Siddhartha Ahluwalia 0:54
Hi, this is Siddhartha Ahluwalia, your host at Neon Show and Managing Partner at Neon Fund. A fund that invests in the best of enterprise AI companies between the US and India corridor, like Atomicworks, SpotDraft, CloudSEK. Today I have with me Sudarshan, founder of Smallest.

Sudarshan, welcome to the podcast.

Sudarshan Kamath 1:08
Thanks for having me.

Siddhartha Ahluwalia 1:08
It’s one of my biggest regrets, you know, I can tell you that not being able to be part of Smallest because as you know, right, we were raising our own Neon Fund 3 at that point of time, we got connected. Right. But but glad to see your journey, really inspired.

And, you know, you have like a set sky as the limit in your journey at Smallest. And what you have been able to create out of India is transformational. So kudos to you on doing that.

Sudarshan Kamath 1:28
Thank you.

Siddhartha Ahluwalia 1:28
And I know that I’ve been a founder from India and still running our own funds, almost like a founder. The ceiling on India sometimes hit you and you have to break of that. Then in your case, you know, before we dive into Smallest, tell me about your journey and how did you break that ceiling?

Sudarshan Kamath 1:44
Yeah, I have grown up in India. I was born in Mumbai, brought up in Pune. I was lucky enough to choose science when I was young and I really liked physics.

So I’ve always been scientific in my younger days. And I when someone would ask me what I wanted to be, I would say I want to be a quantum physicist or like I want to go study large, large bodies and things like that. I got into college in India and IIT.

And at that point of time, you realize that there is very little opportunity to pursue these sort of deep research projects. No one is doing that. Everyone is doing competitive coding.

Right. And so I used to write code since I was very young because my father was in Infosys and he got me a computer when I was young. So I could write code and I wrote like database and a bunch of other things.

So I was like, OK, let’s enjoy code. But what do we what do we do about it? Like, why are we doing why are we competing?

And I did not want to do coding competitively. I wanted to build something. And luckily, I came across computer vision as a topic.

So computer vision is pre AI days. Today, you call it vision language models. But at that time, it was just a very basic approach to understanding images and videos.

Got the opportunity to work at like Samsung’s one of startups that was funded by Samsung in Korea. Worked in self-driving cars like this is like when OpenAI was just coming up or like pre OpenAI days. And so that got me like really interested, like, you know, AI seems to have like a broad set of applications that, you know, we can have.

And so no one was giving a job in AI at that point of time, except some of these foreign companies. And I’m like, OK, I want to do this. I’m going to do this.

Siddhartha Ahluwalia 3:14
And this was which year?

Sudarshan Kamath 3:16
This was in 2017, 16, 2017, basically.

Siddhartha Ahluwalia 3:19
So you were in first year of college?

Sudarshan Kamath 3:21
I was in second year, first, first or second. Yeah, basically. So, so, so, yeah, so I was like, I want to really go into AI and figure out AI, but no jobs in AI in India.

Right. So I got a job in Korea. I was like super excited about going there.

I like I was studying, doubling down into AI, etc. And then I was not able to get a visa for some reason. They were like, you know, if you are below the age of this, Korean government changed the rule that you have to hire Koreans, basically.

So I was like, OK, what do you do now? So I got the opportunity to work in a startup in India that was building something in EdTech plus AI. So you must have heard of toppr.com.

Yeah. Right. So, so that was like completely different from my studies on computer vision, pure research, etc.

But we built something which was like the largest repository of like EdTech material in the world where we indexed all textbooks from various syllabuses, CBSE, ICSE across India. And we overtook Byju’s in traffic and we got acquired by Byju’s for like $150 million back then. And so we were also building this pre-chat GPT, this AI bot, which could basically even ask any question and then it can come up with answers.

And then you can also connect with a human to answer those questions. So I enjoyed that. And that got me like really curious about startups, basically.

Of that toppr team, there are 10 startups that have come out. Largest is Kuku FM. So Kuku is gonna do an IPO very soon.

But yeah, everything I know about how a startup should operate came from toppr. One thing was very clear, though, is if I want to build something, I want it to be in deep tech. And so I started again looking for opportunities in AI.

And luckily there was this professor called Arjun Jain. So Arjun was working with Yann LeCun. Yann is one of the founding fathers of AI.

He was one of the first people in the world to do deep learning. So back when AlexNet came out, he was in the heat of things and he built this company, Arjun, which he sold to Mercedes. And then he used to act as a consultant for all of these automobile companies for self-driving vehicles.

So I was like, this is my opportunity. I’m going to work with Arjun. And so I kind of got there.

I started working with him. We started building the Hydra model for self-driving cars over the next two to three years. We built like a model that can run in real time on the vehicle hardware for level three self-driving across multiple countries, Germany, US, Japan, etc.

And that was where I learned how to productionize AI at scale. And also how that was the point where I started realizing how powerful small models are. Because what we used to do is we used to train these cloud-based models that were pretty big.

And then we used to make them smaller through this process called distillation and make them run on the vehicle hardware in real time. And what that taught us was that most of the pre-training data for these large models was junk and garbage. But when you use that pre-training model as a teacher for this small model, and we could go into what teacher means, but it would basically act at similar accuracies as the larger model at one hundredth or one-tenth of the size.

So that kind of got me thinking around why is that behavior there and do LLMs also have similar properties, right? And so that’s when I started doing some more research around small models and I bought the domain Smallest AI around then. This is like I got it for like $100.

And the idea was basically how to…

Siddhartha Ahluwalia 6:43
Which year?

Sudarshan Kamath 6:44
This was 2021-2022, right?

So this is like even one or two years before I started the company actually, right? And so then that idea was always there. I was doing like a lot of research and eventually I decided to basically quit my full-time role and actually go…

Siddhartha Ahluwalia 7:04
Where were you doing full-time role?

Sudarshan Kamath 7:06
I got into this legal tech plus AI. So we were trying to build like a Harvey AI competitor basically called Zolvit Vakilsearch. One in 10 businesses in India uses Vakilsearch for legal tech compliance, etc.

So this was like a place where you could get like an accountant in like 10 minutes and basically also for lawyers, we have a product where you can automate your case management and things like that. So we were one of the first people who had actually adopted vector databases in India because we were one of Pinecone’s first customers and we were building out RAG and we were building out all these systems. But I think that was not a very successful outcome.

We realized that if you have to build for the legal market, you have to build from the US basically or for the US at least, right? Everywhere else is… The market is just too tiny for you to get any money out of that sadly.

But learnt a lot there about the whole how do you separate memory from model layers. So that was a place where I got to experiment that because you have a large amount of legal judgments and all of that has to be indexed and it can’t be stored in the model context, right? So you have to…

The idea of context engineering, which is becoming popular right now, was being worked upon by us back then to manage these legal searches and answers on that basically. So yeah. After that, I was pretty clear like I’ve seen like a startup going to an acquisition.

I’ve seen like a deep tech product going from zero to production. I have seen another startup go to like a series B and so I wanted to start something of my own. I’m not talking about like the other small attempts I have tried on my own because they were very, in my opinion, juvenile.

But one thing was very clear that if I wanted to start, you have to just go at it. You can’t do… You can’t half as a startup.

You have to do it full time. You have to put your money behind it. I put the first 100, 150k into smallest of my own and we basically got like some GPUs and we started like training models.

Siddhartha Ahluwalia 8:59
This is which year?

Sudarshan Kamath 9:00
This is 2024.

Siddhartha Ahluwalia 9:01
You quit your job in 2024 and then?

Sudarshan Kamath 9:03
I quit my job 2023. I was like figuring out what to do for like 6 years and then 2024, second half is when sort of… We were training some models in 2024, second half is when we are like, okay, something is working and our training was working.

So 2024, second half, we basically, we were experimenting with small models and we realized that one of the applications of this small models was to build real time systems because they have very low latency like 100, 200 milliseconds of latency. And so we started looking at, okay, which applications have low latency? One was your finance and I really don’t like finance.

So I was like, I’m not going to be good at this. But voice AI seemed very exciting because voice AI was just coming up 2024.

Siddhartha Ahluwalia 9:47
ElevenLabs started becoming popular.

Sudarshan Kamath 9:48
Yes. ElevenLabs was started, I think last year, maybe 2023. I’m not exactly sure of the timeline, but so voice AI was coming up, but ElevenLabs was completely into offline dubbing and things like that.

They did not have a real time model. Right. So we were one of the first people to build a real time text to speech that ran under 200 milliseconds.

So we, we put a demo on LinkedIn that went viral. And that’s how we like, it took two weeks to basically go from two kids writing code in a bedroom to, you know, raising a pre-seed round basically.

Siddhartha Ahluwalia 10:20
How many VCs had you spoken to back then for your pre-seed?

Sudarshan Kamath 10:24
Yeah, that’s a very good question. So we did not, we did not know how to do this VC startup. We thought like, I mean, you know, you just like go and ask for it.

And like we will get money. I had no clue what it takes. What does a VC look for?

Nothing like that. So we would get random analysts coming and messaging us saying, hey, you’re building something because we would put like stealth startup in our LinkedIn. Right.

So someone would come message. We must have spoken to every single VC at that point of time. And then they would randomly come and reject us.

I’m like, I’m not even like raising. Like, what do you mean? But, but like someone would come say, oh, you’re too early.

You should do this. You should do like, dude, I’m not asking you. Like I’m, we are working on it.

Right. So, so, so that would happen. And then we knew we had something when this text to speech model went out, basically, because that like we were getting like literally like with, there was a two week window where like my LinkedIn following like two X, three X or something like that.

My like, like the LinkedIn DMS were completely filled. X DMS were completely filled. Emails were coming in.

I was in customer calls back to back. What do you want? What do you want?

And then the product was just an HTML page with the model with one GPU running behind it. Right. And I had like two, three enterprises coming saying, we want this.

I was like, what do we do? That’s when I was able to get some formal interest. And I, I remember I, I had gone to Bangalore.

Siddhartha Ahluwalia 11:49
You were in Mumbai back then?

Sudarshan Kamath 11:50
I was, I was in Mumbai. I, so, so a little bit of history before this. So I had dropped out of UCSD as well.

So after Bosch, I had thought like, should I do my education? Because like I told you, I was like a slightly scientific person. Right.

So I was like, should I go back to studies? But when I got into UCSD, COVID was again, I think the third wave or something like that was happening. And I was like, I don’t want to do remote.

Plus I got this other legal tech offer in India. So I was like, I’m going to drop out. But so I was planning to come to the US and, and raise again, because I had some network connects, et cetera here.

But I decided, let’s like, let’s try to see what the Bangalore ecosystem has. And I was at that point, not aware of the stark difference in Bangalore and US. Right.

So, so, yes. So, so I was in, I was in Pune. I went to Bangalore.

Pune is my hometown. So I went to Bangalore to basically talk to VCs. Right.

And I remember the first meeting we had, that was the, I think, a really good meeting we had luckily with this VC called UpSparks. So UpSparks basically met us in office. They’re like, we want to talk.

And we were like, okay, who are these guys? Like, let’s talk. And we gave a pitch.

They probed us a lot on the tech. They wanted to see whether we were like just bullshitting or did we actually understand what we are training. And obviously we knew our shit.

So we gave them a AI 101 lesson. Right. And after the pitch, I was like, okay, like, where is the check?

So I, I was expecting like, okay, like I pitched myself. Now we have other VCs to go to. So can you just tell me a yes or no right now?

He’s like, no, that’s not how it works. We need like a day or two to decide. Okay.

So I’m like, can I come like tomorrow? Like what? So, but I remember they were the first to give us like a term sheet basically.

And the moment UpSparks term sheet came like the next day, we had Vaibhav from Better Capital who came on a call. His was the smoothest process. He came on a call in the morning.

By evening, we had a term sheet. Right.

Siddhartha Ahluwalia 13:39
So how much did you raise from each UpSparks and Better?

Sudarshan Kamath 13:41
UpSparks was like not much like 100K if I am not wrong. And Better was another 250 or so.

Siddhartha Ahluwalia 13:46
Total round was 350K.

Sudarshan Kamath 13:48
350K. And then a couple of angels of 400 or something like that. This is at like a 3 million post money sort of around.

Right. You have no idea what the valuations are. Is 3 good?

Bad? I’m like, oh, you’re giving me money. Like, no, this is like too good, you know.

But then I remember I was running around Bangalore. I had meetings with like every single VC in these two weeks when we had like gone viral. And 314 Capital, I had never heard of them.

Okay. I ended up at their office late for the IC meeting. So I came like 15 minutes late.

My bag was torn. So it was like I had an open bag. I had a torn shirt for some reason because it got like stuck in the auto.

So the bag and the shirt both tore. 15 minutes late. Coming in between the IC meeting.

Pitching. And then I remember there was this one person sitting who I have never met before. And after all of this, we came out.

They are like, we are very interested. And like the right the next day, we got a term sheet.

Siddhartha Ahluwalia 14:40
And that was like 186?

Sudarshan Kamath 14:42
That was they had offered something like that. We asked them to reduce the check size. We said we can’t take one at six.

It’s like too much dilution for us. So we I think we ended up at like some 500 at six or something.

Siddhartha Ahluwalia 14:52
And this is what within one week of UpSparks?

Sudarshan Kamath 14:53
Yeah. Everything is happening in like one or two weeks of this thing. So finally, we ended up saying we don’t want to raise more.

We just closed like 900-ish.

Siddhartha Ahluwalia 15:04
600K from 314 and remaining from

Sudarshan Kamath 15:06
Exactly.

Exactly. Right. That was the plan.

Now, what happened is we got to work. We got like a small house in Bangalore.

Siddhartha Ahluwalia 15:13
And this was an India HQ company or US company.

Sudarshan Kamath 15:16
So that’s a very good point. Now, the next question that came to us like during this time was where do we incorporate? Right.

Because I had an India company. There was like no revenue or nothing in that, which is a good thing because we had two options. Either continue getting revenue there, get invoices there or flip.

Flipping was going to take two months, two to three months. But for us, it was very, very clear. That was a bet I took that it’s okay.

Like even if we don’t raise these term sheets go away, it’s okay. But let’s flip. So I decided to set up a Delaware C-Corp.

And then we went through the whole ODI process, etc. And then we set up the India as a subsidiary. And the core logic was we wanted to build a global company.

I had like no interest to build a local company. The reason Silicon Valley is big is because all over the world, people use services from here. And we wanted access to global capital.

We wanted access to global talent. There was no way we could do that, unfortunately, by being there. We obviously want to help India.

I have benefited so much from great people from India. But I think there is an indirect way to do that even with having a subsidiary. You can literally hire the best talent, pay them a lot more than everyone else and help India that way as well.

But my thesis was very clear. I want to work on something more ambitious. It is not just Indian languages.

Indian languages for me was like just an operational problem to solve, getting data sets, etc. But it was not a scientific enough problem that helps us answer what’s like the next level of AI and how can small models be the future, etc. So those kind of problem statements, my hunch was only US would answer.

And I think that was one of the best decisions we have ever taken in the company.

Siddhartha Ahluwalia 16:57
You got all the money wired to the US entity.

Sudarshan Kamath 17:00
Yeah, so the money that was coming to the US came in like the moment the entity was set up came in like three days. Basically, you just sent a safe note, signed it, got the money in the bank. The money that was to be wired in India went through like a whole bunch of processes.

We had like physical paperwork signing. Like, I don’t know what all like we had to do. But yeah, like that is that was like a mess basically for us.

Siddhartha Ahluwalia 17:20
This is late of 2024. Late 2024. And then what’s the next that happened after that?

Sudarshan Kamath 17:26
We kept building. We got a small house in Bangalore. It was in Indiranagar by the metro station.

We got our first couple of employees from random sources like these were not friends. One strong hypothesis I have is sometimes not hiring from your comfort zone is better. Because otherwise you end up having like the same learnings basically.

So I generally look for like people who I don’t know but are also ambitious and can teach me something. So we ended up hiring based on like Twitter and like LinkedIn folks. Ex-founders basically who had tried to build their company shut down.

Very unique profiles. Like the second employee I had hired never went to college. He was training model since he was like 16 years old.

Learned everything on his own. Another guy had built a services company scaled it to like a million or so and then shut it down. Very unique profiles we hired for and we were able to build like a good team initially.

I think till this point of time the VCs did not even know that we would be able to build a team, a product, nothing. They just took a bet on like just two hungry people and let’s see what happens. But I think as this started happening and we started getting interest and they saw us working.

We got offers to do you want to raise another round, do you want to like extend etc. For us, it was very clear that we don’t want to raise from India more. So we actually had great offers at like pretty good valuations in India.

But we were like no, we want to come to the US and we want to raise from here. So I had my own process etc. ongoing.

Siddhartha Ahluwalia 18:52
But did you raise any more amount after that?

Sudarshan Kamath 18:54
Yeah.
So what happened was we started the O1 process and then parallelly. Luckily, I was thinking what to do because we had like 900k and come to the US, it is like slightly more capital. So I don’t know if you know her but Aarthi Ramamurthy just DM me on X.

I’m like wow, like I was just watching her podcast like last year and like why is Aarthi messaging me? And she just turned out to be very polite, very smart, very down to earth. She was just like curious about what I am building.

She was building her own fund and like she was also investing from her capital. She basically got on the call and she was like I love you guys. I want to invest, right?

And I told her look we have this offer from Indian VCs etc. What do we do about it? So I want to raise from the US.

So she said I mean that’s the right decision. Let me get some angels together for you. You’ll get some more runway and why don’t you come to the US and invest.

So I did that. Aarthi connected me to a bunch of angels. Vaibhav connected me to a bunch of angels.

We raised another 900k extension that was at a higher valuation. And we had like 1.8 of which most of it was in the bank. So I was like let’s go to the US now.

So that’s how I first landed in the US. Now I’ve lived in multiple countries, right? So for me US was not like a foreign land or anything.

It is another developed country. But what’s different is like the tech density. I feel like that is the biggest differentiator that I saw here.

And anyone could be an American, right? So that I really like. I’ve lived in Korea but I could never be like Korean.

I’ve lived in like Germany, Singapore. But you would not be natives there. You can’t be a native in China.

I’ve lived in all of these places. But US, you can be American, right? This Bay Area has been built by smart people from all over the world coming and like building amazing company.

And the other thing I’ve seen here is talent is appreciated from anywhere. It’s not like you have to have an IIT or a Stanford degree. This is full of people who had nothing.

No background, no credentials. But just raw dog themselves and like build an amazing company. And then now they are very well respected, right?

So I thought like this was the place to be. And we wanted to then sort of raise a seed around because we started getting some traction. We had some revenue.

Siddhartha Ahluwalia 21:11
And what was the use case that you were solving for which you were getting traction?

Sudarshan Kamath 21:13
So everyone was building voice agents at that point of time. And we had a model that could power the voice agent. So we had a text-to-speech model, which basically you could use to give natural realistic voice to the voice agent in under 100 milliseconds.

So initially it was at 200 milliseconds. It came down to 100 milliseconds by the time our seed round was there. And then we had some customers using us for that basically, right?

We also had started other series of model trainings because to train a text-to-speech model, you actually need a speech-to-text model to annotate the data. You need another model to generate synthetic data. You need accent conversion to diversify the data.

So that for one data set, you could convert it into multiple accents and diversify it. So we had like a portfolio of models that we had developed. And then we had a place where you could orchestrate these models to build the backend infra for a voice agent.

So let’s say anyone is building like a AI receptionist company in healthcare. They could use smallest AI to basically build on top of it and build those voice agents. And in the future, now what we are building is basically making these agents even smarter so that the model layer can basically always stay up to date in real time.

And all the user information that it gets from multiple interactions that gets stored in like infinite memory. So we are trying to now separate the compute intelligence layer from the memory layer so that these voice agents can become smarter and smarter as it has more interactions. But that is so at that point, we just had these models.

We were doing well. We did not have a lot of revenue, etc. We had like some small accounts here and there.

And we went out to pitch, raise a seed round with like zero revenue. Yeah, almost zero revenue, right? But the difference I saw like here while raising was they are not looking for…

Obviously, people need revenue. This is validation. It’s validation.

It gives you the power to negotiate. It gives you the power to say we are real. But people are looking for why you?

What’s different? Like what are you building? What’s your vision of the world?

Can you execute the vision? And so I met with Tim Guleri from Sierra Ventures.

Siddhartha Ahluwalia 23:16
How did you get to know them?

Sudarshan Kamath 23:18
So there was a partner at Sierra Ventures, Ashish. So Ashish is my college senior. And in our like WhatsApp chats, my co-founder was asking about, Oh, has anyone trained this model?

And like this issue we are facing some QDA errors, etc. So Ashish saw who is this guy who’s like constantly asking about…

Siddhartha Ahluwalia 23:36
On your college group.

Sudarshan Kamath 23:37
On our college group. Like IIT Guwahati, the college I’m from, it’s not like that entrepreneurial in nature.

Siddhartha Ahluwalia 23:44
Mostly IIT Delhi and IIT Bombay are.

Sudarshan Kamath 23:45
IIT Delhi, IIT Bombay, even IIT Madras has…

Siddhartha Ahluwalia 23:47
Even Kharagpur.

Sudarshan Kamath 23:48
Kharagpur.
These guys have. But Guwahati is still like slightly like newer. And also, I mean, although it is 30 years old, but it’s still considered new.

And it’s also remote, right? So not a lot of like startup opportunities, etc. So very few founders to sort of look at.

But we were one of the few folks who were like asking around how to build. Oh, they’ve raised around. So I think that led to a lot of interest from our alumni plus our juniors because, Oh, there is someone we can like look up to.

And like we can also actually build something crazy. Right. So Ashish had reached out to us.

He had expressed interest that we want to invest. I mean, I was completely clueless about what does it take to raise a US round? What are they looking for?

But then I met Tim and Tim and I had like a great conversation. Like Tim was very technical as a partner. He has been a founder himself multiple times.

He sold his companies. He knows a lot of people in the contact center space. So he felt like a perfect partner for someone like us.

He has also invested in like companies that were previously out of India like Makemytrip, etc.

So, so he had that India angle as well to it. So he seemed like the perfect partner to build like a very big company with. Yeah.

Right. And so, um, I think the conversation basically became fairly, fairly technical. And then we were, uh, we had an IC meeting and at that point of time, I had two options.

Either I, I raised from Sierra or I just raised from an Indian fund and get done with it. I told Sierra, look, I really want to go with you guys. I, I think I would love working with them.

I’d love working with you, Ashish. I would, um, I think we can build a big company here. I told them the same thing, but you need to tell me like today.

Right. So, uh, so I remember after the, after the IC discussion, they wanted to have like a dinner that same day. Um, we had a dinner, uh, we spoke about.

Siddhartha Ahluwalia 25:36
One company where we invested recently, uh, uh, Sierra offered them and then they choose engineering capital as the lead. Okay. So what I heard is, is, uh, Sierra’s usual style is over a dinner close over a dinner.

Sudarshan Kamath 25:48
Yes. Yes. Yes.

I think that’s a good style. Like, honestly, like it, uh, because at the end of the day, I feel if you’re not able to have a dinner together…

Siddhartha Ahluwalia 25:07
There’s no point of in building it for 10 years.

Sudarshan Kamath 25:56
Yeah, exactly.

Right. So, so, um, he, they, we met, we spoke about this. Um, then she offered me a term sheet.

Um, and yeah, basically we negotiated and closed. That gave us a lot of relief because now you have..

Siddhartha Ahluwalia 25:23
Serious money in the bank.

Sudarshan Kamath 26:12
Serious money and serious US money, because suddenly, uh, the biggest difference I saw was we got introduced left, right, and center to very serious by buyers of our product. For example, service now, for example, right.

Siddhartha Ahluwalia 26:25
You have Service Now as a client today.

Sudarshan Kamath 26:27
Yeah, we are working with them. We are hopefully going to make some big progress, but like, this is just one example. And there are 10 other examples.

Siddhartha Ahluwalia 26:35
You got introduced to Service Now.

Sudarshan Kamath 26:36
Yes. Yes.

Yes. Yes. Yes.

Some of them, we, uh, we will publicly announce like in March, April, but like, these are like large enterprise accounts. I can tell you like, if, uh, you know, if we do our job, well, just from the introductions we got, we could get to our ARR targets for this year, right. Completely different than how I was operating in India, where we did get introduced, but the introductions were not like that strong that, uh, I still had to do a lot of work to actually do the stakeholder management and all of those things.

Right. Um, and also I just feel the companies here are very, very clear of what they want to buy, what they want to build. Um, and what is the evaluation process for buying?

Uh, right. So people would generally not waste time because they don’t, they don’t have time to waste.

Siddhartha Ahluwalia 27:19
So for example, one of the telephony companies that you mentioned, yes. Was it also an introduction from Sierra?

Sudarshan Kamath 27:24
Yes. Very, very strong introduction. Um, I obviously like introduction, by the way, it does not mean they would buy.

So the first call I got on with this telephony company, the CTO grilled me, you know, he, he grilled me. Um, he was like, uh, do you know this? Do you know that he’s a very deeply technical person?

Um, and then I was able to answer those questions because I am also deeply technical. I enjoy what I’m building. I know my products.

Right. So, so that impressed them. And then we kept at it.

Obviously we had to do our work. Sierra helped us, uh, sort of, uh, make sure we understand how their buying cycles have looked like, what do they care about, et cetera. And then eventually, uh, they’re like, okay, we want to, we want to do a deal with you.

We want to, we like the product we’re going to, and then we went into negotiation and that, that’s where, you know, I, then I could handle it basically now, but I know what the sales cycle looks like for an enterprise. And like, this was my first enterprise sale in my life, because in toppr I’ve built like, it’s more B2C, B2 student, whatever you call it. Um, VakilSearch was again, B2B, but like small accounts, et cetera.

This was like first enterprise sale of, of like my life. Right. Um, I have worked with other enterprises in India before, but they always like spent like a lot of time evaluating something seemingly simple.

And like, for example, let’s say they have to spend, um, 20 lakh rupees, 30 lakh rupees, the amount of time and people that would be involved in that process would easily surpass that 20 lakh or 30 lakh they have to spend, but they would still wait than actually buy and just be like done with it, you know? So the, so I felt like here, it’s very different. It’s very expensive to just drag things along because the salaries you pay to people, all of that is so much.

And 20, 30 K is not a big deal. You know, like it’s, it’s like, it’s fine. So here you could easily make a 20, 30, 40 K, 50 K sale.

Siddhartha Ahluwalia 29:05
So that, that telephony enterprise, that’s a much bigger deal, by the way.

Sudarshan Kamath 29:08
So when it started, it was like 10 K a month or something like that, which is a big deal.

Siddhartha Ahluwalia 29:13
I mean, it’s a 100K annually. Yeah. But today did it get converted to a seven figure deal or?

Sudarshan Kamath 29:17
It is, it is converted to a seven figure multi-year, multi-million deal. But we, we will announce it formally sometime like in April or so once we sort of clear a few things, but yeah, it is, it is that like, all of those things are in paper right now.

Siddhartha Ahluwalia 29:32
So Sierra happened in April last year.

Sudarshan Kamath 29:34
Sierra happened in June, July. Right. So July, July is when Sierra happened.

And yeah, all this happened.

Siddhartha Ahluwalia 29:40
And how much time and end to end evaluation to wiring of money in that case?

Sudarshan Kamath 29:45
Like three, four weeks. And, and the only reason it took, it would have been done in a week, but the only reason it took three, four weeks or more was because we had money in India and then the whole, like the process with like India valuation, et cetera, took time. Right.

But if it was US funds, it’s just like extremely, extremely quick. Like I have, I know people who have wired in a week basically. Right.

So, so yeah. Sierra happened in July. I think the money came by September.

This deal we closed by December. Right. Like that’s, and so imagine like introduction to seven figure deal happened this quickly.

Siddhartha Ahluwalia 30:19
And by December it was a seven figure deal.

Sudarshan Kamath 30:21
Yeah. But by December, so I just closed like last month basically. So, so now it is, now there is a scale up period, et cetera.

So it’s a, it’s a like a multi-year deal with a scale up period. So the book, they are like the booked amount is seven figures, but yes, the payments et cetera will come over.

Siddhartha Ahluwalia 30:36
Yeah.

Sudarshan Kamath 30:36
Yeah. Yeah. But this year I will make six figures from them and yeah.

Over like next year over the next five years, I will be making multiples of these basically. So now going from like nothing to having a serious like conversation with an enterprise where you’re talking about these things is a very different place to be. And now we have multiple such deals that are happening at this point of time.

And we know what it takes. We went from a, we want to sell to everyone to like, we have a specific type of buyer. Like we should, our API should be integrated.

They should have used this much. They were, these are the valuation metrics to send to them. This is how we price.

This is how we competitively win the price deal. All of that sort of playbook we figured out because of that first deal that happened. Right.

And for us that set the bar to how deals should work and, and, and, you know now we are setting up a repeatable process sort of to make that happen. I think if we were in India, this would never have happened. Never happened.

There are probably 20 companies in India who can give us that kind of a deal size and they would spend an incredible amount of time and effort.

Siddhartha Ahluwalia 31:41
They would waste so much time. Yeah. You know, yeah.

And then did you raise any round after that?

Sudarshan Kamath 31:47
Yes. So we, we just closed the series A, we could call it a pre-series A, but like we closed with Seligman Ventures. So Seligman is their ex-partners at Thomvest.

They have invested in Cohere, Harness, and some of these companies and they are very seasoned. It’s a very large fund. They can support us in next round as well.

It was a preemptive round for us because we were growing quickly. So we thought let’s, let’s do it.

Siddhartha Ahluwalia 32:12
How did the conversation happen?

Sudarshan Kamath 32:15
So the deal started, you know, I think the deal started happening. And then also we got in between like an acquisition offer for like 150 plus million dollars.

Siddhartha Ahluwalia 32:24
September 2025, you closed Sierra?

Sudarshan Kamath 32:28
Yeah. Right. When did the acquisition offer happen?

A month after that, probably like October, September, October.

Siddhartha Ahluwalia 32:33
Yeah. And then when did Seligman?

Sudarshan Kamath 32:34
December.

Siddhartha Ahluwalia 32:35
So very quick cycle…

Sudarshan Kamath 32:36
Very quick cycles, right?
So, so there’s, there’s a, we got an acquisition offer. We got, we got this deal. And then not just this deal.

Now we have multiple other deals that are parallelly sort of already at six figures, scaling to seven figures. So we got an inbound saying that, look, you guys are obviously growing. From Seligman.

We want to do, we want to sort of, you know, preempt this. Let’s let’s work together. And so we basically ended up having a conversation.

We, at that point, we did not even have like a proper mechanism for booking our ARR, et cetera. So we cleaned all that up quickly. I remember I was in the middle of AWS re-invent.

AWS is an active reseller of Smallest, by the way. So AWS had invited us there and we were like doing a lot of one-on-one with customers. And then we ended up in between that, creating the deck, the whatever the financials, et cetera, required.

They are like, why is your ARR like so messy? But we are like, okay, we are like cleaning it up because you have India sales, you have US sales, you have partnership driven sales. You have taxation is different.

Us India, all of that now getting cleaned up. Now we are sort of going to like a formal CFO driven process. So we are getting a CFO.

But all of this happened in the last like three to four months.

Siddhartha Ahluwalia 33:47
Must have been crazy, right? You just closed around. Yeah.

There was an acquisition which you passed on and then you got a term sheet and then the term sheet close in a month or how much time?

Sudarshan Kamath 33:55
Term sheet close, like literally, like we just took a week to sort of negotiate on something.

Siddhartha Ahluwalia 34:00
From the first conversation. Yeah.

Sudarshan Kamath 34:01
The first, I mean, when we got the term sheet from Seligman, it took us a week. I mean, they obviously wanted us to sign them. I think we were kind of completely okay with most of the terms.

We just had to make sure we are on the same page regarding a few things, nothing major. We closed it in a week. And yeah, like that’s, that’s basically how quickly things moved.

Right. So I went back to India after that because we had to tell it to the team. There are a lot of people who have been there with us early on.

We want to make sure that with this fundraise, they are also rewarded. Their goals are understood, et cetera. So we did that.

Siddhartha Ahluwalia 34:30
But did you take secondary in this round?

Sudarshan Kamath 34:32
No, no.
I do. I will take a secondary at some point of time, but I feel comfortable once our, like process is like set and scaled right now. This I want to see how repeatable this process is.

As I’m going from a founder led to a more, let’s say GTM led sales motion. And if I’m able to sort of slightly take a step back and say, Hey, this is a machine that’s just working without me, then I will definitely take a big secondary. But right now I feel there is some work which I have to do to sort of make that happen.

Plus there is some, yeah, it’s just a personal thing, I guess. Like a lot of people do take secondary on series A. I felt maybe, maybe it’s too early for us.

Like, yes, we got a series A, but, but it was preemptive. We were going to do a hidden Q1. So and also I feel the company needs the money right now.

Siddhartha Ahluwalia 35:20
So you mentioned you’re also getting now offers for preempting again on top of that.

Sudarshan Kamath 35:24
And you are getting offers to do the next round. But this time, when we raised the round, we wanted to be a very well thought out big round with strategic investments and things like that. So we are speaking to some of the strategists, et cetera, so that we can not just get the capital, but also accelerate the go to market with the next round and build like a strong partnership driven thing.

Our model, business model can scale extremely quickly with partnerships basically. So, you know, like AWS, for example, all these clouds can be great partners. You have ISVs, you have all your BCGs, McKinsey’s of the world who can basically cross sell us.

So we want to get all of that going. And some of these guys to also invest in us and maybe some of the large voice players in the world.

Siddhartha Ahluwalia 36:05
Like Eleven?

Sudarshan Kamath 36:06
Eleven would not be an investor. I would say, I guess they are, they’re growing quickly, but they are still not that big.

Siddhartha Ahluwalia 36:13
But they are at 300 million ARR, 6 billion or 11 billion in valuation.

Sudarshan Kamath 36:16
11 billion in valuation. I think they would, they would definitely see us as competitors, if not anything.

Siddhartha Ahluwalia 36:24
But who is your competition? Do companies like BOLNA are your competition?

Sudarshan Kamath 36:27
BOLNA would be a customer to us. So it’s people who are building the voice models in real time.

Siddhartha Ahluwalia 36:32
But voice applications are not your customers.

Sudarshan Kamath 36:34
Voice applications are not your competition. So we help people building voice applications basically. Right.

So, so companies like BOLNA, VAPI, RetailBlend, all of them.

Siddhartha Ahluwalia 36:44
Is retail a customer?

Sudarshan Kamath 36:45
Yeah. We just, we are just getting into it. So, so far our GTM motion was very enterprise first.

We were just selling to like two customers, three customers to get like our ARR. Now we are opening it up to everyone. And as part of this strategy in this quarter, we will be listed on all of these.

Siddhartha Ahluwalia 36:58
And how big is the team right now?

Sudarshan Kamath 37:00
The team is close to 40 people right now.

Siddhartha Ahluwalia 37:01
Just 40.

Sudarshan Kamath 37:02
Oh yeah.

I mean, also this was not this much. So for example, when we raised our seed, it was 12 people, 10 or 12 people. So these 25 people have been added in the last three months, basically.

Siddhartha Ahluwalia 37:14
And how many India, how many US?

Sudarshan Kamath 37:16
10 in the US.

Siddhartha Ahluwalia 37:17
And this is all on the sales side?

Sudarshan Kamath 37:19
No, we have research leads as well here. We, because you know, in India you get young talent who is smart, but in US you get seasoned talent who have spent five, 10 years studying about a particular subject. And so we generally hire lead researchers in the US basically.

Yeah. So yeah, but the team is fairly young. It’s, theoretically speaking, most companies would not, I think, need a 40 member team at our stage, like at least in the US, I’ve seen lean teams, but training models is, and just doing research is like an.

Siddhartha Ahluwalia 37:54
So you get compared on your model efficacy against ElevenLabs? Yes. Who else?

Sudarshan Kamath 37:59
Cartesia is one who we get compared against, but there is some nuance to it. Like for example, Eleven has a lot of focus on content creation and like dubbing. And if you want to build like an Instagram video or like a AI movie, Eleven would be a great player.

We don’t focus on content at all. We are focused on the real time space. So very low latency.

Siddhartha Ahluwalia 38:19
What is the real time competition?

Sudarshan Kamath 38:20
Right now the only other player who’s focused on real time is Cartesia, but they are heavily dependent on state space models and we don’t train state space models.

Siddhartha Ahluwalia 38:28
Can you explain in layman terms?

Sudarshan Kamath 38:31
So state space models, the computation required to basically calculate the outputs is something called sub quadratic. What it means is let’s say you have, input sentence, which has like 10 words in it. So a transformer would take 10 into 10.

So a square of it’s a hundred that’s quadratic complexity, but a state space model would take like less than a hundred basically. That’s like the simplicity of it. It makes a very big difference when it comes to larger values of input text.

So for example, imagine you have 10,000. So if it is square, so it is 10,000 into 10,000, which is like a hundred million basically, versus if it is like a less than that. So it could be a hundred thousand or 1 million.

Right.

Siddhartha Ahluwalia 39:14
So how do you, let’s say this is one, but let’s say a customer has to consider.

Sudarshan Kamath 39:20
So that’s our biggest problem with state space models. Like we have, we believe like because of this sub quadratic computation, it actually does not compare all the tokens with each other by, by definition, and also does not have a great way to model long-term dependencies really well. And we’ve seen that happen like in our evaluations.

Right. So our strong belief is like state space models are not like that big a leap over transformers. And there is transformers do have a great base and we see transformers itself scaling, but the way the transformer is today used to sort of just put a lot of pre-training data in it and scale it up, that’s not how they’re supposed to be used.

A very interesting way to think about how, you know, attention and these things can be used is like these prediction-based models that Yann Lecuntalks about, JEPA, et cetera. In our opinion, like those models are actually closer to how we can build AGI than like state space models, which is just a play on compute capacity. How do I say this very simply?

The models that Yann proposes, the JEPA based models, they basically have a predictive view of the world, which means let’s say if I throw a ball at you, you’re going to probably move your hand here to like defend yourself, right? Because you’re predicting that the ball is going to come, right? But let’s say if you’re not able to catch it, it still hits your face.

You realize, Oh, I was wrong. And like, you’ll correct yourself basically. So you’re inherently predicting the next state of the world, given all the times frames till now.

And then if those states are incorrect, you do a correction of the model weights and learn from it, right? So that predictive view of the world is what we are missing in LLMs that we want to learn about. And so we are trying to build like a predictive view of voice and like conversations that we can have so that you can guess what the answer should be rather than how LLMs typically operate.

Like, that’s a very simple way to say it, but I hope that it’s a good answer.

Siddhartha Ahluwalia 41:24
But there’s an ex-enterprise, a telephony enterprise. Why are they choosing smallest again in layman terms or Cartesia?

Sudarshan Kamath 41:28
In like very simple terms for them, what it means is faster latencies, more parallelism. So on a single GPU, you can fit in a lot more of our models and better quality.

Siddhartha Ahluwalia 41:40
Do companies have to use their own GPUs or you also provide GPUs as a service?

Sudarshan Kamath 41:42
We provide our own like cloud services as well, but a lot of enterprises prefer deploying on their own GPUs because it allows them the capacity to scale up, scale down infinitely.

Siddhartha Ahluwalia 41:52
Let’s say the telephony company that you are working with.

Sudarshan Kamath 41:54
Yes, we are doing on-prem deployment for them.

Siddhartha Ahluwalia 41:56
So most of your customers today are on-prem?

Sudarshan Kamath 41:57
We recommend it because for us it’s 100% margins, right? If that happens. So we really like folks deploying it on their devices.

Siddhartha Ahluwalia 42:03
But then it’s so much bespoke support that you have to provide.

Sudarshan Kamath 42:05
No.
So the good thing is when we say on-prem, this is actually virtual private cloud. So it’s, it’s just, they have a like a Docker, which they just download into their servers and run it. It’s not us going to their place and like with a CD or a pen drive and like doing that.

We can do that. We don’t prefer doing that. But we, maybe we, that of course that will require some bespoke support, et cetera.

But but for the virtual private cloud, it’s very simple. They, as long as they have internet, they can download it and run it and we can support just like we support our cloud basically. So, and for us, we don’t have to manage GPU capacity.

The biggest problem is GPU capacity for customers scales at random times based on their usage, right? So you as a vendor can’t keep predicting when they’re going to scale up, scale down, but they can. So they let their team manage how they want to scale up, scale down.

Right. And, and the biggest issue is today, people end up pricing the models extremely expensively because they have to manage the scale up of GPU capacity. Let’s say someone says, I want a hundred concurrency.

So what that means is you will keep, let’s say, the number of GPUs always on that can manage a hundred concurrency, but you don’t want to do that. You only want it on when they’re actually at peak capacity and then reduce it otherwise, which is a very difficult problem to solve. So it’s just easier to give it to the customer and get them to run it basically.

So, so that works for us. And it’s also better for Infosec compliance. So when you go to banks, when you go to like financial institutions, you generally just want to make sure no data is flowing out of that system.

Right. So, so we can, we can provide that as well.

Siddhartha Ahluwalia 43:32
Got it. And so you’re saying that today you are at X revenue. Can you share some ballpark numbers on that?

Sudarshan Kamath 43:37
Yeah. It’s between one to 2 million ARR. Yeah.

And we’ve grown like, like 10 X in the last three, four months, basically.

Siddhartha Ahluwalia 43:44
So basically I assume you are at a 100K when you got Sierra and after Sierra, because you could deliver with the introduction that they could provide. It’s just scaled it.

Sudarshan Kamath 43:53
Exactly. Yeah. Yeah.

Siddhartha Ahluwalia 43:54
Do this process-wise, are you enterprise ready today?

Sudarshan Kamath 43:56
Yeah. Yes, we are. So we, with that introduction, we kind of learned a lot.

We got tons of feedback from like a single enterprise. And what our team did is basically, instead of like going to every customer and like doing requests, we said, give us all of your requests. And we want to make sure our model like just works in all of those aspects of it.

Siddhartha Ahluwalia 44:13
So our enterprise building custom voice application or do you go yourself and build those applications?

Sudarshan Kamath 44:18
They build custom voice applications. We don’t help them build custom voice.

Siddhartha Ahluwalia 44:21
You just provide the model.

Sudarshan Kamath 44:22
We just provide the models. We have an infra layer where they can orchestrate the model, but they do the prompting. They do the engineering.

We do have a forward deployed team that does sometimes guide people. But as we go higher and higher in enterprises, we don’t see this function scaling internally. We see someone like a Capgemini or an Accenture building like a practice around the product.

Siddhartha Ahluwalia 44:42
Are you enabling those companies with capabilities on Smallest like Accenture?

Sudarshan Kamath 44:46
Yes, we are. We are in talks with some of these. So they are the, the goal is to build a center of excellence where they can use these, this agentic platform with the models.

And let’s say they have a business process and some big insurance company that they have really good domain knowledge about. They understand the KPIs, the stakeholders, they take the platform, white label it, use it as an infra layer, add maybe some of their own things and then automate the whole process with us. So we can never do that because in my opinion, that requires deep verticalization.

And today we are not a deep verticalized company. We are a very horizontal infra provider company. And for that, you can’t be doing too many things.

So we need partners. 2025, 2026 is also going to be a very partner led go-to-market strategy, trying to build distribution through other people reselling our products.

Siddhartha Ahluwalia 45:31
Understood. So, so this year is basically, you know, are you going to invest more in product or just in the scaling?

Sudarshan Kamath 45:36
We, we are doubling our research team. So we are going from like right now we are 15 researchers. We are going to 30 researchers and purely because research gives you ROI in like six months.

So you can’t like hire someone right now and then expect if you have goals for 2025, 2026, like in September, you have to get those people right now.

Siddhartha Ahluwalia 45:55
And what are those folks researching on?

Sudarshan Kamath 45:57
A lot of different topics. We have a thesis as well on our website on how we feel like AGI will be achieved. But basically trying to do some more like ambitious projects and some more applied AI projects around small models.

Applied AI projects include like, for example, things like how could you build a text to speech model with even higher concurrency than what it, what it is before. So on a single GPU running more instances it could be, how do you get, make these voices laugh, cry, like become more human in multiple languages. Those are like applied AI projects more ambitious sort of research projects are basically for example, one topic we work on is asynchronous thinking.

So humans actually you are not waiting for me to finish my thought to start thinking. Right. So you actually start thinking the moment I say the first word.

But LLMs today wait for the whole context, you know, like you give the entire chunk and then they start thinking. So to enable this human like behavior in AI, you need to work upon asynchronous thinking. So the AI also starts thinking given the first voice that comes in.

And so that is one topic of research that we are working on. Continual learning is one topic, which is basically how can you get the model to keep improving as it has more conversations with the customer. So these are like few topics.

And then one, I think the topic I love the most is separation of compute from memory. So today LLMs are storing more and more information as they become larger in size. And it’s very difficult for LLMs to learn something new because the moment you train it on some extra data, that is forgetting of the older data, you know, data that it has trained on.

And it’s just super expensive, inefficient, et cetera, to run. The way we see the future is you will separate memory from intelligence. So the memory will be captured by infinite layers and the intelligence will be captured by finite layers of, you know, intelligence.

And that small model that has those finite layers that can basically be quickly trained through continuous learning, basically. So that separation of memory and intelligence is one thing we are working on. You must have seen a lot of companies that are coming up in the context engineering memory space.

They’re basically building the infinite memory. And we are focusing on the small models that power the compute.

Siddhartha Ahluwalia 48:03
So is it a winner-takes-all market? Or can Smallest be a 300 million ARR company, Eleven by the time you reach 300, they’ll be at a few billion and Cartesia will be at a few billion.

Sudarshan Kamath 48:13
I think the markets will diverge at some point of time. So for example, Eleven is sort of doubling down on the content space, but the real time space, like mimicking how humans do conversations, it’s like a very different problem than mimicking how movies are made. And so I think the markets, both are massive markets.

So I think this is definitely not a winner-takes-all and we don’t even know what winning means at this point of time. I mean, winning at the highest point is basically building like a her, you must have seen the movie, right? There’s like a very empathetic voice which basically powers every single conversation of yours.

So that could be winning. But, but yeah, it’s a huge market. Imagine like every single human in the world has an AI that can talk to it every single day. Like we are not texting this podcast, right? Like we are talking to each other.

And imagine like now you have an AI who you feel like having such a long conversation with. Today that does not exist.

Siddhartha Ahluwalia 49:06
I think but people have made AI enabled podcasts where you can talk to Steve Jobs.

Sudarshan Kamath 49:11
Yeah, that so that’s a good, good, I think market that has come up in like the last one or two years. But it is still a one way sort of street and you would still rely on that for very specific information getting sort of things. But it is not a very generalized AI that can basically like talk to you about it doesn’t have a personality basically, at this point of time.

Siddhartha Ahluwalia 49:32
Because you said it doesn’t have an infinite memory, like as soon as the compute goes up, the memory goes down.

Sudarshan Kamath 49:36
Yeah, exactly. It doesn’t have infinite memory. I mean, that is basically constrained to the information you’ve given that AI around the podcast, you know, like about Steve Jobs, if you read the book of Steve Jobs, it will probably just talk about that.

But what did it learn from that? Can it apply to its own long term goals? Does it have a long term goal?

Like it does not, right? So, so that’s a very specific, great use case for voice AI. And we will see many such use cases for voice AI pop up.

So voice as a market itself is going to be pretty big. But like I’m talking about like the gold standard of why this is like a massive, like a multi trillion dollar market is because you are basically gonna have every human as a consumer as this.

Siddhartha Ahluwalia 50:15
But then, if it’s such an attractive market, why not Anthropic and OpenAI playing in this market actively as ElevenLabs?

Sudarshan Kamath 50:21
I think there is a different focus. So some of our investors in like our three seed words from OpenAI, as I spoke to them, they are building a different kind of intelligence. And I think that’s an important form of intelligence as well, which is the large models that can basically take huge amounts of context, process it, and then come up with something meaningful out of it.

Think of it this way, like, when humans made aeroplanes, we tried to learn from birds, and we are doing much, much better there. We are actually flying long distances much longer than how much faster than what a bird can fly. But we are still not like as agile or light as like a bird, right?

So we have mimicked flying, but through a different form of instrument. Very similarly, right now in AI, we are doing LLMs, which is a different form of intelligence than human intelligence. So they have something and they are playing the sort of, let’s make LLMs bigger and bigger and smarter.

And that is a market as well. But to build human like AI, I think it has slightly diverged from that goal. Can they do it?

Yes. I mean, at any point, big companies can do anything, right? But I think it’s just about focus at this point of time.

Siddhartha Ahluwalia 51:27
So today, the scale that you are at like one to two, what does it require you to build like a hundred million dollar company? Time and money both?

Sudarshan Kamath 51:36
Well, very frankly, my predictions about how much money I need have always changed with time. I think that’s the issue with most first-time founders, right? Like when I had raised like the first 900, I thought this was enough to build a hundred million ARR company.

Over time, I realized, of course not, right? Like you need more if you’re doing pure research. Because one is actually having the product, but then the other is an enterprise trusting you that, oh, yeah, this is a serious company that can handle like n number of requests from us to sort of make sure it works.

Having said that, I still think the capital we have right now is enough to build a hundred million-ARR company. So I’ll not comment on capital, purely because we already have a great technology, which is selling at scale. And now it’s about go-to-market and like building that playbook so that other people can sell this product and distribute this product.

Now, very mathematically speaking, let’s say you want to do sales, right? So to reach 10 million ARR, you need 10 people doing 1 million ARR each this year, right? So it’s a very calculated.

If to do 100, you need maybe 30 people like this who are doing 30 million ARR, maybe a partnership model where they have 10 partners doing 5 million each, and then like some pay-as-you-go revenue, etc, right? So I feel like having calculated all of that and this ARR that will come from growing from 1 to 10 to 15 this year, that will fuel back into the company and basically act as the lever to sort of go to the next milestone. Having said that, one important milestone for us could be a strategic coming in who can accelerate our distribution.

Siddhartha Ahluwalia 53:09
Example?

Sudarshan Kamath 53:10
Example, Amazon, for example. Like if you are default on the marketplace, if you are default on the Google Cloud and Google starts selling you, for example, Accenture, let’s say they take your models to all of their clients and build agents, that can accelerate our ARR by quite a lot, right?

And so we want to build that, we want to make that happen.

Siddhartha Ahluwalia 53:27
And are you trying to make that happen right now?

Sudarshan Kamath 53:28
We are already in touch with some of these folks and we want to make sure that we do that. But those things take time because Accenture wouldn’t pick any random product. They want to make sure the product works.

It’s aligned with their internal goals. They have an internal team that understands the product. They have a roadmap of investment for the product.

They have identified the client for them. And then they sort of start.

Siddhartha Ahluwalia 53:45
So it’s like at least a 24-month cycle.

Sudarshan Kamath 53:47
Yeah. Yeah.

Siddhartha Ahluwalia 53:48
They have to themselves build a 40-member, 50-member team inside Accenture.

Sudarshan Kamath 53:51
Exactly.
I mean, if they do, generally, if they are able to do one sale quickly, then they accelerate that and they invest more. And so generally, it’s about that first win together, in my experience. And we are already working with some folks to get that first win, which are in advanced stages right now.

So hopefully, we’ll get that first win very soon. And that should accelerate some of these partnerships.

Siddhartha Ahluwalia 54:12
Got it. So right now, what’s your biggest focus for this year, 2026?

Sudarshan Kamath 54:19
We need to get the research right. Like, yes, we have the ARR targets, etc. And I think we can achieve that.

But the biggest miss can be if, let’s say, the models that we have in mind, we are not able to ship them on time. Right. And so we are focusing heavily on research to an extent where I personally know what everyone in the company is working on.

And like, I have my own thoughts around what the focus should be, etc. Right. So sales in this AI-first world happens naturally if the model works.

Siddhartha Ahluwalia 54:48
Like, why do you say that? Because today, enterprises are getting bogged down from top-down motions by Microsoft, Salesforce on AI adoption.

Sudarshan Kamath 54:56
I think it’s coming down to niches. Like, for example, if you look at image models, there are like few companies that people know are great at image. Like, for example, when stability AI was there or like some of this, I think, if I’m not wrong, Kling is doing pretty well in image and video recently.

These companies have a large word of mouth and just self-serve distribution angle. And then that self-serve leads to a lot of bottom-up usage in the company. And imagine like 100 people in a company are already using your product, then you can easily do top-down after that.

Right. So that’s going to be our strategy as well to make sure like the product is popularized and the model just works. Because if you have good word of mouth amongst developers, if people can just easily use your APIs, you can see cursor, for example, scaled through self-serve.

You can see a lot of new age companies are going from 0 to double-digit ARR in like three months by self-serve growth. Right. So I think the world of self-serve is coming up.

Enterprise will always be there, but it’s easier to do if you have a great self-serve product, in my opinion.

Siddhartha Ahluwalia 55:51
And what are the popular applications that are getting made using your models?

Sudarshan Kamath 55:54
People are building like open source, whisper flow alternatives. People are building customer support, voice agents. People are doing debt collection.

People are doing AI note takers. So very horizontal.

Siddhartha Ahluwalia 56:06
Can you take an example of a few companies that have built these applications? Like what company has built what application?

Sudarshan Kamath 56:11
So there is a like without taking specific names, but like there is a financial services company based in India, which is doing large scale debt collections using us like, you know, every day we are like at 100,000 calls plus using that.

Siddhartha Ahluwalia 56:22
And you build the application for them or they build the application?

Sudarshan Kamath 56:24
They have built the application. They are using the platform to do those calls. There is a telephony company that has all of their like meetings, internal meetings getting like recorded in US using our voice models.

There is a large customer support company that is very recently started using our voice models to power customer support agents. Right.

Siddhartha Ahluwalia 56:44
So something like kind of Sierra.

Sudarshan Kamath 56:45
Yes, exactly. Something like Sierra. So we are in talks with them as well.

So market is very horizontal. All of CX, all of telephony, all of financial institutions and more. And also B2C is like a strong thing as well for our market when you’re selling models.

And hence, all you have to do is make sure the model is really good. As long as it’s really good, it matches user expectations. It can handle multiple languages, accents.

It can do pretty well in the platform.

Siddhartha Ahluwalia 57:11
Understood. And what are the risks that you have today that can stop you from growing, both market risk and internal risk?

Sudarshan Kamath 57:20
Well, if there is an AI bubble and something breaks, first of all, I don’t think there’s a bubble.

Siddhartha Ahluwalia 57:24
Why do you say so?

Sudarshan Kamath 57:26
Because there might be micro bubbles.

Sudarshan Kamath 57:28
For example, I feel like some AI products which are like being sold are just sold in a very temporary period of time.

Siddhartha Ahluwalia 57:36
Example?

Sudarshan Kamath 57:36
Example could be like, there are like coding agents, like the top white coding platforms are amazing.

But then there are a lot of companies that are just trying to build these very specific coding agents, which apparently will be better than Cursor, Lovable. That’s not going to happen, right? So I feel like there are few winners in that space.

And the others are just going to die down. But they have been funded and they will probably go through some acquisition or something to continue, right? But the voice space is very real, because the voice space is an existing market that is getting disrupted.

You are not inventing a market. You already have a dictation in mobile apps, which sucks. You already have that desktop note taker, which was really bad.

Now with generative AI, you have much better dictation. You already have IVR based voice agents, which are really bad. Now with generative AI, it’s becoming very natural, right?

So you’re not inventing a new market. You are taking an existing market, making it much better. And by that, you’re also making it larger.

And so this is a very real market that is here to stay. But coming back to your question, can a large major global event take off funding from AI? I think that’s a reality.

Yeah.

Siddhartha Ahluwalia 58:40
For example, one of the large companies like OpenAI, you don’t know the internal risk because they are not public companies yet. One of these companies burst.

Sudarshan Kamath 58:48
Yeah, exactly.

Siddhartha Ahluwalia 58:48
And then the next day, it’s a bubble that’s burst. And all the funding that is happening in AI will stop.

Sudarshan Kamath 58:54
Yeah. I mean, a lot of the buying, let’s say, would also happen through when you do like pay as you go, happens through other startups or other developers who got money because of VC funding into other startups. And then imagine all of them kind of go bankrupt.

Then suddenly your customers are not. But that will take a massive major event. And today, at least it looks like folks are doubling down.

OpenAI is making progress. Enterprises are adopting AI. Revenue is growing.

So it does seem real. I am not a financial expert of global meltdowns.

Siddhartha Ahluwalia 59:21
But the thing is, we don’t know the internal gross margin. For example, if OpenAI is doing 2% GM and still need 100 billion more to fund it.

Sudarshan Kamath 59:31
Yeah, I’m not an expert. I think the collaboration between NVIDIA and OpenAI is interesting. I think with the latest model, they were talking about even providing feedback to NVIDIA on how the chip should be designed to provide better inference capabilities.

To give you an example, NVIDIA chips are generally very bad for voice inference. If you have 40 gigabytes of RAM in a NVIDIA GPU, you can generally, if you want to do real time, like 100 milliseconds, you can use maybe 2 or 4 gigabytes of it for our models. And beyond that, if you start scaling, the latency starts increasing.

And that is because in NVIDIA, literally, the physical distance between memory and the cores of the GPU is so high that the computation happens very quickly. But the time it takes for the next data to load from memory is more than the computation time. And so that becomes a bottleneck.

It’s a memory bottleneck. Now, there are other chip architectures that are coming up that are basically having the compute and memory literally in the same physical space so that you don’t have to go through these cycles of compute memory. And obviously, these are not at the scale of NVIDIA, but they are taking up some, they have some traction.

And we are also using some of these to reduce the cost of voice. Today, the cost of voice is like 1000 times more than the cost of text. We want to bring it at par, basically.

Siddhartha Ahluwalia 1:00:49
Got it. And what’s your, one is research that you said, but what’s your biggest ambitious thing that you are doing right now?

Sudarshan Kamath 1:00:57
So I think, obviously, I’ll not give you the research answer because every time someone asks me this, I go back to another research idea that I have. But we are literally creating the list of top 50 dream customers. And going aggressively through like a top-down motion, a very targeted motion of how could we get them through a Palantir-style approach.

Like we know every single thing about the customer. It’s not just who is the buyer, what do they do, what have they purchased before, what do they like, are they planning to buy, what is the problem we’ll be solving for them if they purchase from us, what are their incentives, all of those things in a very like a PhD kind of manner. You’re not doing like spread outbound sales where you’re like messaging like a lot of people on LinkedIn and figuring out some of them will work.

You decide these 50 accounts I want to sell to this year, and somehow I need to get in, right? And I can get in because the technology is good, but what are the other things they worry about? And I just have to make sure I know them.

And this is something like coming from a nerdy background that I have is very alien to me. But I have learned it, I think, over the last one year or so. And hopefully we’ll do a good job.

Siddhartha Ahluwalia 1:02:01
Can you share a lesson that might be helpful for the founders listening? Yeah. So some of the largest enterprise accounts that you cracked today.

Sudarshan Kamath 1:02:07
Yeah.

Siddhartha Ahluwalia 1:02:07
How did you crack them apart from introductions? What did you do specifically to make sure you land and win in those accounts?

Sudarshan Kamath 1:02:13
So I’ll talk about this US-based telephony company, right? So and I’ll name it later after this podcast as well. So we got an introduction through our investors.

And I was like, I had just raised the seed. The money was not yet in the bank. I was kind of getting on a call.

I was like, Oh, if this person does not like me, maybe the term sheet gets pulled. I don’t know. So I’m kind of like a little nervous.

But at the same time, I know my things. Like I know what I am building. I know the problems I can solve.

So we have this amazing voice model that converts voice to text called Pulse. It’s a speech to text model. It runs in real time.

64 milliseconds latency. A lot of other features like emotion detection, gender detection, bunch of other things, right? So it’s like 100 different parameters that you should know really well when you’re going to do the sale, right?

Get on a call with the CTO. The CTO is an engineer himself. He is like a public market CTO.

But he is still an engineer. He knows his technology, his teams in and out. So he grills me on like every single aspect of the product in the first call.

And the VC who is investing in me is actually on the call with me at this point of time. And both of us are getting grilled together basically. But I don’t take these things personally.

Like you’re asking me a technical question. If I know the answer, I’m going to give it to you. I’m going to explain to you where we are strong.

I’m going to tell you where we are not. So I’m very like honest when I’m selling to just be sure that your expectations and my reality is matching. So that call happened.

One hour call. We both were sweating after the call. And I felt it went okay.

But once you’re grilled, you’re not really sure, right? Because someone is not being like polite to me. But I just kept pushing that.

Hey, let’s create like a channel together. Let me give you the models. Okay, nothing happened 2 weeks, 3 weeks.

Then suddenly, after week 4, they’re like, okay, we are evaluating the model. So give us the model. So we gave the model.

They evaluated it. The benchmarks came really, really well. We were outperforming their existing provider.

And we were like cheaper, faster, everything. So they’re like, okay, we are interested in doing the deal. Right.

And then from that point onwards, we went into pricing. The pricing negotiation was interesting because I did not know what’s the benchmark, how to negotiate. Do I have any?

I mean, they are a startup. I’m a startup. They know that if I get this contract, my startup would be valued 3x higher.

So they know the leverage they have on me. But what I valued is they never pulled that trigger on me. Like here, they were like, they know that they’re like, okay, if you deliver this, whatever your series happens, etc.

That’s great. Plus, we’ll invest as well.

Siddhartha Ahluwalia 1:04:42
Did they invest in the series?

Sudarshan Kamath 1:04:43
They are. Yeah, we are in talks for that. So yeah.

So it went from them like grilling us to the next conversation having… They were like chatting with us. The CTO was really happy.

Just yesterday, I met him in his office. We spoke for two hours about how the company grew. And like, what are the problems they were facing with speech?

Who’s working in their team? How are they working? What is their roadmap?

How else can we help? Like just generally being friends at this point of time. I think this is not a typical thing that people might face with every company you go to.

But this is like his personality, right? And I respect that, right? First, you check whether someone is capable.

And then you end up spending…

Siddhartha Ahluwalia 1:05:24
First, they grilled you for one hour. Tested your capability.

Sudarshan Kamath 1:05:26
Yeah, and now between that time till now, they had a list of requests that we had to make sure we are fulfilling for them. And we were shipping those updates every two days for them. The entire team was working like including my co-founder on it.

Like literally hands-on reviewing the code. And again, this enterprise. So you can’t like ship half as stuff.

You have to make sure tests are running. So we made sure we understand their whole org. Who understands what?

How are they testing? What are their priorities? Delivered all of that and more in this two to three-month process, which led them to sign the contract basically.

And then the contract negotiation was just about… I really like this because they were very honest about it. I told them, look, I want you to be at least a million-dollar account for me.

It has to be. You guys are obviously getting a great deal. But I want this to be a million-dollar account.

Siddhartha Ahluwalia 1:06:14
A million-dollar multiple year or a million-dollar single year?

Sudarshan Kamath 1:06:16
Single year. So he said, okay, so we can’t do like this year because XYZ reason. And he was very transparent about it.

He was showing me the reasons why I can’t do it this year. But he said, yes, your technology is important. And we think this year, your technology will scale in the company.

So let’s do like a little north of half a million or so this year. From next year onwards, and we’ll also do like a 5-6-year contract where we’ll start giving you more and more money.

Siddhartha Ahluwalia 1:06:38
And what made them confident? Because in fact, they were an enterprise, right? So they could have easily given you a one-year contract and walk away with it.

So what made them confident on signing a five-year deal?

Sudarshan Kamath 1:06:47
So I think it’s not easy for them to switch technologies very quickly at their scale, right? Imagine like a very large set of products are getting powered by this. Now it’s a generative AI model.

So it has non-deterministic behavior, right? You want to make sure that non-deterministic behavior is handled and taken care of across all the downstream products that they are building. Now if something happens and if you are changing the model and some other downstream product breaks, which let’s say gets into the news because it’s a public market company, that could affect their stock price, right?

So that was one of the reasons why they’re like, okay, if you are in, then I want to make sure you are in for long term. But it’s not like they didn’t put in their conditions in front of us to before giving us this deal, right? They do have an initial scale-up period with an exit clause.

Then they have a non-exit clause after that for next one year. Then it’s like in between, they have some exit. But that’s fine.

Again, you’re dealing with a first multi-year, multi-million deal.

Siddhartha Ahluwalia 1:07:48
And you’re also learning a bit.

Sudarshan Kamath 1:07:49
You’re also learning.

So you would do anything to do that. And the biggest thing, more than the ARR, more than whatever, is now you know your model works at scale for a very important enterprise.

Siddhartha Ahluwalia 1:08:00
So maybe a Fortune 1000 or Fortune 500.

Sudarshan Kamath 1:08:02
Yeah, exactly. Now you can go to two others and very confidently say this works, right? Very similarly, we are working with a similar account in India, again at similar stage, etc.

Same thing, right? Just like, obviously, initially, there was some storytelling, there was some feedback, etc. But what I love about my team is they’re so quick to respond, take ownership, and just deliver.

And think of things as problems and solve it. Other people, generally what would happen is you would start worrying, oh, this is a problem. How do I solve it?

But here, you’re basically thinking of everything from first principles. And you’re saying, okay, this is what they need. Let’s figure out a way to solve it.

And we go right to the level of the NVIDIA architecture to even write our own kernels. So we are not scared of touching any part of the stack. Versus, let’s say, you don’t have the models, then you’re like, oh, this is the model limitation.

There’s nothing we can do. We are not. We train the models.

We can evaluate differently. We can optimize the models for various hardwares. Every single thing is built in-house.

And that’s, I think, our super problem.

Siddhartha Ahluwalia 1:09:03
Well, Sushant, I would love to do a part two sometime. Obviously, we are limited by time.

Sudarshan Kamath 1:09:06
Yeah.

Siddhartha Ahluwalia 1:09:07
But it’s so amazing to see cutting-edge innovation coming from India. And with such a high ambition. I think what you are building, if things fall right in place, then today from India, we give example of companies like Postman.

I think we might give example of Smallest in a couple of years.

Sudarshan Kamath 1:09:24
Hopefully, soon. Yeah.

Siddhartha Ahluwalia 1:09:25
Thank you so much again for doing this podcast.

Sudarshan Kamath 1:09:27
Thank you so much for having me.