Ladies and gentlemen, first up we have a very special startup that we're featuring on the show, that we're putting the spotlight on, and that's Fractal AI. Well, it's a leading unicorn in the AI space. It's made waves with the release of Fathom R1 14B, which is a 14 billion parameter open-source model purpose-built for reasoning. What's drawing major attention isn't just the model's performance — in fact, it is outpacing Meta's O4-mini-low in self-consistency tests — but the efficiency of the model as well. In fact, the major point that really makes it stand out from the rest is that it's priced at $499. Now, with open models becoming more and more viable for enterprise-grade reasoning and performance inching closer to closed-source joints, is this the turning point for accessible AI? How does Fractal see developers and enterprises adopting Fathom R1 14B in the wild? We discuss all of that and more with the co-founder of the company, the CEO of Fractal, and of course he's also an adviser to the India AI Mission, Mr. Srikanth. Welcome. Thank you very much Srikanth for joining us here on ET Now. Appreciate you taking the time out and speaking to us right here. Now, of course, since we're talking about your new model — you've just recently launched that — it's an open-source reasoning model, and I was just talking about the pricing of this model as well, but how else is it different from the other models that are available in the market today? And how did you design the pricing strategy for this?
Thank you, Sumita. First of all, there's a big difference between pre-trained models and post-trained models. So Fractal's Fathom R1 14 billion parameter model is a post-trained model, which means that you are actually taking a model and it is going to think at the time of you ask a question. Pre-trained models are like high school students who have read every book in the world and can answer any question very quickly. Post-trained models are designed to think when the question is asked. They're able to reflect on their thinking and they're able to do more mathematical reasoning and things like that. And that's really the idea of a post-trained model. So Fractal built a post-trained model and the cost of training the model was $499. It's free to everyone. Anybody can use it. It's open-sourced, open-weights with open datasets. So anyone can download. I mean, it's just been about 10 days since we launched it. There are more than 13,000 people who have downloaded it already. So it is really getting some traction because we have been very efficient in terms of how much training compute went into it. You may have heard of billions of dollars of compute being used to train other models. So this is much more efficient. At the same time, it gets you O1, O3 performance — O1-mini and O3-mini performance — and it even reaches the O4-mini levels as well. So that's the special thing about this specifically. It is good at answering mathematical questions.
Right. It's a very interesting point that you've made, Srikanth, and the fact that it thinks as you ask it a question. So I want to understand from you — as a company that has other AI offerings as well, including coaching tools, text-to-image tools — which segments are you specifically targeting this particular model to, your new model that you've just released, and what are the growth expectations that you have from this model versus the other offerings that you have?
Yeah, the objective of Fractal in doing this is to actually contribute to the India AI Mission. India AI Mission was launched a year and a half ago with 10,000 crore budget and so on. And we thought that the real issue here is: can India build a model that is state-of-the-art and can exceed the state-of-the-art as it exists today? And that's why we thought the real opportunity for India is to do post-training rather than pre-training. So we decided to build this model. We actually submitted a proposal to India AI Mission. We sent two proposals. One proposal in March was about building a large reasoning model. We called it a large reasoning model which could be larger than 14 billion — it could be something like a 70 billion parameter model. And then we also said in an April proposal that we could build a medical reasoning model which could do specific questions on healthcare and so on. India needs healthcare reasoning. We don't have enough doctors. We don't have enough radiologists. So we thought that would be another option. But the idea behind building these models was to show that India can build these AI research products and then put it out in the open and can actually meet the state-of-the-art. So that's really the objective of doing this. We have, of course, a very significant business serving Fortune companies around the world and also building several B2C AI products. This is our research offering that we have released. This also is a proof point that we can do something like this very efficiently in many other situations as well.
Right. So you are focusing on research — specifically research and development — with respect to this particular model. And you know, the several experts whom we've spoken to, they have specifically focused on the aspect of research when it comes to development of AI models in the country. And that's a good start that's been made. But Srikanth, when I was talking about your other models — when it includes the text-to-image model, the AI-driven healthcare model — what are some of the future models that you'll be working on, and what are the growth expectations that you have from each of these segments? How are some of these segments really performing for you as a company? If you could give us a comparative analysis.
Yes. We released, first of all, we released an AI coaching tool called MarshallGoldsmith.ai, which takes the coaching advice of Marshall Goldsmith, who's a legendary executive coach, and puts it into a hallucination-free coaching adviser that's available free of cost to everyone. What we have seen is that it has become such an interesting, important tool that many HR departments of various large companies have asked us to make it their adviser tool inside their companies. So we have done that for some of them as well. Secondly, we have a tool called Kaleido.ai, which is a text-to-image model. Again, it's free to use for consumers. What we have seen is this ability to convert text into images is very helpful in, let's say, real-time advertising. So many of our big enterprise clients are using it behind the scenes to do personalized advertising and really do the kind of work that agencies do — take six weeks or so — and crunch it down to an hour or two hours. So that's really where we are seeing massive traction. We also did Via.ai, which is a healthcare AI product. Again, it's free to everyone to use. And the idea is it will give you answers — you can ask a question on any medical question before you meet a doctor or after you meet a doctor. It'll give you really sound advice and medical information and prepare you for a doctor visit and let you know about the side effects and so on. Again, it's free for everyone to use. It's at Via.ai. But what we are also doing is using this kind of information to see if we can, for example, do it for pets. We work with several pet care companies and can we use something like this for a pet use case as well. So this is the idea of Fractal AI research — to build these business-to-consumer AI products that is a good proof of what we can do, and then use some of that capability in also serving some of the big companies that we serve today.
Right, Srikanth. But you know, the big question when it comes to conversations around AI is: how are you looking at reducing bias? Because, you know, this is of course from the very early conversations that started around AI, but it's still very, very relevant given the fact that AI models could be prone to bias. So how are you making sure that bias is reduced and eliminated?
That's a wonderful question. This field is called responsible AI, and what essentially you have to do is you have to make sure that the data that you're using is not biased, the way you define the problem — what you call as a dependent variable — that has no bias. And after you build the model, you have to validate to make sure that the way it is working, it's fair, it's transparent, it is explainable — things like that. So there are a whole set of principles. We have a set of eight principles which are responsible AI principles. What we do is after we build any product, an independent team runs the responsible AI checklist — a set of multiple tests — to make sure that the model is safe and responsible. And for example, when we did Kaleido, we had 43 different checks applied by an independent team, and then when the score is green, we know that is responsible enough to be released. So whatever Fractal releases has this responsible AI certification — internal certification — in it. And we've also been advocating to the Government of India to have a responsible AI framework for the nation. And as I'm also the NASSCOM Vice Chairperson, as NASSCOM we have consulted with the Indian government in creating a responsible AI framework for our country. And the idea is, if we have a test — a toolkit — which any tool can go through, then the products that come out of that are likely to be much more responsible than otherwise.
Right, Srikanth. Let's now switch our focus and look at global developments and shifts that are happening. How much of an impact, as a company, are you facing with respect to these headwinds — given the fact that these policy headwinds, particularly if you see the tariff policy introduced by the United States — how much of that is impacting companies, some of whom are your clients as well? And what kind of a shift in spends are you seeing in the near term due to these disruptions that have been caused by some of these policies that I just spoke about?
Yes. The long-term impact of policies is something that one can quantify and see — there may be something good coming out of that. But in the very short term, a flip-flop on policies means that there is massive uncertainty. And whenever there's uncertainty, typically companies around the world freeze — they take the worst possible assumption in mind and then they take those kinds of steps. They might fire people, they might cut down spending, they might delay investments and so on. So when that happens, everything slows down, and therefore you may have a stagnation plus a recession — stagnation plus inflation — which is called stagflation. So this is the kind of scenarios that may happen if the tariff policies come into effect in the very short term. And because the tariff policies are going up and down, companies are in a state of sort of freeze mode watching what's going to happen, in which case some of the spends also get frozen. So whenever companies around the world suffer, everyone suffers in the process. So that's just the nature of the beast. Having said that, I think some of the policies have moved in the right direction. Some of the early announcements that were made by the administration have been unwound, and therefore I do believe that overall things may work out okay and we might avoid some of the most painful scenarios that were possible when the announcements were made on Liberation Day last month.
Right. So, Srikanth, you say that the long-term disruption — the long-term impact of these disruptions — is yet to be seen and eventually you're hoping that everything will fall into place. But what's the near-term outlook that you're looking at, and how is that then impacting your deal pipeline?
So far, you know, we are experiencing a massive tailwind because of the excitement around AI. One of the big issues of AI is that while there's so much excitement about AI in the consumer world — for example, ChatGPT mentioned that they have 800 million active users. We've heard from Gemini that they have something like 1.5 billion AI overviews every month in the Google search — the AI overviews, 1.5 billion of them — and they have probably about 400 million active users of Gemini. Microsoft announced that they're processing 100 trillion tokens in a quarter, which means that that many times the model has been used or that many words have been used in asking questions. Gemini on the other hand is 50x larger than that — 50 trillion tokens. So everything shows that the consumer world is using AI in a very big way. But that's not yet translating into enterprise use because there are concerns like safety, responsible AI, there are privacy and other compliance-related considerations as well. So we see a massive opportunity to translate AI excitement into enterprise performance. And companies around the world — whether it's chief AI officers, CEOs, CIOs — they're very interested in making sure that they're using this in order to drive performance, which is higher revenue growth, higher productivity, and higher profitability. These are the kinds of things they're looking at using AI for. So there is massive tailwind for AI. On the flip side, we have the tariff-related headwinds as well. The balance of this is still a significant tailwind for AI companies as of today. So we are in a good spot in terms of leveraging these tailwinds and building a strong business for Fractal.
Right. And which segments are driving demand when you talk about the tailwinds with respect to AI? If you could throw some light on that as well.
