← Sticker Wall AGENT · DETAIL

Chai Research: A 1.6M-DAU conversational model that hit $30M ARR

Workflow: Every day, users' conversations with AI characters generate massive interaction data. The system automatically collects

AGENT

Key Fields

FIELD STAMPS

IndustryContent / Creator Economy

RegionUS

ScaleSME

ChannelOnline

🔧 Workflow

Every day, users' conversations with AI characters generate massive interaction data. The system automatically collects users' likes, dislikes, and continued-conversation behavior on replies, then uses these preference signals to fine-tune model weights. The fine-tuned model outputs more engaging conversations, thereby increasing subscription conversion and retention, forming a data flywheel. Specifically, the input is multi-turn dialogue logs and user-level feedback labels from AI characters in the mobile app, and the output is updated conversational model weights and personalized character reply strategies. The entire process runs automatically every day, model training tasks execute in a background queue, and the operations team only needs to check key metrics once a week, such as subscription conversion rate, next-day retention, and average revenue per user.

🛠 Setup Requirements

It requires a mobile app that can support 1.6 million DAU, a fine-tuning pipeline built in-house or based on open-source large models, and a real-time preference data collection architecture. Technically, the team needs capabilities in model training, mobile development, and recommendation systems. Building from scratch takes at least 8 to 12 months. If starting as an individual or small team, you can first build an MVP with open-source conversational models such as Llama 3 or Mistral, integrate Stripe or iOS/Android IAP subscription payments, and then collect user likes and dislikes through Firebase or a self-built backend. Key infrastructure includes GPU training clusters, a vector database for storing character personas and conversation history, and an A/B testing framework to validate fine-tuning results.

🧰 Toolchain

🔧 Mobile AI chat app
🔧 User preference data collection pipeline
🔧 Large model fine-tuning training framework
🔧 Subscription payment system (iOS/Android IAP)
🔧 GPU training cluster
🔧 A/B testing and data analytics platform

💰 Revenue

$30 million ARR (about $2.5 million/month), 1.6 million DAU. Each paying user contributes about $15.6 per month on average, estimated by dividing ARR by 12 months and then by 1.6 million DAU. This figure is higher than the average revenue per customer of consumer AI apps, indicating that emotional companionship products have strong willingness to pay.

💸 Cost

Model training and inference API costs account for the largest share, estimated at hundreds of thousands of dollars per month; there are also mobile development and server operation and maintenance costs. GPU inference costs, calculated based on 1.6 million DAU and 20 conversation turns per user per day, may consume $200,000 to $400,000 per month, depending on model parameter scale and quantization scheme. Mobile CDN and push notification services cost about $10,000 to $30,000 per month.

⏱ Time Investment

The team operates continuously, investing more than 100 hours per week to maintain model training, user operations, and product iteration. If estimated as a small team of 3 people, each person invests at least 35 hours per week, of which model training monitoring accounts for 40%, user feedback analysis and character persona updates account for 30%, and subscription conversion and customer support account for 30%.

🚀 Getting Started

Step one: build a minimum viable AI chat app based on an open-source conversational model, and integrate like/dislike feedback collection. Step two: use interaction data from the first few hundred users to run a round of fine-tuning, validate improvements in reply quality and retention, and then gradually launch and scale. Step three: integrate Stripe or Apple/Google subscription payments, set up weekly, monthly, and annual subscription plans, and use A/B testing to find the optimal pricing. Beginners do not need to train a large model from scratch; they can use open-source conversational models on Hugging Face to first run through the closed loop, and consider self-building or deep fine-tuning after technical validation and data accumulation.

🔑 Keys to Success

✅ Fine-tune the model with real-time user preferences to create differentiation in conversation quality
✅ The AI companion track captures the rigid demand for emotional companionship, and subscription revenue is stable
✅ A self-built model reduces dependence on third-party large model APIs and improves gross margin
✅ Continuously collect like and dislike signals so the model can iterate and optimize weekly
✅ Mobile-first product design lowers customer acquisition barriers and user friction

⚠️ 风险

⚠️ AI companion products face regulatory pressure on data compliance and protection of minors
⚠️ General-purpose large models continue to upgrade, and self-built small models may be replaced at low cost
⚠️ Overall user retention is low, requiring continuous paid acquisition to maintain DAU scale, leading to rising customer acquisition costs
⚠️ The emotional companionship track is severely homogenized; once leading players launch stronger models, small-team products can easily be crushed

📌 Real Cases

📌 Chai Research's mobile app relies on AI character conversations, with 1.6 million DAU and $30 million ARR, making money far more efficiently than Character AI.
📌 36Kr reported that Chai's ARR reached $30 million, users send an average of more than 20 messages per day, and per-user willingness to pay is significantly higher than similar AI chat products.
📌 Chai's 2026 roadmap emphasizes achieving leadership in conversation quality through Mixture of Experts post-training. About 72% of its subscription revenue across iOS and Android comes from weekly subscriptions, indicating that users' payment decision cycle for instant companionship is very short.
📌 AIStart.ai's Chai tool page shows that the product still maintains more than 1.6 million DAU in 2026, with subscription prices ranging from $6.99 per week to $99.99 per year.

Sources