The domestic native AI application market has reached a significant turning point indicative of industry trends. On May 4th, ByteDance's AI application Doubao updated its service description for a paid version on the Apple App Store, officially announcing the launch of a three-tiered paid subscription model while retaining its free basic version. The tiers are: Standard (continuously billed at 68 RMB monthly or 688 RMB annually), Enhanced (200 RMB / 2048 RMB), and Professional (500 RMB / 5088 RMB). This move not only signifies Doubao, as the AI application with the largest domestic user base, taking a substantial step towards C-end commercialization but also signals the end of a phase where domestic large language model applications relied solely on free services to build scale.
According to the latest QuestMobile Spring 2026 report, the monthly active user (MAU) scale for domestic native AI apps reached 440 million by Q1 2026. Among these, Doubao holds an absolute leading market share with 345 million MAUs, a figure exceeding the combined total of the second-ranked Qianwen (166 million) and third-ranked DeepSeek (127 million). However, this massive user base brings not pure scale effects but exponentially increasing computational power consumption and sustained pressure on profitability.
The core logic behind Doubao's trial of paid services is to alleviate the cost pressure from high-computation scenarios by charging core productivity users, thereby establishing a sustainable business model. As a leading application among major players that is the first to break the "completely free" tacit agreement, Doubao's paid attempt carries significant industry implications. Morgan Stanley, in its latest research report, characterized Doubao's initiation of paid services as a "critical juncture for the industry transitioning from a user education period to a commercialization phase."
Inevitably, some users may migrate to other free platforms in response to Doubao's paid mechanism. However, whether this means competitors like Qianwen and Yuanbao will seize the opportunity to capture this traffic and maintain a long-term free stance remains a question. Furthermore, whether this user migration represents a boon or a liability for the receiving platforms, given the computational costs involved, will only become clear after Doubao's paid model is officially launched.
Reflecting on Doubao's growth trajectory reveals an enviable path for any internet product. Throughout 2025, its MAUs surged from 99.8 million in Q1 to 227 million in Q4, an increase of over 127%, making it the first domestic native AI app to surpass 100 million daily active users (DAU). Entering 2026, growth continued unabated, adding approximately 100 million active users in Q1 alone, with DAU peaking at 145 million during the CCTV Spring Festival Gala period.
More notably is the quality of growth. Data from domestic mobile ad monitoring firm AppGrowing indicates that Doubao's marketing expenditure overall trended downward in 2025: 161 million RMB in Q1, 117 million RMB in Q2, halving to 65 million RMB in Q3, and slightly recovering to 92 million RMB in Q4. Concurrently, Doubao's average 30-day user retention rate from January to November 2025 reached 44%, significantly leading the second-ranked player in the industry. This "low acquisition cost, high retention" growth model propelled Doubao far ahead of its peers to become the top domestic AI application.
However, the underlying economics of large model applications differ fundamentally from traditional mobile internet products. For traditional SaaS or social software, marginal costs approach zero, meaning the cost of serving an additional user is minimal. But for generative AI, every interaction and every inference tangibly consumes the computational power of GPU clusters. According to a computational cost breakdown from CNDS, hardware depreciation accounts for a substantial 58% of Doubao's single inference cost structure, with power consumption comprising about 29%. This implies that the 120 trillion daily tokens generated by 345 million MAUs directly translate into daily bills amounting to millions in power and hardware depreciation.
Under profitability scrutiny, a model where "high DAU equals high liabilities" is clearly unsustainable in the long term. As model capabilities upgrade, particularly with the introduction of multimodal and long-context parsing abilities, the inference cost per task is rising exponentially. The computational power required to generate a professional PPT or render a few minutes of video is often tens or even hundreds of times greater than that for casual conversation. Continuing to provide completely free services to all users without distinction risks having the platform's financial model overwhelmed by a minority of heavy-usage users.
Many Doubao users have reported to Wall Street News that during recent evening peak hours, the app frequently experiences network queues for complex tasks like in-depth research and video generation, indicating that the computational resources for its free services are under strain. Under these various pressures, the "free" model has become unsustainable. Sources close to ByteDance previously indicated that the company's net profit in 2025 fell by over 70% year-on-year, with substantial AI capital expenditure being a major drag.
Doubao's future model will adopt a tiered structure with a permanently free basic version and paid premium services. Sources close to Doubao revealed that paid features will primarily focus on complex tasks and productivity scenarios, such as PPT generation, data analysis, and video production. In other words, Doubao aims to use its three-tier pricing to precisely identify the heavy computational power consumers within its 345 million MAU base.
From a pricing strategy perspective, Doubao's monthly fees of 68 RMB, 200 RMB, and 500 RMB position its entry price nearly 40% higher than domestic competitor Kimi's 49 RMB/month. Morgan Stanley notes that this pricing level targets professional users rather than the mass market, with the intended audience being creators and knowledge workers. This suggests Doubao's core monetization logic is not primarily about converting free users to paid users, but rather about filtering for high-value users with strong payment willingness and ability.
Morgan Stanley's analysis of Doubao's paid service estimates that even under a conservative conversion rate of 0.3% to 3%, its annualized subscription revenue could range from 101 million to 1.5 billion USD; a neutral scenario suggests approximately 426 million to 684 million USD. While this scale remains limited compared to ByteDance's core advertising business, it is sufficient to open the first window for user monetization.
In reality, Doubao's commercialization is not sudden. ByteDance's AI development platform "Coze" began charging professional developers starting in 2024 and launched a personal plan in late January 2026, with prices ranging from 19.9 RMB to 99 RMB per month. Doubao's move to charge fees can be seen as a natural extension of the systematic monetization of ByteDance's AI ecosystem.
Following the announcement of Doubao's paid plans, market focus swiftly shifted to competitor reactions. With Doubao clearly establishing a paywall, many users began worrying about potential limitations on the free version, inevitably sparking industry discussions about user churn. For instance, some tech-savvy users or heavy productivity users accustomed to free access to advanced features might choose to migrate to currently fully-free alternatives like Qianwen or Yuanbao when faced with monthly fees ranging from 68 to 500 RMB.
Short-term data fluctuations will likely confirm this traffic spillover effect. Many users have already started searching for alternatives. A user involved in professional skills training told Wall Street News that she frequently uses Doubao to generate teaching PPTs, saving significant time, but if it requires a minimum payment of 68 RMB per month, she will begin looking for alternative apps. However, being accustomed to Doubao, she decided to generate a batch of materials before the paid version officially launches.
However, based on the cost dynamics of the large model industry, user migration might also become a "sweet burden" for the receiving platforms, as high-complexity tasks correspond to substantial inference costs. If Qianwen or Yuanbao choose to continue offering computationally intensive services like long-context parsing, deep data mining, and video generation for free indiscriminately to absorb users migrating from Doubao, their server and electricity expenses could skyrocket. In the current macro-environment where internet giants generally emphasize cost reduction, efficiency improvement, and pursuing self-sustaining business units, no company can indefinitely bear the computational burden shifted by competitors without limits.
Doubao's pioneering move to break the "free tacit agreement" might trigger a wave of paid services among major internet companies' AI applications. An AI industry professional told Wall Street News that charging for complex usage scenarios is inevitable in the future to sustainably provide higher-quality services. As a seasoned user who pays to test various domestic and international AI models for work, he believes users aren't entirely opposed to paying; if a model is sufficiently intelligent and genuinely boosts efficiency, many productivity users are willing to pay.
From an industry evolution standpoint, Doubao's率先收费 might mark the beginning of the domestic large model industry's exploration of paid services. As a national-level application with nearly 400 million MAUs, Doubao seems positioned to break the previous industry dilemma of "whoever charges first dies first." In the short term, it remains uncertain whether Qianwen, Yuanbao, and others will follow suit. However, in the medium to long term, the costs associated with serving these high-computation users will inevitably manifest in financial reports, making it highly likely these players will follow Doubao's lead, introducing "basic free + advanced paid" models based on complex scenarios. After all, free services are ultimately a temporary market education tool; value equivalence is the ultimate form of commerce.
In this context, Doubao's commercialization path can be cross-referenced with industry pioneer ChatGPT's paid trajectory. OpenAI also experienced significant cost challenges early on, gradually establishing a strict tiered subscription system. Currently, ChatGPT offers not only a basic free version but also an 8 USD/month Go version, a 20 USD/month Plus version, a Team version, and a high-end Pro version costing 200 USD/month. Comparatively, Doubao's pricing strategy is logically similar to ChatGPT's but has been localized for the domestic market in terms of price points. Doubao's Standard tier at 68 RMB/month is priced lower than ChatGPT Plus, aiming to lower the payment threshold for domestic users and cultivate subscription habits. Its Professional tier at 500 RMB/month, while high in absolute terms, appears relatively restrained compared to ChatGPT Pro's 200 USD.
ChatGPT's financial data has already demonstrated that C-end users possess a strong willingness to pay if the model's inference capabilities tangibly enhance productivity. Doubao's current trial essentially tests whether this logic holds true in the Chinese market. For the industry, this also signifies an elevation in the competitive dimension. Over the past two years, domestic large models engaged in fierce price wars in the B-end API market, severely compressing commercial value. Now, in the C-end market, if Doubao can successfully validate a high-average-revenue-per-user (ARPU) subscription model, it will guide the industry away from pure scale competition towards a comprehensive contest of capabilities and business models.
Returning to its paid model, in the short term, it might trigger some user migration to other free platforms and even spark debates about pricing合理性. However, in the long run, it promotes a return to rational pricing mechanisms in the AI industry. The ultimate test will revert to product capability. When a product is sufficiently powerful to substantively介入 productivity processes, users may be willing to pay for the computational resources and R&D成果 consumed. This could very well become the foundational consensus for the healthy development of the AI industry in its next phase.