired somebody to just kind of like do it incrementally at the edges where there's, you know, where there's the need. And, you know, so I'd say that's how you kind of stay focused. You just keep saying, "Okay, where am I differentially able to better focus my time by taking something off my plate?"

You know, so many people say, "Do not start a company with your family because it always ends in disasters."

My dad said that too.

There you go,right? The best parents say that, even my parents said that. Why did you want to do this with your brother? And how did you learn to disagree and actually just still love each other, hopefully at the same time?

Yeah, well, my dad said that because his dad was in a business with his brother. And, you know, so my grandfather on my dad's side started an office supply business with his brother. It was called Service Office Supplies and it was in Detroit.

And it was servicing a lot of the big automotive makers there. It was a pretty sizable, you know, business. You know, it eventually sort of, I think, didn't end up working out or, you know, but they ran it for a while.

And the thinking there is like this. Okay, so one, when you're a solo founder, I'll speak to all the solo founders who are out thereright now listening to this. It's so lonely. And I know what that feels like.

You know, Lambda started in a small apartment in Chinatown and you wake up by yourself. You start working on things. You go to bed, you know, you're largely by yourself because you're really just getting something started. And the sort of sum of the sort of emotional morale of a company is just your own emotional morale.

And so if you're having a great day, you're having a great day. The company's having a great day. If you're having a bad day, then the company's having a bad day, the entire company. Whereas when you add another person in there and you're sort of adding together those two signals or averaging them together, if you're having a bad day, maybe they're having a fine day and you kind of balance out and makes you feel a little bit better too.

So I think that that's definitely one aspect is that it's important to, I think, from an emotional regulation perspective, have a co-founder. And then, you know, Michael and I really, that trust is so implicit and default,right? You know, that is the foundation of a great team is that trust and that trust is there.

We both have each other's really best interests in mind. So I think that that's a huge advantage. And there's some great examples of family businesses, whether it's Anthropic or Stripe and Lambda as well. And so there's some great examples of successful, scalable venture-backed businesses that are siblings.

And Michael and I had experience working together too previously. So, you know, when we were growing up, we would do some web design stuff together and, you know, we both started our own sort of online businesses together. This is like in the sort of early 2000s.

And so we had some experience doing it and it was successful then.

You know, you've both grown this company from basically nothing into a ginormous company today. And at some point, the CEO of the company is no longer the best person to run it or the company outpaces them. And I think this is, you know, always okay,right?

And not enough people talk about it. You actually recently announced you made the personal decision to go from CEO to CTO. Why was now theright moment to do this? And why did you actually want to do it?

Yeah, well, so now I'm on the other side of it and I've never been happier. It's amazing. I love what relief,right? There's so many very, very just hyper-stressful things about being the CEO and things I really disliked about the role.

And Michel, who's come on, has done such a phenomenal job. He's come in, really listened to the company, listened to people who are in the trenches day to day and sort of really coming and becoming a part of that team.

And it's been absolutely phenomenal. So now I've got that benefit of the hindsight,right? Obviously, it's a lot scarier when you don't have that benefit of the hindsight when you're kind of making these decisions. But the way that I kind of look at that, so, you know, I was the CEO for the first founding 14 years of the company and then in May handed it off and passed the torch over to Michel and now I'm the CTO.

The main things I think about were, well, one, you know, what got me into this in the first place? And as you can probably no doubt tell, it was the technology. It was about building products and about identifying a need in the market and satisfying that need and just delighting customers.

That is the thing that I absolutely love. And once you get to a certain point, and especially in as capital-intensive of a business that we're in, a lot of it becomes a real fundraising treadmill. So like the company is like just constantly, whether it's across the equity, credit, equipment leases, and operating leases, you know, these are all kind of similar types of underwriting and fundraising experiences.

It's just like this constant treadmill of fundraising that to me is not as exciting. It's one of these things where I always, you know, when I talk to other people, I kind of never consider myself a really great fundraiser, but then I kind of think like, okay, well, I have raised, you know, a couple billion dollars across the credit and equity spectrums, but it was never easy for this company.

It was always a huge slog and it was something that I frankly didn't really enjoy as much. I looked a lot at like kind of where some of the people that I look up to who I wouldn't consider my, they're not my peers, but they're just people who I look up to, like whether it's Bill Gates or Larry Ellison or Larry Page.

Not a coincidence that a lot of them, you know, eventually stepped out of the CEO role into something that was maybe either more product-oriented or more technology-oriented. It just felt like theright time. When you're not enjoying something, you know, for long enough, you look yourself in the mirror and you're not like kind of tap dancing to work or like just super excited, then it's like kind of probably a good time.

If you're enjoying today's episode, hit subscribe at the bottom of your screen. The guests on this show are getting bigger and bigger and we need your help to grow this channel. So if you're enjoying today's guests or any previous episodes, hit subscribe and I promise you it'll be worth it.

Now back to today's episode. When you were thinking about stepping aside and, you know, mixing up things a little bit, who's the first person you had to tell, I guess you suppose your brother first. And what does that conversation actually even look like?

I'm a very transparent communicator with not just my brother, but like a lot of the people that I work with, maybe sometimes to my detriment in terms of,right, like just kind of very transparent and open about like, hey, look, you know, really not liking this or looking around and saying like, well, do we have any other options?

Seeing some potential, but, you know, maybe deciding, well, we don't really have any other optionsright now. So we're kind of stuck in this situation where, you know, where maybe I was doing something I didn't want to be doing.

And then, you know, having some conversations with some of our key investors about it. And again, this is not advice, it's just more like how these things happen. I'm just very transparent with the people that I work with.

And there's another aspect too, by the way, which a lot of people don't fully appreciate, which is you as a CEO and a founding CEO are oftentimes, more often than not, providing services for the company at far below market rates.

And what I mean by that is this, okay? Yeah, yeah. We're all vested out. I mean, like, you know what? I vested, I finished vesting my shares in 2016 or something,right? But like, you know, you're not making or getting paid or oftentimes asking for some crazy, you know, market, let's say options package or salary that would be equivalent to an outside executive, an outside chief executive.

As part of that dynamic of providing it to the company at below market rates, quote unquote, the company needs to grow into being able to afford

a high-quality outside executive,right? Do you see what I'm saying? Like, I think that's like, that's one of the things I've thought about in this process is that, you know, that's like, that's very common, especially for founder CEOs. And so I think that if the company's too early stage, I think that sometimes that probably doesn't work out as well.

You have to be at theright stage, but you also have to find theright person. How long did the CEO search take? And for anyone listening who may be thinking about something similar, same type of stage company, looking to maybe move to a different type of position, can you maybe just explain what does that CEO replacement process look like?

Yeah, well, you know, the board and everybody establishes a search committee and, you know, you work through your networks and you try to think about what kind of people, what kind of profile. Every business is unique and different.

You know, Lambda is an exceptionally capital-intensive business, one that requires major infrastructure experience, one that requires sort of a deep set of networks within the investment communities. And Michel had all those things, plus amazing operating experience, plus amazing investing experience from South Bank International, plus, you know, just this telecommunications industry, very similar.

And there's a lot of things that rhyme with the telecommunication industry with Lambda in terms of the sort of amount of capital investment, the race to achieve scale and a certain position within the market. So there's a lot of similar things.

And so we've got some great board members like John Donovan, who used to be the former CTO and then eventually CEO of AT&T. We've got Michel, who used to be the CEO of Sprint. And so they used to compete, you know, they used to be competitors, I guess.

And it's been cool to see that sort of some of those teams come together.

We've mentioned a few times, this is a very capital-intensive business for us. And what I found really interesting is you don't have the typical type of investors, VC investors who are on your cap table. You've chosen a very different approach.

Maybe can you explain a bit about that?

I think that there's just inherently within Silicon Valley. And I mean, it's not for lack of networking or for lack of talking. I talk with every single one of the name brand, you know, top VCs. And I pitched them every round.

I was very polite to continue to go back, hand in hand every single time, swallowing my ego, even if they've said no six times, maybe this time will be different. But, you know, I think that just goes and speaks to for a very long time, there was extreme skepticism within Silicon Valley of this sort of industry in general, I would say, the industry in general, whether it's OpenAI and Anthropic,right?

I mean, these were previous to ChatGPT exceptionally bizarre investments for the vast majority of the investors. You know, you have to remember in 2021, everybody was investing in crypto. So I think that that's one of those things where Lambda specifically has so many things that are very unique about it in terms of we were a hardware business for a long time.

We had, you know, hundreds of millions of dollars of revenue. Great, awesome, profitable. And many of our rounds in a very avant-garde, as some very prominent VC said, punk rock profitable. You guys are punk rock profitable. But

I think that people didn't like investing in hardware companies. People didn't like investing in capital-intensive businesses. And, you know, I think that that's the result of that is that they missed out and then also that we ended up going through alternative routes for people who understood the business and were willing to invest in that kind of a company.

If we just look at the opportunities there, it's ginormous rise. Gigawatt factories, GPUs are at an all-time high in terms of demand. The scale of opportunity is unlimited. Where does this thing even go next? Because it's huge.

So having seen the progress of these models over the last 14, going on 15 years, and seeing that every single time, there's kind of three legs to this, three legs to this stool. It's like model capacity, the amount of compute that you're using in both the training, maybe the reinforcement learning, the post-training processes, and then the data sets.

And it just, we've consistently seen you can expand model capacity, you can expand the amount of compute, whether it's through reinforcement learning, synthetic data set generation, all these different techniques, also just massive expansion of the corpuses,right, through these projects of scanning and scanning books.

We seem to be on a continued path of every generation of models going to get better and better and better. And I don't, you know, what is the likelihood that we're somehow, oh, this is the last model that's going to end up no longer scaling?

And we are absolutely, as an industry and as an economy, making these massive infrastructural investments in order to ensure that at least the compute side of that expansion continues on, which if the compute side continues on, then the models will naturally expand to fit and the data sets will, I believe, continue to expand.

But the physical part of that, which is the infrastructural buildout, which involves both power and compute, is, you know, continuing on, continuing to get funded. And investors are starting to really like GPUs as an asset class. You know, if you flash back to the 2000s or whatever, it's not uncommon if you're familiar with like the insurance industry, for insurance companies to take their float and invest it into really rock-solid, stable assets.

You know, they'll buy a power plant or they'll buy a toll road or something like this. This is what NVIDIA GPUs are becoming today. They're becoming an asset class that's super investable that investors really like. They like the return on it.

Lambda has had V100s, which are a GPU from 2017 in our cloud, still nearly fully sold out and occupied by our customers,right? And so if you go back just a few years to people making, you know, short sellers making these insane statements like, oh, there's only a three-year usable life to the GPUs.

Well, like we have real hard cash flow evidence that the V100 is a GPU that can be launched in 2017 and still be generating money. I think that we might still have a V100 in our cloud by 2017.

And so that'd be a full decade of use.

That's super fascinating. These GPUs, if they become some form of asset class where the typical any consumer financial provider can invest in them, this becomes super interesting.

Yeah, both the institutional investors with any asset, as it matures, it goes from private institutional investors, public institutional investors, and then eventually people in retail get access to these assets. You eventually have an entire ecosystem that evolves around it, whether it's a spot market, which can enable a futures market, which can enable all kinds of things that people on the outside sometimes think, oh, well, this is just financialization and, you know, there's not really a lot of value, but there's a huge amount of value creation to having a futures market,right?

I mean, farmers in the United States rely on a robust futures market in order to hedge uncertainty, in order to have a place for them to have an offtake for their crops. And that allows them to get financing and do all kinds of stuff that's super valuable in the economy.

And as the compute market matures in the same way that, let's say, the energy markets in the United States have matured, I think that we're going to see a lot of very positive benefits from that.

There are so many opportunities, but I imagine there are also so many challenges. What keeps you up at night? You know, when you see headlines like Elon Musk a couple of days ago announcing he wants to build the largest gigawatt factory five times, I think, the size of the largest one already,right?

Does that keep you up at night or you just like, no, we are building and we're very focused in our lane?

Well, this is the amazing thing about the world is that, first of all, it's huge. The good news is that, you know, we've got a GDP, global GDP of 126 trillion or something like this. First of all, one, Terrafab, I mean, that seems like that's going to be an amazing American alternative to things like TSMC.

And I think I absolutely applaud that. I think that, you know, we love and use NVIDIA chips and maybe sometime in the future, some chips from NVIDIA will be created on a Terrafab. I mean, who knows what the future could look like?

I certainly think that people would have found it surprising that SpaceX AI was able to provide compute to Anthropic and Google and others just a few short period ago. Elon Musk is somebody who I absolutely look up to.

I look at what he's doing for humanity and feel a sense of like, okay, Elon's really on it and is going to make for an amazing super science fiction future. And that's actually one of my, just on this topic of Terrafab specifically, when I was at University of Michigan, I had an AI class with Ben Kuiper, who is a professor there.

And he used to say, hey, some of the best ideas come from science fiction. And that always stuck with me. And I oftentimes would read, whether it was, you know, when I was doing the augmented reality stuff in the beginning, I'd just read a few different books, including The Diamond Age by Neal Stephenson.

And I had been watching Serial Experiments Lane, which has, there's some sort of augmented reality aspect to it, as well as Deno Coil, which is an anime about augmented reality. And, you know, Elon is one of those founders, same with Palmer Luck, it seems, who all of us, you know, try to use science fiction to inspire us to build things in the world.

So going back to like the original question that you had, it was like, well, how do you feel? Does it keep you up at night to have somebody competing in the market or to doing some of these things?

Not at all, because if you do the math and you say, allright, well, if the GDP is at 126 trillionright now and AI can go from 3.5% annual growth to 4.5% annual growth, you know, in 50 years, we'll be at a one quadrillion dollar equivalent global economy.

There's a lot of space, you know, there's a lot of space for, you can have dozens of trillion dollar AI labs at that point, or you can have dozens of trillion dollar companies at that point, if not even more, and maybe even a company that is hundreds of trillions of dollars of market capitalization, I could imagine.

I think we're going to have that. You know, we just had, we talked on the show, Kovan and LinkedIn, and we got talking around like, actually, how many of these AI models are they going to be? Right now, you can argue there are maybe four or five, but we actually think maybe there's going to be 10 or 15 in the future.

Some are going to come from the US, some will come from China. Do you have any thoughts around that? Where do we think are we going to land in the number of models we're going to have?

I think it's very, very difficult to predict. I would say I would always bet on more as the amount of compute in the world expands and people find different niches. Today, it's the case that some

small handful of companies, because of the capital intensivity of training these models, you know, some small number of companies have gone about doing it. But I think that a lot of it might be due to the fact that there's just not as broad of an understanding about how to build these kinds of businesses.

Now, I do think that that market that they're in, because there's not necessarily this, like, I guess you'd say intrinsic network effect in the business, I do think that those types of businesses, ones that don't have some sort of intrinsic network effect, tend towards oligopolistic in their market structure.

I say this about Lambda in the compute industry. It's true about if you look at the hyperscale or data center industry, it's true if you look at the oil and gas industry, it's true if you look at these industries that don't have some sort of core network effect and are mostly economies of scale and mostly capital formation modes.

And most, you know, you look at the modes, the modes kind of oftentimes determine the market structure. I think that that suggests that there will be certainly more than one.

You know, so many of these opportunities also, it just comes from having great partners. You've been partners with NVIDIA for a long time. What do you think has made that partnership so enduring? And yeah, I'm sure you've met Jensen a few times.

First of all, just I'm really grateful for the support that NVIDIA has, you know, given Lambda and helped us grow up as a company. I think that the core fundamental thing is this. I was training on a GTX 580 workstation under my desk in 2012.

I built our first cloud or our first little cluster that saved the company using GTX 980 or, you know, GeForce 980s. And then our first version of the cloud, early, like sort of in 2016-ish, was 1080 TIs, which is like one of my favorite GPUs that's ever come out in the history of GPUs.

Amazing product. 11 gigabytes of VRAM. How far we've come. And I just was always really passionate about providing tools to other researchers and engineers. And it's very clear that that's also what Jensen and the entire NVIDIA team is passionate about, is enabling amazing things to happen in the industry with regards to breakthroughs in science, breakthroughs in technology, breakthroughs in AI and how that affects positively the world.

And so I think there's a big cultural alignment between the two companies. And that's kind of how I look at it. The other aspect about it that a lot of people I think don't maybe always appreciate is that, you know, building NVIDIA was really hard.

It was really hard. And building Lambda was really hard. It really wasn't, it was not a walk in the park. It was so much having to sort of eat shit for so long. And no one believes in you.

You get constant no's from investors. And I think that that's something that I really, really look up to in NVIDIA in terms of how difficult it's been to build that company and how hard one has to work, how much commitment and dedication you need to have to building a company.

And so I think that that's like part of the reason what I feel like that partnership and that there's some shared cultural values between the companies.

It's so true. I forget when NVIDIA launched, 1980s, 1990s, but it took them a long time,right? I mean, I would say just the last five, ten years is probably when they've become.

90s was really nice, yeah.

Yeah. And really, maybe in the last five years, they've become headline news every single day.

I mean, when I was first building my first computer, I was choosing between an AMD and an NVIDIA. And I ended up getting the, I ended up getting an AMD CPU and then an NVIDIA GPU, which was avant-garde at the time because, you know, everything was Intel at the time.

Yeah, I mean, just to talk about pivots,right? And I think this is another interesting thing of like, I think the most talented founders, if you look at any of the really big companies, NVIDIA, Apple, Intel, Microsoft, all of them started with a product, introduced a new product that completely took over Amazon's another great one, completely took over.

And then, you know, maybe have done that a number of times. So Apple started with the Apple I, Woodbox, Apple II, huge success, Macintosh, but then eventually they introduced the iPod. And the iPod became the vast majority of the revenue.

Then they introduced the iPhone, and then that became the vast majority of the revenue. And talk about pivoting,right, you know, into totally new spaces and refounding the company. Microsoft started with Microsoft Basic for the Altair. Eventually, they moved into disk operating system to MS-DOS, then eventually Windows, and then eventually an entire, you know, cloud, Azure, and, you know, they've got so many different products, you know, not even to mention the Office suite and everything.

And I guess the point I'm trying to make here is that like NVIDIA has, you know, started off making chips for GPUs and eventually realized that this AI thing that they were enabling and scientific compute from CUDA was the thing that's going to take over and be the biggest part of the business.

And it's just really amazing to see if you actually look at the history of all these companies, how much pivoting is involved and how much the company completely changes over time.

And there's a great thing about building a company,right? When you say Windows, you know, ages us a little bit. I remember, you know, going back at home when my mom and dad's computer, the white thing, you see like the little Windows sign crop up and, you know, fast forward 20, 25 years, boy, this like tech world is very different.

It's very different. Yeah, it actually is quite, quite interesting to think, you know, kind of how much things have changed from, you know, booting up your Windows, hearing the Windows startup sound of Windows 95, at least for me, that was my first one, and to where we are today.

And it makes you realize, I mean,

I've had a talk about this, about neural operating systems and kind of where I think things are going in the future. And I think that that's definitely the sort of direction is that you're kind of going to get this very squishy and kind of probably will be very creatively interactive software that we interact with in the future.

I think it's going to take some time though, just because it's so compute intensive to do something like that today. But yeah, it'll be interesting to see what things are like in, I guess, if Windows 95 was, gosh, 30 years ago.

Wow. Times have changed. Our kids' generation in the future, what that world looks like is obviously going to be so different. And GPT only launched four years ago, five years ago. And how far we've progressed in that time is just like absolutely bonkers.

I think about that a lot. I always think about that from the context of what kind of life or what kind of society or what kind of technology my kids are going to be interacting with. And it's one of those things where the world is always just such an exciting, it's such an exciting and new place.

And humanity always figures out, like it just becomes the norm for them,right? You know, they don't even, they take things like ChatGPT for granted.

And, you know, when you see now YC companies or founders who are 18, 19 years old, they're building, you know, $1,500,000, $100,000,000 annual current revenue companies. I mean, this never happened four years ago,right? So let's see what happens.

It certainly was less common,right?

Much less common.

The democratization of writing software is really interesting. That's actually one thing I really think about, just even in the more medium-term kind of things, which is what you realize, okay, is that there's this entire set of people who you work with at your job today.

They're your normal colleagues that, you know, they could be in the legal department, they could be in finance, they could be, and there's this like this little sort of like cadre of people who are like, oh, hey, check this out, you know, look what I built with Claude.

And you're like, oh, you know, as somebody who, you know, studied computer science in college and had exposure to programming in a structured, I guess you'd say, environment, you're like, oh, these are the people who are like all meant to be programmers, but like somehow just ended up getting majoring in something different or, you know what I mean?

And it's really cool to see because you can actually imagine a future very, you know, basically it's already here today where it's not really so much about did you study computer science or, you know, did you teach yourself how to program, but can you think in a very structured way?

And were you kind of that person who was sort of meant to be a software engineer, but like didn't end up falling down that career path? And I think it's really interesting to see that set of people.

I agree. It's going to be super exciting to see what happens in the next five or ten years. I'm really excited also just for Lambda. You know, I think it's going to be really exciting over the next few years and can't wait to see where you guys go next.

Thanks. Me too. Yeah. I can only imagine what it's going to be like looking back. I mean, certainly if I was, if I flash back ten years previous, I can tell you that I would be very, very surprised and very proud of where we've gotten to.

Amazing. Stephen, thank you so much for coming on the show. I really appreciate you and yeah, we can't wait to see where things go next.

Ollie, thank you so much for having me. I really appreciate it.