China's AI App 'Doubao' Loses 6 Million Users After Announcing Paid Version

(Central News Agency, Taipei, June 4) Two years after Chinese internet giant ByteDance launched 'Doubao', it announced a paid version in early May. However, this AI application, which holds the top spot for user numbers in China, lost over 6 million customers before even collecting any fees. This underscores the dilemma facing Chinese AI companies: under pressure from soaring costs, they also face consumers unwilling to pay for AI services.

According to reports from Yicai and the South China Morning Post, Doubao officially announced yesterday that it plans to launch a professional version targeting the productivity needs of professionals. This version will include professional services such as software development, data analysis, professional design, process automation, financial analysis, and scientific research.

Doubao emphasized that functions used by the general public daily, including search and Q&A, writing and image generation, and voice and video conversations, will remain free to ensure user experience and habits are unaffected.

This marks the first time a product with the highest monthly active users (MAU) among Chinese AI agents has introduced a paid subscription. In early May, Doubao launched three paid versions on the Apple App Store, with the most expensive professional version costing 5,088 RMB (approximately NT$23,670) per year, expected to start in late June. Following this news, Doubao experienced its first user decline.

According to data released by the AI Product Rankings (AICPB) on June 3, Doubao's MAU in May was 330 million, a decrease of 1.81% from April, losing approximately 6.07 million users.

The South China Morning Post quoted Li Bangzhu, founder of the AI Product Rankings, as saying that in the fiercely competitive Chinese market, user numbers determine everything. The 1.81% decline is a worrying signal for ByteDance.

Li said, "The era of free AI services in China is far from over, so Doubao's commercialization may be premature."

Meanwhile, Doubao's strongest competitor is rapidly rising. Alibaba's 'Qwen' saw its MAU surge by over 13 million in May, reaching 234 million.

Li stated that the trend indicates Qwen is steadily eroding ByteDance's advantage. "When user switching costs are still very low, whoever starts charging users first is essentially handing the initiative over to competitors."

The report said the decline in Doubao's user numbers highlights a broader challenge facing Chinese tech giants. In the world's largest internet market, consumers so far seem reluctant to pay for daily AI services.

The report pointed out that the fundamental pressure driving Doubao towards paid services comes from the high computing costs of large models. Every AI conversation and task generation consumes real computing power; the larger the user scale, the more the computing bill swells. As of March this year, the daily token usage of Doubao's large model had exceeded 120 trillion, a 1,000-fold increase from the 120 billion at its launch in May 2024.

Converted into actual costs, this is no small amount. According to data from Chinese AI application analysis firm Xsignal, Doubao's annual token inference cost is approximately 8 billion RMB, and its annual chip computing cost is approximately tens of billions of RMB.

According to reports, ByteDance's profit in 2025 fell by more than 70% compared to 2024. Insiders revealed that the core reason was a significant increase in AI investment in the second half of 2025, with high spending on computing procurement, infrastructure, and R&D dragging down the annual profit performance.

The report said that the scale advantage built through subsidies needs to be converted into real revenue, but the traditional 'free + advertising' model is largely ineffective for large models. AI handles complete task flows, making it difficult to insert ads into conversations without disrupting the user experience. This leaves Chinese AI agents generally facing the dilemma of being 'well-received but not profitable'. (Editor: Yang Shengru / Chen Yanjun) 1150604