LLM MaaS: API Capability Packaging and Usage-Based Billing

📌 Background

In 2026, the commercialization of domestic large language models entered deep waters, with leading vendors shifting from selling computing power to selling model services. Zhipu's first financial report after going public showed that its MaaS ARR reached 1.7 billion RMB, making it China's largest revenue-generating LLM company; Alibaba Cloud also announced revenue sharing for Qianwen LLM users. The MaaS model, billed via API calls, has become the mainstream monetization path for LLM companies.

👤 Target Customers

Developers, enterprise clients, and channel partners who resell capabilities via APIs. The payers are enterprises and individual developers who need to invoke model capabilities to build applications or businesses.

💰 Revenue Streams

1) Usage-based billing per token, offering subscription packages at different tiers; 2) Enterprise customized private deployment and solution fees; 3) Platform revenue sharing or channel revenue splits.

🧮 Cost Structure

GPU computing power and training/inference costs. R&D personnel compensation and model iteration investments. Sales, ecosystem building, and API gateway O&M costs.

🛡️ Moat

Leadership in model capabilities and iteration speed. API ecosystem and developer stickiness. Cost advantages brought by token pricing and scale.

🔑 Keys to Success

• Lower API invocation costs to attract developers
• Establish a complete toolchain from models to applications
• Expand channel partners for resale revenue sharing

⚠️ Risks

• Price wars leading to gross margin compression
• Policy regulation and data compliance risks
• Migration of key clients to open-source or self-developed platforms

🏢 Cases

• Zhipu AI's MaaS platform (ARR reached 1.7 billion RMB)
• Alibaba Cloud Bailian platform's Qwen API service
• Moonshot AI's Kimi commercialization exploration

📊 SWOT Analysis