Doubao Hits 382M Users but Generates Under $1M Daily Revenue, Collapsing Traditional Internet Traffic Logic

According to QuestMobile data from June 2026, ByteDance’s AI chatbot “Doubao” reached 382 million monthly active users (MAU). This represents a 172.1% year-over-year increase, surpassing the combined user base of the second-ranked Qwen and third-ranked DeepSeek, establishing an unparalleled scale among AI chatbots in China.

However, reports from LatePost present a different picture. Doubao’s daily revenue is under 1 million yuan (approximately 20 million yen), with the vast majority dependent on e-commerce transaction fees embedded within the TikTok (Douyin) marketplace. Meanwhile, daily computational resource costs run into tens of millions of yuan. The metrics—382 million MAU, under 1 million yuan daily revenue, and tens of millions in daily costs—point not to a revenue model, but a significant cash drain.

On June 24, Doubao launched three tiers of paid subscriptions: Standard (68 yuan), Enhanced (200 yuan), and Premium (500 yuan). While some view this as the “first step towards commercialization,” as noted by a report from Huxiu, what this move exposes is not the potential for C2C revenue, but the collapse of the foundational premises of traditional internet business in the AI era.

The Premise of Traditional Traffic Logic

For the past two decades, the Chinese internet industry operated on a constant principle: user numbers are an asset, convertible into advertising revenue.

WeChat’s 1.2 billion MAU generates over 100 billion yuan in annual ad revenue. TikTok’s 700 million daily active users (DAU) record single-day livestream commerce GMV exceeding 10 billion yuan. Traffic captures attention, attention converts to ad slots, and ad slots yield revenue. This chain was scripted into every internet company’s business plan.

However, AI products overturn this premise. According to disclosures from Volcano Engine, as of March 2026, Doubao’s large language model exceeded a daily token usage of 120 trillion—a doubling from three months prior and a 1000-fold increase compared to May 2024.

As a Huxiu article details, “each question-answered pair, each image generated, each office task consumes computational costs.” Whereas the marginal cost of internet products approaches zero, the marginal cost of AI products is significantly detached from zero. This signifies a fundamental difference between the two business models.

Daily Costs in Tens of Millions,

Annual Losses Potentially in the Hundreds of Billions

A May 2026 report from Guolian Minsheng Securities estimates ByteDance’s total daily cost (including Doubao’s free C2C service) ranges from 132 to 240 million yuan. Calculating on the upper limit suggests the annual loss scale could approach 100 billion yuan (approximately 2 trillion yen).

A technical analysis report by CSDN breaks down Doubao’s single inference cost: hardware depreciation accounts for 58%, electricity for 29%. Every interaction incurs real cost, and the rate of consumption consistently outpaces the rate of monetization.

ByteDance’s 2026 capital expenditure has been raised from an initial plan of 160 billion yuan to over 200 billion yuan. According to the South China Morning Post, the company is also negotiating overseas financing of approximately $20 billion. In April, it procured around 350,000 Huawei Ascend 950PR chips for $5.6 billion. These funds are being invested not in “purchasing traffic,” but in “acquiring computational resources.”

ByteDance’s internet era monetized user dwell time via recommendation algorithms. A user watching for one minute generated one minute’s worth of revenue. In the AI era, Doubao consumes one unit of computational cost per user query, yet there is no mechanism to insert in-feed ads into its responses.

ARR of $4 Billion, 90% of Revenue from B2B

According to Huxiu reports, ByteDance’s large language model business has achieved an Annual Recurring Revenue (ARR) surpassing $4 billion (approximately 27.2 billion yuan), exceeding the combined ARR of other major domestic model companies. However, dissecting the structure reveals the majority is not C2C.

Zhipu AI’s ARR is around $1 billion, derived entirely from API and coding plans. DeepSeek’s ARR is $400-500 million, with approximately 90% from API revenue. Kimi’s ARR reaches $300 million, with API revenue accounting for over 70%. China’s first wave of large language model companies are fundamentally businesses selling infrastructure to the enterprise (B2B) sector.

As reported by 36Kr, Volcano Engine holds a 49.5% share in China’s public cloud Model-as-a-Service (MaaS) market, with daily API call volumes exceeding 180 trillion. Volcano MaaS’s 2025 sales were approximately 1.5 billion yuan, and the 2026 target has been raised to 10-15 billion yuan.

The monthly revenue of the video generation model Seedance exceeds 1 billion yuan, annualizing to roughly $2 billion, with a gross margin of 70%. In Q2 2026, 90% of Feishu’s new corporate customers also purchased Feishu’s AI features. B2B represents the real revenue, while C2C is merely the fractional remainder.

The Four Strategic Intentions Behind

Launching Paid Subscriptions

Huxiu’s analysis suggests Doubao’s paid subscription launch conceals a four-stage strategic intention.

First is experimentation. Baidu initially attempted paid models in November 2023 but reverted to free in April 2025 due to pressure from DeepSeek. Closing the free entry point in a market where alternative products are available at zero cost means pushing users to competitors. Doubao will retain a free tier, gradually segmenting users.

Second is a signal to B2B customers. Monetizing C2C proves to enterprise clients “that someone is willing to pay for Doubao’s capabilities.” In MaaS sales, “already having achieved monetization” itself becomes a valuable talking point.

Third is exploring price tolerance. It gauges the C2C price tolerance range at minimal cost, testing which features trigger churn at which price tiers. This provides more accurate data than any market survey.

Fourth is offsetting computational costs. Frequent users of the Premium tier partially shift inference costs from “ByteDance alone” to “user-shared.”

None of these four stages includes “profiting from C2C subscriptions” as an objective.

Conversion Rate 1.8%,

Willingness to Pay Depends on Output Value

According to a report from iFeng, interviews with Doubao’s paid users show willingness to pay is contingent on the quantifiable output value. A self-media blogger subscribed to the 68 yuan Standard plan, experiencing efficiency gains in article extraction and analysis. An e-commerce operator stated the 200 yuan Enhanced plan generates “an additional 20,000 yuan in monthly income.” Conversely, a private high school chemistry teacher uses the Standard plan for virtual experiment procedure design but frankly admits, “my income isn’t high, so even the lowest plan feels a bit extravagant.”

Industry survey data indicates domestic C2C AI paid conversion rates typically range from 1% to 3%, with mass-traffic products like Doubao limited to 1%-1.8%. Specialized productivity-focused products like Kimi or DeepSeek see conversion rates of 2%-3% with higher Average Revenue Per User (ARPU). The root of this difference lies in the depth of user engagement.

The Structural Contradiction of AI Products

A Huxiu article points out that AI products possess an inherent structural business contradiction. Traditional internet products monetized user dwell time through ad sales. AI products aim to complete tasks efficiently, causing users to leave once the task is done. Lacking space for ads, they are forced to rely on a low-conversion subscription model.

Resolving this contradiction is a challenge not just for ByteDance, but for the entire AI industry. C2C user acquisition costs are high and inefficient, while free users solely consume computational costs. ByteDance has already integrated Feishu into Doubao, building a closed loop leveraging C2C traffic, Feishu’s enterprise entries, and Volcano Engine’s computational resources, aiming to cover computational costs through B2B enterprise services.

Editorial Opinion

In the short term, Doubao’s paid subscription launch will prompt competitors to follow suit. Domestic major model enterprises like Baidu and Moonshot AI will be compelled to redefine the line between free and paid services. With the numbers showing that standalone C2C monetization is difficult, resources will accelerate towards concentration in B2B (MaaS/API).

In the long term, AI product revenue models must shift from “user dwell time” to “the value of task resolution.” As the ad model fails to function, enterprise AI agents and workflow integration will become primary revenue sources. The expansion of Volcano Engine’s MaaS business, as cited in Huxiu reports, is pioneering this structural transformation.

The question the editorial board poses concerns the possibility of infrastructure innovation that fundamentally alters the cost structure of computation. If inference costs can reduce to 1/100th of current levels, the premises for C2C monetization could change. The massive investment in computational resources, evidenced by the large-scale procurement of Huawei Ascend 950PR chips, can be read as a gamble on cost reduction. Whether this gamble succeeds will likely determine the success or failure of the AI business.

References

• ” 3.82亿月活每天收入不足百万，AI的流量逻辑已经死了 ”, by 听风译码 — 虎嗅网, 2026-08-09T22:15:13.000Z (ARR)
• Source URL: https://www.huxiu.com/article/4881831.html?f=rss