Building Voice AI For The World: The Cartesia Story With Karan Goel | Elevation Capital Podcast

Is it fair to say that India is one of those markets which exposes the weaknesses of speech models the most as compared to any other market? Like, there's constant code-switching in Indian languages, right?

If you can be successful building your product for the Indian developer and the scale that they're operating at, I think you can be very successful anywhere else in the world. In audio, it's all about keeping people engaged. If you can keep the user engaged, you're more likely to have a successful call. If you're pausing too long, it disengages the user. Whether you're training language models or whether you're training genomics or audio models or video models, all of it boils down to how do you build these architectures and then how you train them on data and compute to build the capabilities. Albert noticed that, you know, he realized sort of that transformers aren't going to solve everything in AI. And one of the big bottlenecks that's going to come is you know, dealing with context.

Most of the other speech models today are built on transformer-style architecture. Cartesia is explicitly built on state-space models. For an end user, for an end customer, what changes?

No customer of ours should really care how we build our models. BFSI is ripping in India. Yeah. In terms of, you know, how much stuff is happening on agents and voice. In 5 years, to like really get to a place where I'm only selling folks one API and it's just, you know, this magic API that you can use like a human.

It feels, at least from my vantage point, that India is not an afterthought for Cartesia. It feels like you've been very explicit in making a bet here. But I would love to hear from you what's the thinking behind India and for Cartesia. Is that the camera that's recording?

Okay.

Are we all good?

I mean, that covers everything. Everything.

Exactly. It's all the same. We are like blending in somehow.

Oh, we are coordinated.

We are coordinated. That's good.

What part of my body is not visible?

So that you can chill. Today we have with us Karan Goel, founder of Cartesia, who's obsessed about building one of the fastest, most realistic, and real-time speech models in the world. What I'm really excited about, Karan, is that your story starts here in India, goes deep into AI research, and then comes back out again to to solve some very real-world big use cases. Thank you for doing this, I'm very excited to have you.

Thanks so much for having me.

Awesome, and clearly we're twinning and then it's all blending in with the same colors. Cool, Karan Goel, so let's start from the beginning, right, you grew up in India in a family with generations of manufacturing business capabilities as I might say. But I know I heard it or read it somewhere where you said that Growing up you weren't one of those kids who exactly knew whether I want to be an entrepreneur or a scientist or a lawyer or doctor or something else, right? But walk us through those stages of going to IIT Delhi then CMU then Stanford, what was unfolding in your parts of life, you know, some parts of your life?

Yeah, I mean I think I grew up pretty much in a business engineering family. So my folks all were trained technically. All the way back to my great-great-great-grandfather.

4 greats.

I guess 3 greats and a grandfather.

Okay.

So I think like, you know, it's sort of like in our blood to build and do engineering and technical work. So, you know, I was reasonably good at maths growing up, not as good as some of the folks I met later in life, but reasonably good. And so I ended up like obviously like pursuing engineering just like most folks in India do. I think that I was excited to obviously go to one of the IITs, that's obviously what most Indian kids would dream of.

Which you just saw recently again.

So I think IIT Delhi obviously was a great experience, I met so many smart people, it is an insane concentration of talent and so it was kind of the place where I discovered AI research. I actually started working on AI research because Professor Mossum, who had come back from the US, he's still at IIT Delhi and he does a lot of work in AI and he's a great, great, great teacher and researcher. He was the one who got me and actually a lot of folks interested in AI research because he was like full of energy, just come back from UW after doing his PhD. He did a lot of, you know, I think great work in training up a generation of Indian students. And not just me, you know, I have a lot of friends who went to IIT Delhi who ended up doing graduate school in the US because of him. So I think that's where I got interested. So that was like 2014. And then, and I think that, you know, AI is one of those things where like you can get inspired pretty quickly because it's just kind of science fiction and like, you know, you're building things that don't seem like they should be real. So I do think that was kind of super exciting for me. I still, even when I was graduating from IIT Delhi, wasn't quite sure what I wanted to do with my life, but I knew I wanted to continue doing research because it was interesting and fun. So I ended up applying to grad school and went to CMU. CMU was You know, I think quite different from IIT Delhi.

Of course.

Just like culturally and also in terms of the focus on, you know, just how much AI research is happening, right? You know, IIT Delhi, I think, you know, is a great school, but there's a lot less in terms of like research volume. But at CMU, it's just like there's the School of Computer Science, you know, there's a department for pretty much everything. I was in the department of machine learning. There's a department for robotics and so on and so forth. So I think there's an incredible amount happening and you learn so much about what problems matter in AI because you're always talking to folks. And so that is what actually got me more inspired to try to do a PhD. I think the other part of it was that I didn't really know if I wanted to work. I was— I feel like every decision to go to grad school is a decision to delay doing real work.

This is your first real job, right?

That is true. I had never worked at a company full-time before. I had interned but not full-time, so I actually did, unlike a lot of the 20-year-olds that drop out, I did, you know, the full cycle of every school you could go to and all the programs you could, you could go to. So I ended up spending about 5 years at Stanford doing my PhD, and that was where I got into a lot of new research that became Cartesia. But it was also where I discovered you know, more of the frontier of work because we were really exploring things that were super new, right? And so obviously happy to talk more about that, but I think that's where I really realized that you could make a career out of doing AI and like spend more time on problems in AI, and it's super interesting as a field to work on. So yeah.

No, very interesting and such an interesting journey that you never thought that AI could be this interesting and this could be like a lifelong thing you would do. You were just pursuing your passion, hopefully, right?

No, I really love what I do. I did a 5-year PhD. At some point, 3 and a half years in, my advisor was like, do you want to graduate? And I was like, absolutely not.

Oh wow.

I just wanted to do more of it.

Amazing. And see, Karan, you spent so much time at Stanford in academia, right? And you were really rethinking how to build sequence models, like really thinking from scratch about model architectures. I'm very curious to understand at what point in time what were some of those things that you started noticing that felt unsolved and then you wanted to go after it and hopefully that's what sort of transpired into Cartesia today.

Yeah, I mean I would say like I have to give a lot of credit to my co-founder Albert, he was the one who really was working on a lot of this research. You know, I think I'm more of a modest contributor to this area compared to the work that he's done. You know, he's kind of more of a monumental figure in the field of AI. I would say that for me, you know, I've always been interested in, especially from an AI lens, like how do you build things that are real, right? And I think part of what's exciting about AI research is the work that you can do to build real models, and a lot of that boils down to for any modality that you can imagine, you know, whether you're training like language models or whether you're training, you know, genomics or audio models or video models, all of it boils down to how do you build these architectures and then how you train them on data and compute to build the capabilities. So the sequence modeling is, which is an area of, you know, machine learning, that area has a lot of research going on. And I think, you know, transformer architectures obviously came out and they were a workhorse for machine learning, and they've kind of really taken off in the last 10 years. But Albert noticed that, you know, he realized sort of that transformers aren't going to solve everything in AI, and one of the big bottlenecks that's going to come is, you know, dealing with context. You know, how do you actually build a machine that can have infinite context? Very simple question. The answers are hard because, you know, if you're using transformers, you have to have a model that's essentially needing to capture all of this context in its, let's say, prompt window or context window. And it's not able to do compression. And compression is kind of something that you're used to as a human. Like you interact with the world, you spend 85 years, you're absorbing all this data, you know, terabytes of it. And so how are you actually able to inference yourself, do things at such low latency so quickly without ever slowing down and there's so much instinct and intuition and all of these things. So how does that kind of intelligence manifest? That's kind of the core question actually. And so I think when we talk about human intelligence, AGI, getting closer to that, you kind of have to ask the question like why is it that humans can do so many things that AI cannot and what is it about the architecture that makes that possible versus not. That is really the key question.

See, that's a very important one and I wanted to get back to to it? Because in a practical way, and getting a little technical here, but most of the other speech models today are built on transformer-style architecture. Cartesia is explicitly built on state-space models. For an end user, for an end customer, what changes?

Yeah, so one thing I say always is, um, you know, no customer of ours should really care how we build our models. At the end of the day, you care about a few things. If you're adopting voice or audio models. One of those things is latency. And latency is sort of like one of those things where it's not just about how fast is this model when I demo it, it's about how fast is this model when I'm scaling an application on top of Cartesia's model and I'm using it on a million calls a day. Like, that's the question. That's where the architectures start to matter because you're really designing for something that's going to run much, much faster at scale. So for example, one of the things that happens when you do inference is when you're doing inference for batch size 1 versus full throughput inference, your time to first token, time to first byte is actually going to change significantly. And the architecture is actually defining how that change happens. Like the graph of like batch size versus latency that you're getting is actually going to depend on the architecture that you design. So these are super consequential decisions that ultimately impact what you're seeing in production. That's one of the places, you know, what we did at Cartesia was build the fastest models in the world for audio. And, you know, these models are fast because we're not just thinking about it from a, hey, how do I serve it closer to the user? It's also about how do I design from the ground up so that every single part of the pipeline that you're building has been optimized to all hell, basically.

Correct. And see, I again was reading some of our docs and, you know, you don't only focus on average latency but tail latency. And these are very small tiny details that really matter in voice. Like again for an end customer why does tail latency matter a lot, like give me an example.

Yeah, yeah, I mean I think that if you look at enterprises and you look at let's say you're doing a use case where you're trying to help someone through a technical issue. And so you're actually maybe you're a large business that's selling them hardware and they need help. Fixing a router, right? I think a lot of those cases like customer satisfaction, being able to actually help your customers be successful, it really matters that every interaction is done right because there's a reputational hit, in many cases there's a hit to revenue, in many cases there's a hit to how people actually want to adopt that product. So I think for us it's really important that when folks are doing calls every single call is done right. And so tail latency matters a lot. If you look at average latencies, average latencies or median latencies can look pretty good, right? Because if half of your requests are 600 milliseconds, or let's say 40% are 600 milliseconds, 60% are 200 milliseconds, you're gonna have pretty good median. But you do care about the experience degrading for a bunch of customers. So that's super important in that space. I would say that delivering that at scale is super hard. And one of the things, if you go on our Slack, which you can't, but I can promise this is what's happening. Our Slack channels, there's like, you always see folks putting screenshots of Datadog charts. And then you'll have like 200 message Slack threads where people are discussing what's going on. Because a lot of these Datadog charts that we obsess about, are cases where you actually see a latency spike for a customer and then you're trying to go deeper to understand like despite all the systems that we've built why is it that that customer is still seeing some spike. So that's the level of obsession that you really need to get into where it's down to the level of like a single request and why it's not doing the right thing. That's what we focus on. So I think that infrastructure to me it's super valuable for the world because real-time delivery of models which is what we're doing effectively is actually one of the biggest areas that I think is going to grow. Because people want intelligence faster. You want it to look at your data faster, you want it to talk to it faster, all of that. So I think these types of systems are super interesting just independent of what we do. And so we spend a lot of time trying to understand deeply like how do you actually make them great.

Yeah, yeah, and see, to your point, right, in voice it really matters because I don't even want to wait even for a— like a second is very big because we're talking about milliseconds here versus in text, video I'm okay for it to render a little bit, take its time, think, I actually want it to think and then come.

In text it's actually interesting because I've I've heard from folks that sometimes you actually want to be slower, not faster, because it feels like you thought through the answer better. So there's weird idiosyncratic effects that happen in different modalities. In audio, it's all about keeping people engaged. If you can keep the user engaged, you're more likely to have a successful call. If you're pausing too long, it disengages the user. So having that is super important. I would say that's one of the things that we are best in class in and we really care about.

Very, very interesting. I think I also wanted to talk about something very specific that we've built at Cartesia, the streaming context model, which is essentially not converting just text to speech, but making sure that our rhythm, our prosody, etc., are also managed across partial inputs. Why is it so hard?

Yeah, I think that generally when you think about it, like, you know, what is text to speech at the end of the day? It's a way for you to talk to the person on the other end of the line. And so the question is, how can I build you a model that allows you to give the best prosody and response as possible? So for example, if I am a customer and I'm getting angrier, you need to be able to match my tone. I need to be able to match the way I'm talking. And that's something that you can only do when you're capturing a lot more of that context. So this is something we're super interested in. And we've been working pretty hard on. And in general, I think one of the things we see with developers that use our APIs is that they're using these things called continuations in the API, which is when you're trying to generate a lot of speech, you need to be able to continue across multiple model generations and have continuity of context there. This is some of the stuff that we do behind the scenes to give them a pretty smooth experience when they're generating responses and so on.

So TCR, Karan, didn't just stop at text-to-speech, but with obviously InkWhisperer, we are getting into, you know, not just text-to-speech but speech-to-text. What is the thinking there? Where are we going with this?

Yeah, I mean, I think our thesis is pretty simple. We are a research company, we are interested in end-to-end models, we want to be able to build models that mimic the capabilities of human intelligence, we want to get closer to that. A lot of the work we do on architectures is really about how do you push closer to that vision for continuously running context-consuming machine that can take action, understand what you want from it, and do this over a lifetime. This is what we want to build. And as a product, we want to be able to give this to companies and have them embed this almost like a teammate or an employee. And we think of voice as sort of our entry point application into that. So when we were building this product, and this is a multi-year thing for us, right, when we were building this product, we wanted to have or peel off parts that we thought were super important for this type of product that we could also help developers solve for. So text-to-speech, I think, is an important component of voice systems that we felt like could get a lot better, especially for developers of conversational systems, which is where we focus. And so part of what we've been doing is expanding the set of products we offer. So going from text-to-speech, which is one model that's doing speech synthesis, to the listening side. There's only 3 components, there's talking, thinking, listening. And so you'll see us expand into all of these areas and the goal is to sort of help people have the best end-to-end systems for that where you can configure them easily as if you're a developer or business it just becomes super easy to kind of close your eyes and use this and it just works. That's kind of where we're going with this.

And do we think that we will at some point in time pick a battle on whether we want to be the best text-to-speech or the best speech-to-text or something in between?

I think that's an interesting question. I would say I think we're more interested in the overall goal for this, which is less about the actual like this API or that API, it's more about what are people actually trying to do at the end of the day. And what they're trying to do at the end of the day is let's say you're doing support, you're actually trying to just automate how you deliver that support to customers. And that means what you actually want is can I have a great conversation? Can I capture all of the context that is necessary to make sure that the conversation is addressing what the user wants? Can I take the actions that are needed inside my systems? All of those pieces. And I think our sort of thinking is over time a lot of this stuff needs to be done by a single model. It needs to be consolidated into a model that's capable of kind of managing all that context. Today if you look at how people build those systems it's actually a lot of engineering. There's a tremendous amount of engineering being done by a lot of companies that are focused on building agents, right? And so we don't think that's necessarily the right place to go in the future, especially as researchers. We feel like actually the right place to be is you can plug in one API and that API just does all of the things that are required to run the conversation well. So that's what the best product for us is. It's like, as long as text-to-speech is the right way for people to build those systems, until we try to replace that, that's the right product. But we think for example text-to-speech as a product maybe in a couple of years may not even exist. So we're more interested in how do you solve this problem end-to-end as a developer.

Very fair, very fair. I want to switch gears now, Karan, and come back to India, why we're here, why you're here. It feels at least from my vantage point that India is not an afterthought for Cartesia, it feels like you've been very explicit in making a bet here but I would love to hear from you what's the thinking behind India and for Cartesia.

Yeah, I think we feel like especially as a company that's based in San Francisco we felt like India as a place where I grew up firstly which is obviously like helping us a lot in terms of actually wanting to spend more time in India but also as like a market is super interesting for Voice. I think it's a huge population center, there's so many folks that are on mobile right, and connected to the internet. There's a cultural aspect of actually using voice to be able to interact with other people, like sending voice notes, for example, right. And I think that's, you know, maybe not common in every part of the world. There's also the fact that it's a highly multilingual country, you know, you can't cover more than 50% of the population with any language really. And so there's a lot of diversity in terms of connecting people to each other when it comes to languages. So I think overall as a market it's sort of super interesting, right, like it's clear that the amount of interaction that's going to happen with AI systems through voice in India is going to be pretty large. It's going to take some time for those systems to get built, deployed, used, people to get used to it, you know, the comfort that people have talking to AI is going to change over the next 10 years. So there's a lot to do but I think for us what feels like the right time to kind of be investing and spending time here. So one of the things we're doing is obviously building a team out of Bangalore and like hiring folks here who are going to work on core AI models and like really focus on like helping us deliver those models to customers, which is, you know, I think it's an interesting sort of company because we're so research-focused that I think we're excited to bring that kind of flavor of work to folks in Bangalore as well.

And I know a lot of people are excited.

Yeah, I mean, I think it's going to be super amazing, and I think, you know, it's one of those things where I wish I had that opportunity to work at a company like that when I was in undergrad. So I think that's kind of our way of maybe passing on that to other folks. But I think it's going to be super exciting for us. The other piece is, of course, we think that like there's so much happening in voice in India. Like we've seen so much organic adoption from Indian customers sitting in the US, so we felt like it makes a lot of sense for us to be here and have a ground game. So those are all of the things that we've been thinking a lot about when it comes to India. I think I'm sure there's going to be— I think especially working in voice, I think there's going to be so many use cases that are hyperlocal that you don't see in any part of the world except here that just get built out because people are super creative and they just figure out ways to build.

Yes, there's so much that we see, right, like the focus on code switching, like there's constant code switching in Indian languages, right, the precision that you need to have in numbers, in accounts, in order IDs, and it should be very specific. I'm sure it's a hard market to crack.

I think that, you know, for us, like, one of the things we care a lot about is how do you hyper-localize your product to a market and make sure that it's authentic to what people need there. I think here in India, like, obviously take one example like Hindi and English, right? English is a super common way to speak, and, you know, being able to code-switch seamlessly is pretty easy for folks. And that's how they expect to talk to others. And so how do you capture all of that nuance inside voice models so that it sounds authentic? I think that's a really interesting problem and we've been working pretty hard on that. I think we've gotten super good reviews there actually from customers who tell us it's the best English they've ever heard. So I think that's been an amazing sort of like step change that we've been able to make. And now we're trying to bring that to all the Indic languages. So we kind of actually already expanded to 9. And now what we want to do is get the quality up to peak. So the way to think about it is if you look at English— English is always the highest quality language, right? Any— take text models, take like audio models, it's always the one that's at ceiling. And so I think for us the priority is like how do we actually apply a lot of force on multilingual to get the quality to where it needs to be. Compared to English. That's what we did for Hinglish, and that's what we're trying to do for all these other languages. So I think that's going to be a big way that I think it gets adopted as well.

Yeah, one thing I'm curious about though is how do you get the data, right? Like all the data that you needed to build voice models for, let's say, the US versus the one that you need to build all these Indic languages that we're talking about. How are you going about it?

So we tend to end up both going out and actually partnering with folks, collecting data on our own. And also finding folks on the ground that are giving us recordings and we're able to use that. That is kind of how you actually acquire the data to be able to do this. And obviously we have also like royalties that we pay out to actors and so on and so forth to be able to use their voices. So I think there's a whole sort of like data engine underneath all of this that's helping us both pre-train and post-train all these models so that we can deliver them to customers.

Very interesting. Is it fair to say that India is one of those markets which exposes the weaknesses of speech models the most as compared to any other market?

I think Indian developers are very direct.

Okay.

They're pretty blunt.

Yeah.

And the scale here is insane. If you can be successful building your product for the Indian developer and the scale that they're operating at, and if you can actually make sure that the way you build your product is because in India folks have the patience to go in and listen to every call that's happening, as an example. This is what I heard from one of our customers is that there's gonna be a whole team that's gonna go listen to a call inside an enterprise or this collection of calls that happened and they can give you super fine-grained feedback. And they're gonna be pretty picky about what they're seeing. So that is a level of obsession with detail and quality that is surprising perhaps to folks, but I think it's kind of actually cultural. And I think if you can build for that, I think you can be very successful anywhere else in the world. So I do think it's a battleground for like actually building super good product.

It stresses you.

Yeah, it really does.

Okay, now I'm going to throw a curveball question at you. See, because I've been tracking the voice AI space in India for a bit now, I think I would have met almost every other company out there. There are a lot of new emerging voice AI model companies coming out of India, and then there are obviously global players like yourself who are targeting India as an important market. And we've been wondering, like, who has an advantage to win and why? Like, what would you— like, what would differentiation look like for different models? I think that's where the question is coming from, that if different models— if we want customers to have different— have more options when it comes to models, where does differentiation lie?

Yeah, I mean, I think when I was in grad school, actually the word foundation models was coined at Stanford and I was part of that 200-person or so team.

Amazing.

And we actually voted on the name and now everybody in the world uses it, so it's kind of great. So cool. Yeah, which is kind of cool. But I think the thesis there, which I think equally applies here, is that models will eat more and more of the stack and get more and more general, right. So I think that's kind of one of the things we believe in, which is task-specific models have a shorter shelf life. You have to build for generality and you have to do it carefully because you can't just say I'm going to build for 10 things, you have to sequence it in some way. But you also have to acknowledge that like a task-specific model is going to be eaten up by a model that's more general. So that applies to this, which is if you go too niche on the capability, then, you know, eventually it's just going to be subsumed by somebody else's model that can do that and do 10 other things. But if you're really shooting for, let's say, having the best model, then— and if you sequence it right, you can actually get there. That's kind of my perspective. So I wouldn't say that I think that it's not— like, I actually think, you know, I know there's a lot of debate in India about what kind of company should get built in India. I've always been of the opinion that folks should do more deeper tech and invest in actually building hard tech as well. So I think that's all great stuff and I think folks should work on that. And I think that the whole point of AI is you can be surprised by something that people invent or build. I think for us it makes a lot of sense to try to build this in this more general way and we think that The pie is super big, it's growing fast, and you should just focus on building the best models possible, and there's enough for everybody to do. But I would say I never want to say, hey, like, don't go do that because we're going to— yeah, I think it's great to have folks that are actually trying to build some stuff like that.

I think where that question was emerging was from differentiation, like how do you differentiate and what are those parameters of differentiation, but I think you answered it well. I think one other question, Karan, is a big one where does Cartesia want to be a model company, an application company, or both? And the reason that question emerges is that there are a lot of voice AI application companies that are coming up and they're always thinking whether the business model will make sense or not.

Yeah, I mean, I think if you look at— excuse me— our research DNA, what are we great at, right? Almost half of the Cartesia team works on research. Okay, how big is the team? We're about 85 folks.

Oh wow, that's a small team.

So, you know, half of the team is working on research and there's a reason for that which is we are 5 founders that were researchers and there's a core DNA around doing world-class research. That is always going to be the number 1 thing that we want to be great at and that means that you can call that a model company, right, we're a model company. But I think our sort of philosophy around what we're building is more product-led in the sense that what we're thinking about is how are our models getting used, how do we make sure that we're covering what people need from us. And so that can mean like actually what people want is as I said like end-to-end employees that do work and they're able to connect to context and all these things. And I think that's what we should shoot for, but we'll do it in our own way is what I'm getting at. Like, for us, we really believe the right way to do that is not to engineer a lot of systems, which is our bet, it's to go build it in the model layer. And that's how we kind of are different from, let's say, a lot of application layer companies, quote unquote, because they're often using engineering to go out and solve those problems. But we sort of feel like, you know, especially coming from a research perspective and having a little bit of a you know, chip on our shoulder about that. Like, we feel like, okay, that needs to be done by the model, and the model should over time just subsume that part of the stack. Now, we don't know who's right, and it's going to take a while for that to be litigated. But I think that to me it feels valuable for folks to be building for specific verticals and applications because that is something that requires a lot of deep work. Like, being an expert in like healthcare or banking or, you know, insurance, etc., is not easy. And Cartesia, we sell to companies that work in pretty much every sector of the economy, so it's not possible for us to build for all of those verticals and use cases. So we're not trying to compete on that part of the stack.

No, that's very fair. I think we'll know at some point in time on where we end up spending more time, but I agree with you that there is so much workflow, there's so much domain expertise that you need in certain kind of verticals. That those companies are best suited to do them?

I would say our philosophy is when we talk internally we say we don't want to be classified as an X or a Y company. We typically look at the problem as it is presented and we say is this the right problem for us to solve and does this make sense for us to add to the stack of things we can do while keeping the focus that we do have. So I'll give you an example. We obviously build models, data, etc., but we also have built a really great platform. We actually built our own inference engine, right, so we actually for core things that we think are super important to the business we just go solve them. And then you could argue like we're also an inference provider there, right, but that's not how we kind of divvy up the thing. We're building an agents platform because we think developers are struggling to build reliable agents and they're doing a lot of repeat work, so we said hey, Like why is it that if there are 10,000 voice AI companies each of them has to go build the same infrastructure, that seems like it should be our problem. And so that's something that we stepped into. So I think that's how we kind of operate, it's just like what makes sense for us to do versus where do we absolutely not want to go fight battles.

Very interesting and see this also came up in some of our conversations, a lot of our portfolio companies use Cartesia and I've heard great things. So there are two questions here, right, one is that pricing pressure is very real in India. Now things here like there are obviously a lot more companies giving out today models than there have ever been. A, are we— what does pricing look like for India, is it how different is it than US? The second one is are we overly subsidizing right now to gain market share?

I think for us like we never try to sell the product to anyone at negative margins ever. We would rather lose that deal than compromise on that, I think just as a matter of sanity. So I think for us like we really go out and try to obviously localize to what the market needs and wants and I think it has to make sense like from a business perspective. The only way that a lower margin sale would make sense which is typically what you would see in India, is you have to have the volumes to compensate for that. So I think that's kind of the game that has to be played. And so when we go work with folks in India, I think partly it's like, you know, for some folks they're still trying to understand, like the volume, the peak volume that we're going to get to is going to be insane. So a lot of folks here are still in the phase of the first set of pilots are successful. And we're seeing some of that usage translate into what we see in our Datadogs. But we haven't yet gotten to the point where those massive scale opportunities have emerged. So that's where we're excited about it. And I think as companies get more sophisticated, like, you know, obviously you work with a lot of voice companies, they're also going to innovate on pricing. Because for a company that's building in India for, let's say, a sector, let's say BFSI, they're going to figure out, hey, like, what is the value that we're adding here? And how do we quantify it. And then how do we charge for that value and not just charge for the number of minutes of conversation that I did. For us that business model can make sense because we're selling to basically everybody that's building that but for them they're going to figure out that value piece and that I think is going to ease up some of those like margin pressures let's say. Because you can't stack one on top of the other and have 4 companies selling per minute pricing, like somebody has— there has to be a release wall. So I think that's partly what I think is going to happen over time. I think you just learn more about the problem and then when people realize hey, like I'm saving so much money doing this you can charge more, it's going to make sense.

Yeah. See, other question I have is that when you sell to enterprises in India where they're using Cartesia as a model one is obviously the subjective voice quality that hey, we're much better, we're really, really good, we're really fast, there's hardly any latency. But the second part is what are some of those objective metrics that enterprises really care about today when they compare us against someone else?

Yeah, so I think as Cartesia, it's actually interesting because it depends on the— so there's two ways that we sell, right? We partner with a lot of companies that are building for a vertical, so vertical agent companies. All of those vertical agent companies have different metrics that they're actually focused on for their customers. So if you're working in the collections space, you know, you will care about something totally different in terms of like how much you're actually getting back from a total rupee perspective versus if you're working in the healthcare space, let's say, and you're doing— let's say you're doing appointment scheduling, right? Appointment scheduling, your metrics are going to be totally different. It's going to be like how many appointments actually got scheduled by the AI versus the human. So I think all of those metrics look quite different for our developers and the businesses that we serve. And so I think almost at Cartesia what we're always looking for from our customers is give us more visibility into what your end customer is doing with it. So if you're working with a bank, give us more visibility into what they're saying so that we can help you be more successful with them. That's kind of how we operate actually. So we'll spend a lot of time, for example, trying to understand like what's driving that end metric.

Fair enough.

And so it could be CSAT, it could be dollars or rupees that are coming back to you. Or it could be something different. So how do we kind of like get visibility into the end metric and drive that, that's super important. And as you said, it's kind of hard to say there's one metric to rule them all, right, like it's actually a lot of different ones. And it's also that they're also trying to figure out what's the right metric for us to focus on. So all of those things keep changing.

And do you think there is a difference in the way you sell in India versus the US, or it's more or less the same?

I think for us it's been fairly similar actually. In that most folks now will do— it's like product-led or developer-led growth, which is they'll go to the internet, they'll find us through word of mouth, or they'll see a post that was made, or they'll hear from another business, or they hear through their VCs or through the grapevine, right, like they'll hear about us and then they'll go try the product. This is something that's super important, which is I think that there's an expectation that I can use your product and understand all of its features without needing to talk to you, which is something that we obviously super focus on. That's the way that most folks get onto the platform. It's that they're going to see that and then they're going to go say, hey, I would love to use it for all of my customers. Let me go figure out how I can talk to someone at Cartesia and actually get help doing that. That's when we start seeing them graduate to sort of more sales. So I think that's how we sell everywhere in terms of like our go-to-market motion. And that's something that we're obviously spending more time maturing, expanding, and so on. So I do think that's one key part of it. The other is I think the Indian community is in some ways smaller and perhaps—

And warmer to you.

Yeah, and perhaps warmer to me. So I think that's also— helpful for us being based in the Bay Area but then also having folks in Bangalore and having a connection to the country. So I think that helps too. So folks have been super nice in trying our products, adopting them, giving us a lot of feedback and helping us really make it great for the market.

See, I think also now coming to the tail end of it but you've spent a lot of time this week with a lot of founders, operators, customers. Just what's your take, what have you observed, any new thing like the way India used to look 10 years back versus what you feel now?

I mean I feel like I'm pretty outdated. I think just the pace at which things are moving is pretty amazing actually. I think that one of the things that's been interesting is there's a lot of desire to do— to build inside companies technically as in like rather than saying like give me the whole solution wholesale. I think a lot more folks are like I'm going to have my tech team do more and I think that's interesting as a pattern because I think that means that a lot of companies think of them as technology companies which is I think something that are not realized. So that's an interesting thing that we saw. I think the other thing that we saw was BFSI is ripping in India in terms of how much stuff is happening on agents and voice in in BFSI. And so that's super exciting, like it seems like that's going to be the beachhead for showing folks what's possible in voice. And so it's super important that that becomes very successful so that people feel like there's a lot of ROI and value in investing in voice as a UX. So I think those are two big themes that I saw that I think were pretty interesting to me and I had not understood or realized before I spent the last 10 days here. So I think that's been great.

Amazing. Okay, one last question, Karan. 5 years out, if Cartesia really becomes what all of us really hope it to become, what would conversations feel like?

Probably more like this. More like this? Okay, fair.

Yeah. But like, what are you— like, what's your end goal? How are you thinking about Cartesia for the next 5 years?

Yeah, I mean, I think for us, like, it's all about the models. It's all about the models. You know, I always tell folks, like, AI is just about models. Where do you think AI happens in the stack? It happens at the model layer where everything, all the data is going in, the compute soup is being, you know, stored around, and then you get something great that you can use. So I think for us, like, we are super excited, you know, in 5 years to, like, really get to a place where I'm only selling folks one API and it's just, you know, this magic API that you can use like a human. I think that's where we want to get to, and I think along the way we're super interested in solving some of those core problems in research that are honestly not just a bottleneck for Cartesia but the whole world, right? I think that's part of our, you know, our culture is to work on frontier problems and figure out how to solve them. So I think that's something that we're super excited to do more of. But hopefully we'll do this again in 5 years.

Yes, 100%.

And I won't be here, my AI is going to do this one.

Oh damn. No, but this was so much fun and thank you so much for spending time with us and being so honest and candid. Loved it.

Thank you.

Thank you.