Doubao must be charged: the era of free AI is over

Core Quick Fact

On June 3, 2026, ByteDance’s AI application Doubao issued an official statement announcing the launch of a professional version paid subscription for productivity needs - Standard version 68 yuan/month, enhanced version 200 yuan/month, and professional version 500 yuan/month. This news quickly triggered a heated debate on social media between the "pay to uninstall" and "there is no free lunch" factions. But when a set of figures surfaced, the focus of the controversy shifted from "should it be charged?" to "why Doubao couldn't bear it first": As of March 2026, the average daily Token calls of Doubao's large model exceeded 120 trillion, and the annual computing power expenditure was conservatively estimated to exceed 30 billion. This is not because Byte wants to collect money, but because the computing power bill cannot be sustained.

A boiling computing power bill

There are essential differences between large models and traditional Internet products. For every additional user of WeChat, Douyin, and Taobao, the shared server cost approaches zero—this is the economic foundation for the establishment of the free Internet model. But large models are different. Behind every question, every picture, and every video generated are real GPU operations.

💸 Cost structure dismantling (Doubao single-time reasoning): Hardware depreciation accounts for 58%, and electricity consumption accounts for 29% - nearly 90% of the cost of each conversation is real resource consumption, rather than a one-time investment in infinitely reusable infrastructure.

The more users you have, the faster you will burn money. Doubao’s user growth curve is almost a cost index curve:

During the same period, the number of monthly active users of Doubao reached 345 million, and the average monthly usage per person in the first quarter was 54.8 times. When hundreds of millions of users consume computing power in units of tens of billions of Tokens every day, free is no longer a strategic choice, but a loss of blood that cannot be sustained.

Byte paid a heavy price for this:

• 📉 ByteDance’s capital expenditure in 2025 will be approximately 150 billion, of which 90 billion will be earmarked for AI computing power procurement
• 📉 Increased to 160 billion yuan in 2026, many media pointed out that it may actually be closer to 200 billion yuan
• 📉 Byte's full-year net profit in 2025 will fall by more than 70% year-on-year. The core reason is the substantial increase in AI chip purchases in the second half of the year.
• 📉 Doubao’s net loss in a single month is in the range of 210 million to 270 million
• 📉 The overall loss of domestic C-end large model applications in 2025 will exceed 18 billion, and CICC predicts that it will exceed 20 billion in 2026

The "Beanbag Paradox" of 345 million users

In a market where almost all large-scale domestic models are advertised as free, the first player to put out a price list is bound to suffer the most controversy. From a horizontal comparison:

Why can’t Doubao, which has the largest user base, be the first to hold on? The answer lies in scale itself. It’s not that we don’t want to accept other competing products, but that the user base has not yet reached the critical point where we cannot sustain them without accepting them. Yuanbao’s monthly active users are about 109 million, and Qianwen’s monthly active users are about 203 million, which is an order of magnitude different from Doubao’s 345 million. When the number of users exceeds 300 million and the average daily token consumption exceeds one hundred billion, hundreds of millions of yuan will be burned for every extra day of free use.

📈 "Bean Bao Paradox": The scale of Dou Bao today is the result of the rapid advancement of Byte. From the public beta in the name of Grace in August 2023 to the official launch in May 2024, the monthly living expenses were only 1.73 million at first. But Byte then mobilized the entire group's traffic resources to fully inject it - Douyin opened two entrances, Jianying provided video capabilities, the Volcano engine supported the training framework, and e-commerce and local life scenes were linked. By November 2024, monthly activity has reached 59.98 million, ranking second in the world after ChatGPT. The greater the traffic, the thicker the bill - Doubao has become a typical example of the simultaneous expansion of scale and losses in the domestic AI industry.

Accurate redistribution of computing power bills

A careful look at Doubao's three-tier pricing reveals that its design logic is not simply the realization of Internet traffic, but the precise redistribution of a computing power bill.

🎯 Core logic: A simple text question and answer token consumes only a few hundred times, while the computing power consumption of a PPT generation or high-definition video is hundreds of times of daily conversation. Doubao charging is aimed at these complex productivity scenarios with high computing power and high value - whoever uses it more will pay more**.

The basic free version retains daily chat, Q&A, basic copywriting, translation and simple drawings, and promises to be permanently free without emasculating the core experience.

Looking further, the commercialization of beanbao does not expect to rely on subscription members to support all profits. According to 36Kr reports:

1. Q3: Combine e-commerce functions to improve payment scenarios, and use subsidies to attract traffic to Douyin Mall
2. Q4: Entering the running period
3. 2027: Longer-term commercial returns

Even if the 345 million users only have a payment conversion rate of 1%, based on a minimum calculation of 68 yuan/month, it can still bring in annual subscription revenue of approximately 2.3 billion yuan. Some analysts estimate that Doubao’s annualized subscription revenue will be in the range of US$400 million to US$700 million under a neutral scenario. Compared with Byte's annual AI investment of hundreds of billions, this is only a small part, but it illustrates one thing: C-side AI users are willing to pay for value, as long as the value is clear enough.

The turning point comes

Doubao charging marks that domestic C-side AI has officially entered the "commercial verification" stage from "subsidy to grab users". In the Internet era, the formula of burning money in exchange for scale and scale in exchange for advertising has failed in the large model - every additional user means an additional rigid expenditure on computing power, and the larger the scale, the deeper the losses.

Alibaba Qianwen and Tencent Yuanbao have chosen to stand still at this moment. Although they have considerations of grabbing users for free, CICC predicts that industry losses will exceed 20 billion yuan in 2026. No one can advance money for users forever. For model manufacturers, including DeepSeek, whose selling point is extreme cost-effectiveness, Doubao's charging action provides an important reference coordinate on the commercialization path-all C-side AI players will face the same computing power bill sooner or later.

💡 Industry Enlightenment: Doubao was the first to come up with charging fees and received the most scolding. It may also be the first to find a sustainable commercialization path.

Follow-up directions worth tracking:

1. 📊 Payment Conversion Rate - The user retention rate and payment conversion rate after the launch of Doubao Pay will test the ceiling of C-end AI’s willingness to pay
2. ⏰ Time to follow up on competing products - When will Tencent Yuanbao and Alibaba Qianwen follow up on charging and pricing strategies? The effects of active charging and passive charging are completely different.
3. 🛒 E-commerce commercialization - The specific plans and effects of Q3-Q4 e-commerce payment scenarios will be the more important revenue increase after subscriptions
4. 📉 Computing power cost curve - When the computing power cost of the large model industry begins to decline, which determines whether the free model can return