Mongolia’s AI journey shows how developing nations can achieve digital sovereignty through local problem-solving and sustainable growth
When global researchers predicted that speech recognition for low-resource languages wouldn’t be commercially viable until 2030, we had already achieved 97 per cent accuracy in 2020. This isn’t a story about one company – it’s about how developing nations can build AI sovereignty by solving real problems, building infrastructure patiently, and getting creative with talent.
After 25 years of building language technology in Mongolia – from basic character rendering to complex reasoning systems – I’ve learned that the path to AI independence requires three things: sustainable development, pragmatic focus on real missions, and a bigger vision. Here’s what actually works.
Start where problems challenge society
Silicon Valley starts with solutions looking for problems. Developing nations must start with problems demanding solutions. Mongolia’s AI journey began with frustrations that might sound trivial to outsiders but were paralysing daily life:
The typing crisis: We literally couldn’t type in our traditional vertical script. Eight centuries of written culture were becoming digitally extinct. This wasn’t a market opportunity – it was cultural death in slow motion.
The dictionary that wasn’t: No comprehensive digital Mongolian-English dictionary existed. Students and professionals relied on dusty paper editions from the 1960s. Every missing translation represented a lost opportunity.
The spellcheck absence: Mongolian text had no spell-checking. Government documents, business contracts, and academic papers were riddled with errors that undermined credibility and caused actual legal disputes. That was one of the biggest social problem until 2017.
The language corruption crisis: Here’s something that keeps me awake at night – we’re raising a generation that speaks broken Mongolian. Young people unconsciously rely on Google Translate for homework, social media, and daily communication. But Google Translate’s Mongolian support is so poor that it’s literally corrupting our language. Students write essays in grammatically incorrect Mongolian, thinking it’s proper because “Google said so.”
The speech recognition gap: Parliament and government operate through meetings, making transcription a critical function. Without speech recognition, protocol-keeping consumed enormous human resources. And solutions had to be on-premise for security.
The service crisis: Banks and telecoms haemorrhaged money on customer service. Without automated Mongolian language support, every query required human agents.
The procurement corruption: At Erdenet Mining, Mongolia’s largest company, manual tender evaluation enabled systematic corruption. Millions disappeared into rigged contracts or excessive delays in purchase completion.
Each problem we solved provided data, experience, and credibility for the next challenge. Our fully automatic tender evaluation system at Erdenet didn’t just save money and reduce corruption – it generated the revenue and trust that funded our speech recognition and language model research.
This problem-first approach built our capabilities organically. By the time we tackled large language models, we had accumulated 20 years of linguistic data, domain expertise, and customers who trusted us.
Harness your diaspora’s expertise
Mongolia has more software engineers in Silicon Valley, Berlin, Seoul, and Tokyo than in Ulaanbaatar. This brain drain, common across developing nations, became the cornerstone of our talent strategy.
The diaspora advantage
We couldn’t compete with FAANG salaries. Instead, we offered something money can’t buy: the chance to build their homeland’s technological future. Our pitch was simple: “Your skills + our mission = your legacy.”
AI Engineers from Germany, Switzerland, Italy, and the US were invited to work on-site, fostering knowledge exchange with local talent. They mentored, contributed code, and validated technical decisions, creating a sustainable talent flow instead of a one-way brain drain. This approach assured young engineers that they could venture out, gain experience, and return to impactful work.
Infrastructure has been our greatest challenge from day one. We move slowly, but we move forward. Our datacenter evolution tells the story of patient building:
We started in 2017, purchasing two gaming GPUs (RTX 2090 Ti). By 2019, we had received angel investment to buy our own L40S GPUs. In 2021, we added A6000 units. Today, we run our own infrastructure with local hardware for sensitive data. We use cloud resources from AWS and Google to train bigger models.
We manage everything ourselves – not by choice, but by necessity. No managed services exist for our use cases. This forced us to develop deep infrastructure expertise that later became a competitive advantage.
The lesson? Start with whatever computer you can access. Our first speech recognition model was trained on two gaming GPUs for one year. Perfect infrastructure is a luxury, developing nations can’t afford to wait for.
Competitive advantage
Cultural localisation
The assumption that AI models can simply be “translated” fundamentally misunderstands how language and culture intertwine.
The mixed language reality
Mongolians don’t speak “pure” Mongolian. Never have, never will. Real conversations flow like this:
Mongolian grammar structures with English tech terms embedded
Russian expressions from the Soviet era sprinkled throughout
Chinese trade phrases when discussing business
Our models had to understand “Kodoo push hiigeed product ownertoo мэдэгдчихлээ.” (I’ve pushed the code and notified the product owner.). When we insisted on linguistic purity, our models failed spectacularly. When we embraced messy reality, they worked.
Release strategy that makes sense
While big tech companies focus on the “top 20 languages” and throw everything else into an “other” bucket, we take a different approach. For each market, we focus on the languages that actually interact. For Mongolia, that means:
Mongolian (obviously)
English (global connectivity)
Russian (historical ties)
Chinese (trade relationships)
This connected-language approach delivers 10x better results than generic multilingual models.
Your data will be securely stored within your country’s territory, allowing you to use it with peace of mind. If governments fail to actively embrace and utilise artificial intelligence, they risk falling behind, potentially leading to significant social inequality within their societies.
Digital sovereignty through pragmatic building
True sovereignty means controlling critical layers while pragmatically using what works.
What we built ourselves:
Character encoding and input methods (100 per cent local)
Finate state automata and transducers as language complexity (100 per cent local)
Speech recognition and synthesis (100 per cent local)
Core language models (100 per cent local)
Application frameworks (80 per cent local, 20 per cent open source)
Infrastructure management (60 per cent local, 40 per cent standard tools)
The honest trade-offs:
We use NVIDIA hardware (no alternative yet)
We leverage open-source frameworks (why reinvent PyTorch?)
We adapt international research (standing on the shoulders of giants)
But we control all critical paths and data
Economic sustainability:
Sovereignty without economic sustainability is just expensive nationalism. Our model evolved through necessity:
Government as first customer, not sugar daddy
Enterprise solutions funding research
Open source contributions building global goodwill
The uncomfortable truths
Let me be honest about what building AI in a developing nation really means:
You’ll never have enough resources. We have 1/1000th of OpenAI’s budget. Make constraints drive innovation. Our GPU shortage forced us to optimise models that now run on minimal hardware
Technical debt is inevitable. Plan refactoring cycles from day one. We still maintain code from 2017 because it serves critical government systems.
International competition will arrive. Build cultural moats early. By the time OpenAI or Google supports Mongolian properly, we’ll have ten years of local context they can’t replicate.
Talent will always be scarce. Create compelling missions. We can’t match Silicon Valley salaries, but we offer something better: the chance to preserve their culture through technology.
Trust is your biggest challenge. When we first claimed 97 per cent accuracy in Mongolian speech recognition, nobody believed us. International researchers said it was impossible. Now, the customers and government say, “Why not just use ChatGPT?” You need rock-solid local success stories. Building credibility took years of consistent delivery.
Also Read:
Asia rises in the AI chip race: China to outgrow US by 30 per cent by 2030
Asia rises in the AI chip race: China to outgrow US by 30 per cent by 2030
The future is distributed
The era where a few companies in a few countries control global AI is ending. The future belongs to distributed, culturally-rooted AI systems serving specific populations with a deep understanding.
Mongolia’s journey from missing keymaps to billion-parameter models proves that developing nations don’t need charity. Don’t blame Silicon Valley for ignoring your market – they don’t owe you anything. Build it yourself.
With sustainable development focused on real problems, pragmatic building strategies, and a vision for digital sovereignty, any nation can achieve AI independence. Don’t be discouraged if someone says you can’t compete with OpenAI or Anthropic. While they are currently reducing their model parameters, we are increasing ours. In the near future, we will converge in the middle. The first to reach that point could be the winner, but nevertheless, business applications using models under 30B parameters will capture 90 per cent of the whole AI market share.
The question isn’t whether to build homegrown AI, but whether your problems are painful enough to sustain the decades-long journey to solve them. For us, watching our youth lose their mother tongue to bad machine translation was painful enough.
But when a nomadic herder can speak to AI in their own language and get help accessing government services? When corruption drops because algorithms can’t be bribed? When does your culture live digitally for future generations? Egune AI is making it possible.
By Badral Sanlig

Published Date:2025-11-13