FOUNDER FOCUSWe speak with early-stage AI founders and executives about AI, startups, and whatever else is on their minds. In this edition, Tech in Asia sits down with Hiep Nguyen, chief operating officer of Vietnamese AI startup AI Hay, which is building Vietnam's answer to ChatGPT. He believes success in Southeast Asia's AI race will come not from building the best model, but from borrowing the playbook of game developers. The interview has been edited for brevity and clarity.  When Tech in Asia last spoke with you, AI Hay had reached 15 million downloads, 120 million monthly queries, and 1 million daily users. How have those numbers changed since then? Right now, we have over 21 million downloads across the App Store and Google Play Store, around 280 million monthly AI interactions, and more than 2 million daily active users. A lot of AI Hay users also have other AI apps on their devices such as ChatGPT and Gemini. The idea is to answer Vietnamese and Vietnam-related questions well enough that AI Hay becomes the default choice. We want users to come to us whenever they have a question about Vietnam or want to ask something in Vietnamese. AI Hay is still pre-revenue. What's your strategy for turning the product into a business, and when do you expect monetization to begin? We aren't selling subscriptions for AI Hay; instead, we're selling stars. Users can collect stars by doing different missions, or they can pay for them. The stars can be used for different tasks: a normal question is 10 stars, deep research is 100, creating a PowerPoint is 150, images are 200, and so on. As gaming culture is ingrained in Vietnam, if you make a product like a game - with microtransactions and credits - users are more likely to adopt it. When you build a consumer AI product for Southeast Asia, you are building for a market with low average revenue per user but high engagement. These people will not pay as much as those in the West, but they will use the product as intensely as anyone in those markets. If you don't own your post-training inference optimization, you get crushed. But because we charge by these stars, we can price the stars proportionately to each task. If a user wants to create an image, that uses an image model, which costs a lot more than a text model. So we can price the stars accordingly, instead of offering a subscription, so people can't go ham. We started monetization in March, and only about 0.2% of our user base is currently seeing any paywall. Rolling it out gradually is a product decision, as at our scale, flipping it to 100% at once would cause problems. The goal is a freemium model where basic needs are always covered but premium features cost a little extra. And a little really does mean a little - these features are cheaper than a coffee. The model isn't about extracting 2 million Vietnamese dong (US$76) per user. It's about 20 cents from 100 million users, consistently, every day. How is AI Hay's expansion into the B2B market progressing? It's going well, but we've realized that we have a consumer-first DNA. If you look at companies like OpenAI and Anthropic, they serve consumers, governments, and enterprises. But the foundation is still a consumer-grade product that people want to use. Our product is easy for people to use. We've started seeing a lot of bottom-up adoption from office workers and public-sector employees. They're using AI Hay on their own, which has led to new opportunities. One of the biggest opportunities we're piloting right now is with city officials. So while we're still focused on the consumer side, we see a clear path to expand into other segments. As a Vietnam-focused model provider, what do you see as your edge as larger models improve at Vietnamese? Language was never really our edge. The larger models will always get better at Vietnamese as they scale. Our advantage lies in two areas: understanding user intent and serving niche, high-volume use cases. Across nearly 300 million interactions, 76% of the prompts AI Hay receives contain fewer than five words. Many users live in rural areas and access the app on inexpensive smartphones, making long text prompts impractical. Even when users don't type, they rarely communicate in standard text. About 80% of longer prompts include a photo or document in Vietnamese - handwritten notes, chalkboards, or scanned PDFs. Other users send voice notes in strong regional accents. So the real challenge is figuring out what someone actually needs from very little input. On top of that, there are highly localized use cases: lottery numbers, university entrance exam policies that changed last month, Vietnamese stock tickers. We have to stay on top of all of them. If you do that well enough, you build a third advantage: cultural mindshare. It's like WhatsApp versus Zalo - both work, but you instinctively know which one you reach for in different situations. Are you planning to raise more funding? We're probably going to start fundraising soon, likely sometime this year. |
