Doubao is about to start charging.
On its App Store page, three subscription tiers have appeared: Standard Edition at 68 yuan (RMB) per month, Enhanced Edition at 200 yuan per month, and Professional Edition at 500 yuan per month. The basic version remains free, while paid services focus on complex, high-compute tasks such as PPT generation, data analysis, and video production.
Doubao's official response states that these plans are still in the testing phase. But the pricing has already been posted, and the structure is very clear.
This is not just another attempt at commercialization. It marks the first time that ByteDance, a "precision-engineered attention extraction machine," has systematically begun dismantling its own long-cherished free strategy.
In March of this year, Volcano Engine disclosed that Doubao had exceeded 120 trillion daily average token usage, doubling in three months and growing 1,000-fold since its launch in May 2024. With over 345 million monthly active users, Doubao becomes both the top AI-native application in China and the largest single cost center in internet history.
On the enterprise side, growth is equally strong. The number of clients with cumulative token usage exceeding one trillion grew from 100 at the end of last year to 140.
By any internet metric, this is a blockbuster hit. But one number remains absent: consumer-side revenue.
Throughout this growth phase, no direct relationship has been established between user scale and revenue. In the context of China's internet over the past fifteen years, there is no precedent. When WeChat's MAU reached 1.4 billion, Tencent's social ad value rose in tandem. When Douyin's MAU exceeded one billion, ByteDance's ad inventory and eCPM grew together.
The iron law that more users mean more profit has governed the industry for fifteen years. Doubao is the first large-scale product to break that rule.
The reason is straightforward. Every conversation consumes real computing resources: GPU inference, model operations, and data calls. These costs do not decline with scale. On the contrary, when 345 million people use the service simultaneously, the bill is more than ten times larger than when 30 million people did.
The user's role has fundamentally changed. They are no longer just a number in the traffic pool, but a part of the cost function, completely overturning the internet's marginal revenue curve. Every additional active user does not create another ad monetization opportunity; it adds another line to the compute bill.
Many people's first reaction is to cover revenue with advertising, as with short videos.
Unfortunately, this path does not work for AI products. The reason is not at the user experience level — though inserting ads into AI responses is indeed a poor experience — but lies deeper in the economic structure.
The advertising model of the past internet relied on a premise: user attention is "segmentable." Advertising is essentially inserting commercial messages into the gaps of the consumption chain. As long as gaps exist, ads can be inserted.
But AI changes that structure. Users do not initiate browsing behavior; they submit task requests. Writing a report, analyzing a dataset, and making a PPT; this is a complete execution process with no interruptible gaps in between. No one can insert an ad between the third and fourth slides of a generated PPT, because the user has requested a complete task. Any interruption breaks the task and directly fractures trust.
Yet the more critical conflict is economic. The advertising model presupposes near-zero marginal cost (serving one more user barely increases expense), so one can offer the service for free and make money from ads across a massive user base. But for AI products, they have a persistent marginal cost that grows linearly with usage. The two are mathematically contradictory.
According to media reports, ByteDance's 2025 revenue exceeded US$200 billion, making it still one of the world's strongest attention monetization engines. But that engine cannot feed Doubao. Advertising revenue grows with "impressions," while Doubao's costs grow with "inferences" — and the latter grows much faster.
Not long ago, Chinese tech website 36Kr reported that ByteDance's net profit in 2025 fell more than 70% year-over-year. Li Liang, Vice President of Douyin, responded on Weibo that after excluding changes in preferred shares and option costs, the decline was "far less severe," but he also acknowledged that operating margins did drop in the second half of the year.
The reason is clear: in Q3 and Q4, the company significantly increased spending on AI compute procurement, model R&D, and infrastructure.
The interesting part is that ByteDance's ad business has not deteriorated. Douyin and TikTok continue to grow.
The problem is that within the same company, two opposite economic logics are operating. On the ad business side, a larger scale brings a higher profit. Every additional user watching videos creates another ad slot to sell. On the AI business side, a larger scale brings a higher cost. Every additional user asking a question requires another inference to run.
Both sides are growing simultaneously, but one side's growth generates profit while the other's generates expense.
This situation is extremely rare in technology companies. Microsoft's Azure and Office at least share the same software distribution logic. Apple's hardware and services both rely on the same ecosystem. ByteDance's advertising and AI, however, are at odds at the most fundamental level.
That is why ByteDance feels the pressure earlier than anyone else. It is not that it lacks money; it is that the way it makes money and the way it spends money are fighting each other internally. The more AI users, the faster they eat into advertising profits. Charging for Doubao is essentially seeking an independent path for the AI business that does not rely on advertising.
This problem is not unique to ByteDance. OpenAI, Anthropic, and Google are all facing the same thing: user scale is growing much faster than revenue models can be established.
But their responses are diverging.
OpenAI has chosen a hybrid free-plus-paid model. Fewer than 6% of its 960 million MAUs are paying, with annualized revenue exceeding US$25 billion, but full-year cash burn for 2026 is projected at US$17 billion. It is raising prices (GPT-5.5 API pricing tripled) while also starting to sell ads (first batch tested on ChatGPT's free tier in February 2026 at a US$60 CPM).
Two years ago, OpenAI CEO Sam Altman said ads would destroy trust. That stance has now been rewritten by reality.
On the contrary, Anthropic chose to charge from day one. It has virtually no free users, generating US$211 in revenue per MAU — eight times OpenAI's monetization efficiency. Its user base is smaller, but its financials are healthier.
Google took a third path, embedding AI into cloud services and covering costs through enterprise customer payments. Google Cloud's Q1 revenue reached US$20 billion, up 63% year-over-year, with operating margins rising from 17.8% to 32.9%. It does not charge consumers directly for AI; instead, it lets enterprises pay for AI through their cloud bills.
Three paths, answering the same question: Who pays for compute?
OpenAI has consumers and advertisers share the burden. Anthropic makes paying users bear the full cost. Google shifts the cost to enterprises. ByteDance is now moving closer to OpenAI's hybrid model.
However, charging consumers has historically been difficult in China. A decade of mobile internet has trained users into habits: socializing is free, content is free, tools are free, and even food delivery and ride-hailing come with subsidies.
Long-form video platforms have struggled for ten years, reaching about 30% paid membership penetration. Music streaming has been around for seven or eight years, with a pay rate of about 20% — far below the 60% seen in mature Western markets.
Doubao's Standard Edition at 68 RMB/month is priced directly against ChatGPT Plus and Claude Pro (US$20), at less than half their cost. ByteDance is doing what it does best: acquiring customers with low prices and building a payment mindset first.
But this time, there is a structural variable that has never existed before: AI completes work for users, not entertainment.
Historically, all consumer-side payments were essentially consumption upgrades. A video subscription lets you see a show a week early. A music subscription gives you higher audio quality. These add a little extra to an existing experience.
What AI sells is different. For 68 yuan, users get a PPT that would have taken two hours delivered in ten minutes, or an analysis that would have required digging through data for half a day run in a single sentence. This is not an experience upgrade; it is labor substitution. Exchanging money for time, money for output.
The psychology behind these two types of payment is completely different. Willingness to pay for a better experience has a low ceiling. Willingness to pay for labor saved has a ceiling that depends on how much your time is worth.
Returning to the three pricing tiers (68 RMB, 200 RMB, 500 RMB), on the surface, they represent plan segmentation, but the underlying logic is task stratification: light assistance at one price, complex tasks at another, professional production at yet another. Yet beneath task stratification lies something deeper: for the first time, a software product is explicitly pricing time.
In the past, paying for efficiency was for enterprises. Companies bought SaaS and ERP to pay for employee productivity. Individual users rarely pay directly for "I can get my work done faster."
AI is breaking down that wall. When an ordinary person can use Doubao to produce something that previously would have required hiring someone or working overtime, they become a "micro-enterprise user." They pay for their output, just as a company pays for its employees' output.
After news of Doubao's planned charges emerged, will free users disappear? Clearly not.
Doubao retains a free tier, ChatGPT has a free tier, and Google Gemini does too. What has changed is the role of free users in the business model.
In the past, free users were a core asset. Platforms wanted everyone to come; more users, more data, more valuable ads. Free users themselves were the foundation of monetization.
In the AI era, free users have become a resource that must be managed. All the free-tier restrictions: limiting call frequency, limiting model parameters, and setting up queues, are essentially a system of precise cost rationing.
The future revenue structure for AI products will likely resemble that of cloud computing. The free tier brings people in. Beyond a certain usage level, billing begins. Basic resources are cheap or free; high-performance resources are pay-as-you-go. Free has shifted from the destination to the entry point.
This is a belief that has loosened for the first time in twenty years. "More free users are always better" holds only for businesses with zero marginal cost. AI has exposed that precondition.
All business model analysis ultimately boils down to a very simple condition: Can AI get the job done well?
If it can, then payment is only a matter of time. Price is not an issue either, because users are buying a finished product, not the right to use a tool.
If it cannot, then the premise of saving users time collapses. Users who pay still have to do rework themselves, the paid experience rapidly deteriorates, and the free tier becomes mainstream again.
This might be the biggest uncertainty in the entire situation. Pricing strategies can be adjusted, customer acquisition methods can be changed, but the product question of whether AI can reliably deliver stable results cannot be bypassed by any business model.
ByteDance's algorithmic capabilities are world-class, and its model iteration speed is fast. But the gap between "can do" and "can do reliably" may be wider than most people imagine.
Ultimately, what determines whether Doubao's paid model succeeds is not the pricing strategy, but the completion rate.
(Source: 36kr)
Related News:
Deepline | AI artists: A threat to fan economy?
Deepline | 'AGI brotherhood' of Zhichun Road: DeepSeek and Kimi's shared pursuit
Comment