Baseten CEO Tuhin Srivastava on the AI Inference Crunch, Custom Models, and Building the Inference Cloud

0:05Hi listeners. Today, Elad and I are here with Tuhin Srivastava, the founder and CEO of Base10, the AI Inference Cloud. We're here to talk about capacity constraints for AI compute, why inference is the last market, how the workload is changing, the open source and perhaps multi-chip future, and what 30x scale in a year looks like. Tuhin, welcome back. Hi. Good to see you. Thanks for having me. All right. You are in one of the craziest markets, AI inference. It's very important. There's a lot going on. You guys have grown 30x over the last year. And I think I can say you're expecting to do more than a billion dollars in revenue this year.

0:47What's going on? Tell us about scale. Yeah. No, it's been nuts. I think what's happened over the last, honestly, 24 months, but it just kind of keeps getting bigger and bigger, is that I think everyone is realizing that you can put AI everywhere. You have all these great options available from closed source to open source models. The open source models have crossed some sort of chasm in terms of their baseline capability. And then I think RL techniques and post-training for specialized models has become mainstream enough. and there's enough examples of it working, the customers realizing they can kind of own their inference more and more.

1:34And what that's meant for us is more the long tail models coming true, customers in-house seeing a lot of that intelligence themselves. And as the application layer just gets bigger and bigger and bigger, and that's growing, we are just someone index on that and we've been around to be able to collect the demand. There's an existential question in here that I think everybody is continually asking of, does the independent application layer get to exist at all versus the labs? Like, how do you, you have to believe this. Why do you believe it? Yeah, look, I think it'd be, it'd be a sad thing if it didn't exist in general.

2:11And I think that's like my, but you know, sadness is fine. I said all the time. Sadness is fine. But that's not the reason why I think the application layer will exist. I think the application layer will exist for a number of reasons. One is because, you know, I think this idea that what is valuable to a company is, you know, the user signal that they can gather, that only they can gather. And to the extent that that is encoded in a model, I think a lot of their business will be at risk. But to the extent that it is encoded in workflows, that is where they will be able to develop mode. So a good example of that is, say, a company like Abridge, where the clinicians edit off the notes and what they do with those notes after the fact, and the thing that happens inside the EMR three steps down, that becomes a workflow that only...

3:11Can you explain what Abridge does? Sorry, Abridge is an ambient scribe that is used by physicians in almost all hospitals in the US. I think a large investor, great company, great team, great product. And they've basically got this very, very deep integration into hospitals, into clinician workflows. and my argument would be here is that actually it's very, very hard for a frontier model company to go to EWOA because they just don't have access to that user signal. And what will happen over time is the folks who have access to that user signal can start to post-train models on that reward signal and start to get long horizon agentic models running that.

4:01And I think to the extent that that is possible and that signal is differentiated and unique and is somewhat rare to get access to, there will be an application layer. And I think, you know, support companies is another example of that where, you know, a support task isn't one-shotted. Usually at a company like Base 10, when a ticket comes in, there's like, what, like one, two, 10, 20 actions that get taken. And that is where, you know, someone can develop a specialized model. So there's almost two versions of this then. There's new companies like Abridge or Decagon or some of these other things that you mentioned that are doing these new types of applications that are using AI and they sell it to customers.

4:44The other is enterprises building things in-house or building their own models. What proportion of the market today do you think is these new application companies versus enterprises just adopting AI? And how do you think that looks in a couple of years. Yeah, I think that's, I think you asked me the same question two years ago. I had to be repetitive. It is crazy. At least I'm consistent. The answer is just that it's crazy that the answer is still, I think, I think if you look by inference count, it'd be 99 % the full. Yeah. And that kind of represents the scope of the opportunity here is that the majority of the market hasn't come online and added AI into the world.

5:26Yeah, most of enterprise adoption is well ahead of us. And I think that's one of the very exciting things about AI. Yeah. There's just so much still to come and people are underestimating that, I think. 100%. But what's cool is that we're seeing the transition happen, right? Before it was like, hey, are they using AI tools? I don't think that was immediately obvious two years ago. I think that's obvious now. But yes, they are. Are they using closed source model APIs? I think they're starting to get there. And then once you do that and then you kind of see what is possible, then comes the whole custom model adoption.

5:54I think that is all that is ahead of us. today. So if the majority of your customer base today is, as you described, the former application companies, AI natives, the fast-growing, I mean, some of them are at considerable scale now, like the abridged cursor, open evidences of the world, what do they teach you? What does that push the company to do? How do you think about serving them versus evolving for the enterprise? Yeah. I think firstly, you just learn a lot by building with the company's greatest scale, doing the most interesting things. We think of it two ways. I think there's the most obvious way, which is just build for the highest scale.

6:41the customers that will push you the most from technologically and everything kind of will fall into play. I think the Stripe evolution as a company showed that was like Stripe now like serves so many enterprises, but 12 years ago, that wasn't the case, but they just built for the frontier and kind of went with them. And the second way we think about this is to just think about building for companies that are serving enterprises. So yes, we don't serve the enterprises, but our customers serve enterprises. Abridge serves every device, OpenEvnance, Decagon, all these, Ryter, Gamma, all these companies serve enterprises en masse.

7:19And what we actually get is like a translation of the requirements from them, which is like, you know, they're like, hey, we need this data retention. We need this web models need to be deployed. This is the types of GPUs or the latencies they're okay with. This is the model requirements from like a transparency perspective that they care about. And so I think that is actually the more nuanced answer is that if you listen to what their needs are, we actually get a full translation of what the enterprise was required. I would say that by serving companies like Abridge and Open Evidence, we're probably pretty well suited to go serve the healthcare system and latent health given that they are selling to them.

7:55How much of a shift are you seeing in terms of the types of open source models that are being used? And so I think we've seen an evolution where two, three years ago, I think the main thing was kind of Mistral and then a few other things. And then Meta kind of came along with Llama and then it kind of really shifted in terms of the misperformance models or of Chinese origin and different ways. Do you see that sort of mix reflected in terms of what's being used by our customers? Yeah, I think customers, at least the customers we are serving, are very, and these are like the fastest growing AI companies in the world that are very forward thinking.

8:25They want to use the best model. And they are optimizing. I think there is a subset of tasks, which I think is small today, where people really start to start with cost. But everyone comes from capability first, because that's really where the economic growth is being unlocked, where the value is being delivered. And then they optimize. And I think that's like actually been, you know, and so with that in mind, you know, you name, like you name it, everything from GPT OSS all the way to Moonshot, to DeepSeaks, to Canopy, Orpheus, which is like really good text-to-speech models. Customers generally want to use whatever's at the frontier.

9:11And I think the difference has just been, I think we have a lot more visibility into how to run these and how to run these really well. And secondly, that they're good now. There have been a number of different concerns raised about the use of Chinese models, in particular security, or is there something embedded in the models or Trojan horses or other things. A, do you think there's any real concern there? And B, people often talk about how there should be U.S. counterweights to this. from a geopolitical perspective, do you think that's something that's legitimate or something we should be worried about?

9:42Or how do you think about the sort of origins of these models versus their uses? Yeah, look, I think these models, firstly, are fantastic. They're amazing. We work with these teams. They're truly awesome. I'd say, look, I don't, it is hard for me. It's hard for me to see, and I could be wrong, but if I network bound these models that they're not magically going to be able to cross those network boundaries. And so data is there. And I've never seen any real evidence, except from some very early models that I think people picked up on very quickly that there is some agenda or bias built in this.

10:25I do think that to some extent is, I think there is importance to the US that we develop our own models. I think that that would be a massive loss if that there are five companies, you know, five different labs in China that are creating open source models. And we're struggling to get one set up. So it's necessary. I also think it's inevitable. And, you know, like the DeepSeek moment a year ago, I remember someone saying to me, and I thought it was like very well said, which is like, and the world's changed a lot. But they said, hey, you know, we should kind of just forget. that this is a Chinese model, we should just act like this came from meta and build with that in mind.

11:13It's like, you know, I think you're kind of missing the forest from the trees. Like there's two scenarios, right? Either America does not ever come up with good open source models. I think there's probably a fundamental problem there or we will get there and we need to be ready for that world. Yeah, that makes sense. It's interesting because, you know, like you, I think it's very important for the US to have a strong open source footprint here. At least for now, it looks like effectively the Chinese government is subsidizing at least a large subset of these models. And that subsidy or surplus is effectively just being passed on to US enterprises who are adopting these models.

11:46In other words, it's a way for the Chinese government to effectively subsidize US enterprise in an indirect manner. And I think that's a little bit lost right now. But it's always interesting to weigh that against some of the other concerns that are raised. I appreciate your comments on this. I think the concern also there just becomes It's like, what happened if we aren't able to, like, if it is fun, like, I think if you think of the economics here, which is DeepSeek by most, DeepSeek is a very good model. You know, like, and like, you can argue whether it's at the absolute frontier or not, but like, let's, let's go back three months and it's there.

12:22And so think about everything. We were doing a whole lot of things three months ago. And so let's just think about that. Well, you know, if you could run DeepSeq, probably 20 % of the cost of running open-handropic models in production with comparable, better latency, probably better reliability. If we don't have access to that intelligence in that form, I think it's just a massive loss. And as a country, we won't be able to innovate as fast because the cost of intelligence going down and control of intelligence, what we have seen just means more intelligence. Intelligence being embedded in more places.

12:56Yeah, an important note here that we didn't mention explicitly is that the state-of-the-art models, the ones that are most far ahead on the frontier, are actually still the closed-source, Anthropic, OpenAI, Google, etc. Yeah. What has been, actually, maybe you can just characterize workload a little bit, like how, of tokens being served on Base 10, like how many of them are from custom models of some kind versus like vanilla open-source today? It is all custom. It's basically... Okay. So like 95 % point. 95%. And I think that's really cool, to be honest. Look, we have two businesses. We have three businesses.

13:33We have three businesses right now. Should we help you count? No, no. So we have like dedicated inference, which is basically custom model inference. Your SLA is your SLA. Then we have shared inference, which is a shared inference endpoint, shared SLAs. And then we have a training business. I'd say 95 % of the tokens today are on the first business. And almost all of them, there's probably, yeah, for almost all of them, the customer is making some modifications to the model with their own data, specialized for the use case. And I think what's even more important is they might be compiling in different ways.

14:14No one is just running the vanilla open source weights. Like you might be customizing it for quality, but you mostly might be customizing it for performance. You made an acquisition of a research team a few months ago. You mentioned post-training customization. What was the rationale behind the acquisition? What is that team doing today? Yeah. So the rationale around the acquisition was, you know, we are infrastructure and product people. We are product people and now are really good infrastructure people. And we didn't have much of a research capability ourselves. And what we saw was the market moving heavily and heavily, like that we could accelerate the market itself with post-training resources, either productized or honestly, even just as resources for that market.

15:09So Paz was a company that was a base 10 customer. So they were post-training models and running them on base 10. And I think what they realized was that they would eventually need to become an inference company. And what we realized was like, hey, we really needed that expertise because it represents a way for us to get closer to the customer earlier and be able to support them all. and just made sense as a, like pairing them together. And just as I said in the opening statement here, which is, you know, as more and more post-trained models have come up, we've realized that the demand for people to either for software loops to do post-training or for post-training expertise is very high.

16:03And we're really, really investing in that. There are also a bunch of Australians, And, you know, I like to think that we had a bit of alpha there. But yeah, that's been fantastic. They're working with all sorts of customers. And it's also very interesting when you start, you know, we were doing a lot of research on the performance side and less so on the post-training side. It's interesting as we've started to do a lot more research on the post-training side, you start to see how linked inference and post-training are. And even when you think about stuff like quantization and when you should do that and how you train the model affects how you need to quantize for inference and how paired these problems are has become very apparent.

16:53And more and more, the post-training inference are both sides of the same problem. So because inference ideally will get more post-training where inference creates data, you do evals, you can now post-train on that reward function that you found with those evals and hopefully just set up that entire look. Plenty of folks from Ant and OpenAI, Sam, Greg, et cetera, have said in recent months that inference is super strategic, inference talent is strategic, capacity is strategic. So between that and post-training, these are very difficult to gather capabilities. I imagine that lots of your customers go to you guys for advice on how to do this progression of moving to custom models.

17:39Like, what do you tell people about the lifecycle and when they should invest in that? Yeah, I think it's, hey, go find, go prove to yourself with the best in class model that you have something worth optimizing. And I think, you know, a lot of, you know, if a customer comes to us, was that meme, which was like, it was like two years ago. It feels like no GPUs, pre-product market fit. It's like no post-training pre-product market fit is what I'd say. Yeah, yeah, yeah. It's what I'd say. So people that you're working with here are very at scale first. Yeah, they have a user signal that they know how to optimize.

18:13And they've shown that they can, you know, they can serve customer value and that they have something special around that value. And once you have that value, it's like, okay, now how can I do that better, faster and cheaper? With the idea being that, hey, if you need to be very good at customer support, you maybe don't need to be that good at coding and that a specialized model might be a better fit for that problem and you can do it better, faster, cheaper. What about the capacity side? You started with unifying capacity across all the clouds and new clouds. How do you think about this when everybody keeps talking about a supply crunch and a multi-year supply crunch?

18:47I think there's so much narrative around the supply crunch. And no matter, as much as we hear about it, I don't think people realize how bad it really is. There is very, very little Slack compute available. We run pretty large clusters ourselves, and we run them in uncomfortably high utilization. When I'm saying we're mid-90s utilization most of the time.

19:21We sit in 18 different clouds now. We have 90 clusters around the world across 18 different clouds. And like, you know, initially we started, we built this technology to be able to like kind of create one runtime fabric that spans all these different clouds and try to abstract that away from our customers as a way to think about reliability, latency, failover, all these things that we think are going to be very important for very mission critical use cases. that same technology, like just our ability to get compute wherever humanly possible has been really, really helpful in our ability to get supply.

19:58And what I mean by that is we can be introduced to a new provider in a different country and have it up and running with the whole base 10 inference stack. As part of the fabric. Half a fabric in half a day, maybe less. and that gives us enormous flexibility. Even for us, it is hard for us to grow. We have a, I think it's, yeah. We have a 4 p.m. standing meeting for the company where we basically like, how do we manage capacity for the demand right now? I think the second part, which people don't really, the second part that people don't really understand is that there are also a lot of suppliers right now that it's kind of grifty.

20:56You know, like I think, you know, they haven't run data centers before. You know, they don't understand SLAs, especially for inference. And so like, you know, even when there is capacity available, there's a lot of, like there's probably we run a lot more and we have redundancy so it's fine but if you you know there's probably like a dozen good like clouds and I'd probably like put like three or four of them in like the the gold tier and I think that just means that like supply like not only are we supply crunch we're supplier and operationally crunched onto people who can who can run these data centers as well.

21:39How far ahead can you actually buy capacity right now? In other words, is there any slack in the market if you buy two years ahead or five years ahead? You mean like contract length or actually like, hey, I want this in January 28? Either one, yeah. I mean, it's more the I want this in January 28 or at least I have some visibility into my future supply. Yeah. You could buy that, but you've got to also remember how quickly the market is moving. And that gets balanced somewhat off the fact that the H100 is such a great chip.

22:17And it's crazy. It's four years, four and a half years old. The price is going up still. Maybe it has a useful life for nine years. So that's good. But at the same time, yes, you can do that. But you're making a lot of bets as part of that. And then in terms of, I think that's the big thing that's changed over the last six months is that the term length that people want has just gone up. So if you wanted 1 ,024 B200s, which is, you know, from a good cloud, right now you're not getting that less than a three to five year contract right now with probably a 20 to 30 % TCV prepay. um so like actually what becomes important when acquiring capacity um is you need to have enough demand to supply it um to serve but then you also need like a low cost of capital um which is which is actually changing the dynamic pretty significantly does that does that impact how you think about going public as a company because arguably yeah i think you'd go sooner yeah exactly yeah i think you need like i i think the and i think there is demand for that um but i think you the pool, the, it also, you know, one of our, one of the, one of the realizations that we had recently and we're, we're software people.

23:42Um, and so we don't, we don't think like this all the time is that, you know, our business has like very interesting work and capital, um, requirements, like, you know, um, and, and, and I think, you know, even, and that as a result of that, it has very interesting financing, um, requirements. And we're not, at least right now we're not even going down to the down to the dirt there's also things you could do in terms of debt or other structures that yeah yeah and yeah i've learned a lot about debt yeah recently given the uh supply crunch uh inference being one of you know the top couple markets you're going after you have um plenty of people who understand this problem and therefore you know some competition how do you uh how do you think about like what are the factors that create a dominant player here or a winning player?

24:30Is it, as you mentioned, cost of capital? Is it access to supply? Is it software? Is it demand? Yeah. Just being excellent at everything. Yeah, it's... Look, I think what's so interesting about inference is... Is it operations? I guess it's actually cloud. Yeah, I think so. Yeah, I think like GPUs as a service is not sticky. I think that's been seen. Like customers generally just see that as commodity. Inference with the software layer included is incredibly sticky. You know, like just like, you know, none of our top 30 customers have ever churned. You know, we're talking like 400 % annual NDR around our business.

25:12And so it's like very, it's very, very sticky. So I think that software layer is very important. The optimist in me is like, oh, there's so much value in the software. And we will build the best software layer for inference that exists. I think, as I think is becoming clear now, access to inference compute is a strategic advantage. And I think that is the strategy that even the labs are going after, which is if we have all the compute, good luck running inference. Yeah, yeah. In a world of constraint compute, the number one thing to own is compute. And so just owning it in and of itself is an asset.

25:51And I think people underappreciate that. Yeah, you can't make a good hot chocolate without milk. Unless you're a vegan. Unless you're a vegan. No one wants a vegan inference. Well, I've got to ask you, people might want alternative milk, right? So, okay, like when you, the H100 is a great chip. People want a B200, they want a GB200. They want, of course, tons and tons of NVIDIA. When you think about making a bet, you know, several years in the future, Do you believe that there is a multi-chip world? What do you think happens from a compute perspective on the chip side? Yeah. I think diversification everywhere is the same way I want to water many models.

26:38We want to water most things. You'd be sad if it didn't happen. Yeah. And I think everyone would be sad. I will say to some extent, which is, yeah, and I think there will be inference-specific chips. I think you have like decode-specific chips, I think. And we're looking at - And NVIDIA said this too. Yeah, yeah. I mean, that was a whole Grok LPU thing. It's like, you know, I think that is very straightforward and makes sense. I think people really, really, really underestimate supply chain stuff with NVIDIA. Like how good they are at that. CUDA, how good CUDA is. the developer ecosystem around it.

27:16And, you know, we, the ability, like, to me, like one of the most important things as an infrastructure company in this moment is how fast you can move. And you can move fastest with NVIDIA today. And I think that is the reality. And like, it just like, given the scale that they operate at, given the scale that they operate at, it's hard to see. it's hard to see the I'm not saying it won't happen like the short term like in the next couple years how anyone's going to be able to compete for that. Especially with, you know, so much of the other players like what you need to be able to compete here is the ecosystem to form around you.

27:57And if you tie up all your supply with one buyer which, you know, a bunch of the other chip providers have done it's actually hard for that ecosystem to form. You know, like if you think about if you're a big lab and you have a proprietary deal with one chip type where you get 90 % of the supply, it's actually in your best interest to make sure you get 95 % of the supply and everything that's built for you and no one else could ever use it. When you think about reacting to the market, what do you think is happening with the actual workloads that you have to go invest in, right? Like obviously code agents and long horizon agents over time have become a big deal.