March 2025 was a busy month for IT entrepreneur Bakytzhan Dos. First came the news that GoatChat.ai, a project he had been developing together with Yerbol Kopzhasar, had been sold to Spanish media holding company Newry Global Media. Just a few days later, Bakytzhan announced that he had raised $5 million for his new project, Nace AI. The company is developing a system that creates highly specialized AI models for specific business tasks. Over the course of this year, Nace AI has already closed a $21.5 million funding round, signed a contract with a Big Four firm, and agreed to work with Yahoo! co-founder Jerry Yang and Intel CEO Lip-Bu Tan.
In an interview for “100 Startup Stories from Central Eurasia,” a joint project by Digital Business and Astana Hub, Bakytzhan spoke about what it took to secure the first $5 million in investment, how small language models (SLMs) differ from LLMs, and what value they can bring to the enterprise segment. We also talked about how difficult it was to land a contract with a Big Four firm, how much the startup is already making, and when it could reach a $10 billion valuation.
“I got turned down a lot because, back then, no one really understood the power of small language models”
– Almost immediately after the GoatChat.ai deal, you announced that Nace AI had raised $5 million from General Catalyst. When did the idea for the new project first come about?
– While I was still working on GoatChat.ai, I started digging into different language models. I realized the current architecture has some fundamental limitations that make it difficult for these models to generalize broadly or handle long, meaningful contexts. This led me to a niche I found really interesting: small language models (SLMs). I saw a lot of potential for them in the enterprise space. They could help companies get real value from AI instead of just burning through tokens without much to show for it.
Architecturally, SLMs are actually quite similar to LLMs, and the way they work isn’t fundamentally different. The key difference is scale and specialization. For example, Claude or ChatGPT can have anywhere from 1 to 3 trillion neurons, while small language models have around 10–15 billion and are designed for much more narrowly focused tasks.
What does that look like in practice? If you take a trillion neurons in a large language model, around 85–90% of them are essentially storing information from a huge library. Things like when Einstein was born or the name of the first dog to go into space. In other words, only around 10–15% of the neurons in an LLM are useful for business. Our idea was to give companies exactly that part. For example, a model that can handle customer support, write SQL queries, and so on. So the main advantage of small language models over large ones is that you strip away all the noise and keep only what is needed.
– How quickly did you manage to close the $5 million round?
– Once it became clear that the GoatChat.ai deal was close to completion, I started looking for investment for Nace AI. By then, we already had a prototype that could create a small adapter for an SLM tailored to three use cases: customer support, SQL queries, and compliance.
I spent four months talking to different funds. I got turned down a lot because, at the time, very few people really understood the potential of small language models. Investors asked a lot of strange questions simply because they didn’t have much technical expertise in this area. So we decided to focus on investors who really understood the space.
General Catalyst was one of them. I pitched the product to their CEO and managing partners. They liked what we were building and decided to invest.
– Sounds pretty easy…
– It was actually a tough road. Before I secured the investment, I spoke to 40 different funds, and every single one turned me down. It was hard.
Another problem is that in Silicon Valley, no one really gives you proper feedback. You have no idea whether your pitch was good or bad. Investors just say, “It’s okay, fine.” At first, I found that really unusual.
Getting to General Catalyst took a lot of sweat and tears. But even if they had said no, I wouldn’t have given up. I came to Silicon Valley with one goal: to build a company here. And no one was going to stop me.
“Right now, large language models are basically using a sledgehammer to crack a nut”
– What does Nace AI look like today?
– Our core product is MetaModel. It’s an architecture for creating small language models that incorporate a client’s internal policies, compliance procedures, and corporate knowledge. In practice, we bring our small model to a client, and they might say they’re working on an AI agent for a specific task. We then build an adapter that fine-tunes the agent to the company’s policies and particular use cases. Put simply, MetaModel is a technology that lets companies quickly customize and personalize small language models.
We have four small language models of our own. The first versions were built using a mix of existing architectures from Google, Meta, and Microsoft, which saved us a lot of money. There also isn’t much point in building an architecture from scratch when so much work has already been done. So we took the parts we needed from several LLMs, stripped out what we didn’t need, and added a large layer of our own knowledge on top. The key was choosing the right data so the models could work effectively for the kinds of requests enterprise companies have.
In the end, you get a mix of datasets that needs to include both public and academic data, such as basic arithmetic, to keep the model from degrading. But the most important part is specialized data. We bought this data from different companies and providers, spent quite a lot of money on it, and then filtered it. Around 60–70% of the dataset needs to consist of this kind of specialized data if you want to build an SLM tailored to a specific use case.
– What specific tasks is each of your four small language models designed for?
– The first one is mainly designed for working with documents. A lot of large corporations still handle documents as PDFs, scans, and so on. We built a model that scans all of this material with a high level of accuracy and prepares it so other models can process it and perform specific tasks with it.
The second is designed for fact-checking. For example, if another model says a document hasn’t been signed, our model goes and finds that document so it can visually show whether that claim is correct or not.
The third is trained to understand finance, because current models aren’t very good at it. For example, they often don’t know basic accounting principles and other important details.
The fourth model specializes in choosing the right tools for different tasks. Right now, large language models are, figuratively speaking, using a sledgehammer to crack a nut. We trained ours to use a hammer instead.
By the way, we built the fourth model entirely from scratch once we had more funding. It runs fully on our own GPUs and was trained on our own data.
“Convincing accountants to give you money isn’t easy”
– Not long ago, Nace AI started working with a Big Four company, signing what you described as a “long-term contract.” What helped you secure the partnership?
– They saw a lot of potential in our technology for financial audits, tax return preparation, and similar tasks. It performed much better than the products they already had and the solutions offered by competitors. It’s pure competition: the best solution wins. We showed in practice what our technology could do, and after that they said, “All right, let’s work together properly now.” And our pilot wasn’t free, either.
It may sound simple, but we pulled off something that felt almost impossible. Convincing accountants to give you money is no easy task. We put a huge amount of work into it. The deal took 10 months to close, and we probably had more than 30 meetings along the way. But in the end, it worked out because I was determined from the start to make the deal happen.
– Who else are you working with?
– Right now, we’re focused on service companies in the financial sector, including banks and lending institutions. Basically, anyone whose business revolves around money. I see a real opportunity to build a profitable business model in this segment and deliver strong value to our users.
Right now, we have three full-fledged clients, with two more in the onboarding process. But it’s important to understand that sales cycles in the enterprise segment are long.
– How do you monetize the product?
– AI companies don’t have that many options when it comes to monetization. Since we sell small language models, there isn’t much money to be made from inference, unlike some other companies. (Inference is when a trained neural network applies what it has learned to new, real-world data to produce an output — Digital Business.)
A big chunk of our revenue comes from updating the models, because they’re working in an environment that never stands still. Take the Big Four example. They operate in 70 countries and have 35–40 global offices. Say the regulations change in one of those countries and a certain process now has to be handled differently. We then update the model so it works according to the new rules.
There’s also a customization side, where we build company-specific data into the model so it can behave more like the owner or the board of directors.
– What’s Nace AI’s ARR right now?
– I won’t share the exact numbers, but I can say our revenue is already at a strong level. We’ve generated cash flow for the company that is roughly on par with the amount we’ve raised in funding. (To date, Nace AI has raised a total of $26.5 million — Digital Business.)
“In many countries, investors waste a lot of founders’ time by asking the wrong questions and digging into things that just aren’t useful”
– Let’s talk about funding. Not long ago, you raised a $21.5 million round led by General Catalyst, and you managed to close it in less than 36 hours. How did you pull that off?
– There’s a bit of a rule in Silicon Valley: if investors really know the space they’re investing in and see a founder who’s ready to push ahead with the product, things move very, very quickly. It could have taken even less than 36 hours. The funds simply needed time to call people I’d worked with before and get a sense of what I was like as a person. That’s what took the 36 hours.
I actually think all venture deals should move that quickly. Unfortunately, in a lot of countries, investors take up far too much of founders’ time. They ask the wrong questions, ask things that aren’t really necessary, promise funding and then drag the process out for ages, or simply disappear. In my view, once you’ve made a decision, you should move quickly and close the deal. It’s in the investor’s interest too, because a good deal can easily attract a better offer from someone else. So it works both ways.
By the way, our round was oversubscribed. We originally planned to raise $18 million, but a number of other investors also wanted to take part. So, rather than leave anyone out, we gave smaller allocations to several companies and individual investors. That brought the total to $21.5 million.
– You mentioned earlier that you planned to use the funding to grow the team. How many people are at Nace AI now, who have you hired, and what roles are you still looking to fill?
– We have 30 people on the team, and 90% of them are engineers. In fact, we don’t even have a proper sales team at Nace AI yet. We’re only just starting to build one.
As for open roles, we’re looking for product managers, technical program managers, and growth managers. Essentially, we need people who can help drive the company’s growth. We’re also hiring actively in Kazakhstan. We see a great combination of energy, talent, and strong English skills here. We also have a subsidiary in Astana and are a member of Astana Hub, which gives us certain advantages as well.
– Is the US your sole focus right now, or are you planning to expand into other markets as well, such as the EU or MENA?
– We’re seeing a lot of demand for the product in Saudi Arabia and Southeast Asia. We’ve also had quite a few banks from Kazakhstan reach out. But expanding into those markets takes time and attention, and we simply don’t have enough of either right now. So for the moment, we’re 100% focused on the US. I think once the company reaches the Series B stage, when international expansion usually becomes part of the agenda, we’ll come back to those markets.
“We’re not at a billion-dollar valuation yet”
– Are you worried about competition from major players like OpenAI or Anthropic, which could move into small language models at any point? How prepared are you for that possibility?
– We live in an AI world where anything can happen at any time. But right now, we know for sure that OpenAI has decided not to pursue this direction. As for Anthropic, they’ve never really shown much interest in this space. Their focus is on larger LLMs trained on ever-growing amounts of data.
There’s also Google and Meta, which are interested in this area, but for us they’re more of a tailwind. In a way, we benefit from what they’re doing because they mostly work on small models as an open contribution rather than as a core commercial business.
Microsoft is probably making one of the biggest bets on SLMs and competing with us head-on. They still have their Phi family of models. But ultimately, the winner will be whoever can deliver the best accuracy and the strongest personalization.
Yes, there’s a risk that a major player could come in and take our market share. Is there a way to stop that? No. That’s AI. Right now, what matters is how fast you can iterate and how clearly you know where to focus.
– Have any of the big players tried to acquire Nace AI?
– We had two offers, and we said no to both. The first came from the Big Four company we ended up signing a contract with. When you start talking to a company like that, they’ll spend the first six months discussing an acquisition with you. Because of their internal regulations, it’s often easier for them to buy a startup than to set up a partnership.
The second offer came from a fund. They never gave us a final number, but the figures being discussed were nowhere near what I believe Nace AI could ultimately be worth. In my view, the company could be valued at more than $10 billion if we execute well.
– A lot of people in Kazakhstan are already calling Nace AI a future unicorn. How far are you from that milestone today?
– I don’t want to speculate about that. But I can say we’re not at a billion-dollar valuation yet.
– What are your plans going forward?
– They’re pretty straightforward. We have three main priorities: grow the number of customers, increase ARR, and raise our next funding round. And we’re going to do it in exactly that order.
As for fundraising, it’s hard to talk about exact numbers at this stage, but we want to raise significantly more than we have so far. We now need serious resources to build out the infrastructure. That’s the next goal, because supporting large customers requires much more infrastructure capacity.