On June 3, 2026, ByteDance's AI application Doubao released a carefully worded official announcement. The announcement outlined plans to introduce a professional version, mainly targeting users with productivity requirements, encompassing services such as software development, data analysis, professional design, process automation, financial analysis, and scientific research. Simultaneously, daily functionalities like search and Q&A, writing and image generation, as well as voice and video conversations, will continue to be offered free of charge.

Previous media reports indicated that Doubao's paid content is anticipated to officially roll out in late June, coinciding with the Volcano Engine Force Conference. Earlier, on May 4, the App Store page had already been updated to display three subscription tiers. The standard version is priced at 68 yuan per month or 688 yuan annually. The enhanced version costs 200 yuan per month or 2,048 yuan annually. The professional version is set at 500 yuan per month or 5,088 yuan annually. With the highest annual fee surpassing 5,000 yuan, netizens frequently draw comparisons to ChatGPT Pro's monthly fee of $200.

On social media platforms, two opposing camps—those vowing to "uninstall if it charges" and those asserting "there's no such thing as a free lunch"—are engaged in a heated debate. However, a set of figures warrants more attention than user sentiment: as of March 2026, Doubao's large model witnessed daily Token usage soar past 120 trillion, marking a thousandfold increase since its initial launch in May 2024. Based on an estimated input cost of 2 to 4 yuan per million Tokens, ByteDance is incurring daily expenses of 200 million to 500 million yuan on GPU data centers, with annual computing costs conservatively estimated at over 30 billion yuan.

This isn't merely about ByteDance's desire to charge—it's about the computing costs becoming unsustainable.

01

A Steep Computing Bill

To comprehend why Doubao is implementing charges now, one must first understand the cost structure of large models.

Large models fundamentally differ from traditional internet products. For services like WeChat, Douyin, or Taobao, each additional user drives server costs towards negligible levels—the economic foundation of the internet's free model. This isn't the case for large models. Every user query, image generation, or video creation necessitates real GPU computation. A cost breakdown reveals that hardware depreciation accounts for 58% of Doubao's single-inference costs, with power consumption making up 29%. Nearly 90% of costs stem from actual resource consumption, rather than reusable infrastructure.

More users translate to faster cash burn.

Doubao's user growth curve mirrors an exponential cost curve. Upon its official launch in May 2024, daily Token consumption stood at 4 trillion. By December 2024, that figure had multiplied tenfold. By March 2025, daily usage surged to 12.7 trillion. By March 2026, it doubled again within three months, reaching a historic peak of 120 trillion. According to QuestMobile, Doubao's monthly active users hit 345 million during this period, with each user averaging 54.8 monthly interactions in the first quarter.

When hundreds of millions of users consume computing power by the billions of Tokens daily for free, "free" ceases to be a strategic choice—it becomes an unsustainable financial drain.

ByteDance has paid a hefty price. Zheshang Securities estimates ByteDance's 2025 capital expenditures at around 150 billion yuan, with 90 billion allocated for AI computing procurement. In 2026, this figure rose further to 160 billion yuan, with multiple media outlets suggesting the actual number nears 200 billion yuan. Insiders reveal that ByteDance's net profit plummeted by over 70% year-on-year in 2025, primarily due to massive AI chip purchases and foundational model R&D in the second half of the year, sharply narrowing fourth-quarter book profits. Douyin VP Li Liang attributed this to accounting fluctuations under international standards, noting that core revenue and business profits grew after excluding non-operating factors like option costs. Even so, the AI business's zero-revenue, high-loss reality remains unchanged.

Third-party estimates put Doubao's monthly net losses between 210 million and 270 million yuan. In 2025, domestic consumer-facing general large model applications collectively lost over 18 billion yuan, with CICC projecting losses to exceed 20 billion yuan in 2026.

The cost of "free" is now manifesting in the hundreds of billions.

02

Why Doubao?

In a market where nearly all domestic large models claim to be free, the first player to introduce price tags is bound to face maximum controversy.

Horizontally, Doubao faces significant pressure. Tencent Yuanbao explicitly stated it has no current plans to charge, focusing commercialization on Hunyuan API and MaaS enterprise services to sustain free personal use. Alibaba Qianwen maintains a free strategy, with its flagship model Qwen3.7-Max fully open across scenarios—its base model team never set DAU-based commercial KPIs. Kimi charges 49 yuan monthly, Zhipu Qingyan 49 yuan monthly, and Baidu Wenxin Yiyan 30 yuan monthly. Doubao's standard version, starting at 68 yuan monthly, sets a higher entry price than its peers.

Why did Doubao, with the largest user base, implement charges first? Scale itself is the reason.

Competitors aren't avoiding charges—their user volumes haven't yet reached a tipping point where offering services for free becomes unsustainable. Yuanbao has approximately 109 million MAUs, Qianwen around 203 million, lagging Doubao's 345 million by an order of magnitude. When user scale exceeds 300 million and daily Token consumption surpasses 100 trillion, each additional day of "free" burns billions more.

Doubao's rapid growth stems from ByteDance's aggressive push. Launched as Grace in August 2023, Doubao officially debuted in May 2024. Initially unremarkable, it had just 1.73 million MAUs by February 2024. ByteDance then funneled Group-wide traffic resources into it. Douyin added two entry points—long-pressing the like icon for queries and a Doubao assistant at the top of message lists. Combined with CapCut's video processing, Volcano Engine's training framework, and scenario linkages with Douyin E-commerce and Local Services, Doubao quickly infiltrated ByteDance's ecosystem. Internally, the strategy was clear: bring Doubao to users, not users to Doubao.

The approach yielded rapid results. By November 2024, Doubao's MAUs surged to 59.98 million, trailing only ChatGPT globally. In December 2025, DAUs officially broke 100 million—ByteDance insiders claim it had the lowest promotion costs among all billion-DAU products in company history. By March 2026, its 345 million MAUs neared half of Douyin's scale. The problem? Douyin monetizes via ads and e-commerce, while Doubao generates no revenue yet consumes computing resources exponentially.

Higher traffic means thicker bills. Doubao exemplifies how scale and losses amplify in tandem in China's AI sector.

03

More Than Just Recouping Costs

If Doubao's charging were merely to plug ByteDance's financial gaps, it would oversimplify the move.

Examine the three-tier pricing logic. The 68 yuan/month standard version covers long document summaries, PPT outlines, and high-definition watermark-free image generation for light office users. The 200 yuan/month enhanced version unlocks deep data analysis, HD video generation, and bulk content creation for self-media and senior professionals. The 500 yuan/month professional version offers dedicated computing power, API access, and priority responses during peak hours for enterprises and heavy users. The free basic tier retains daily chats, Q&A, basic copywriting, translation, and simple image generation—promising permanent free access without feature cuts, usage limits, or speed throttling.

The core of this pricing system is to make heavy computing users bear the costs. A simple text Q&A consumes a few hundred Tokens, but PPT generation or HD video output requires hundreds of times more computing power. Doubao's fees target high-computing, high-value productivity scenarios—the heavier the usage, the higher the payment.

In essence, this isn't traditional internet traffic monetization but a precise redistribution of computing costs.

Furthermore, Doubao's commercialization doesn't rely solely on subscription revenue. According to 36Kr, if progress goes smoothly, Doubao will further integrate e-commerce functions in Q3 2026 to enhance paid scenarios, driving traffic to Douyin Mall with subsidies and entering the operational phase in Q4. These moves lay the groundwork for long-term commercial returns beyond 2027. In 2026, Doubao won't prioritize paid user penetration as a core metric. ByteDance knows that even with just 1% conversion among 345 million users, the lowest 68 yuan/month tier could generate approximately 2.3 billion yuan in annual subscription revenue. More critical growth lies in e-commerce referrals, B-end API calls, and Volcano Engine's enterprise services.

Analysts estimate Doubao's annualized subscription revenue could reach $400 million to $700 million under neutral scenarios. Compared to ByteDance's thousand-billion-yuan annual AI investments, this is a mere drop in the bucket. Yet it proves a point: C-end AI users are willing to pay for value, provided the value is clear enough.

Epilogue

People's Daily once commented on "free": "Free" never means no cost—it merely suspends costs in midair, awaiting a moment to land.

When computing bills swell to structural loss levels, "free" becomes a looming threat hanging over the entire industry.

Doubao's move to implement charges first isn't about ByteDance harvesting users—it's about AI's cost logic finally overtaking subsidy logic. The internet-era formula of burning money for scale and monetizing via ads fails for large models. Each additional user means rigid computing expenses; greater scale deepens losses.

The competitive landscape is shifting. Alibaba Qianwen and Tencent Yuanbao may stay passive now to grab users with "free," but they'll face the same computing bills eventually. CICC predicts industry losses to exceed 20 billion yuan in 2026—no one can subsidize users indefinitely. Doubao, taking the first charge and enduring the most backlash, might also pioneer a sustainable commercial path.

Chinese AI is moving beyond subsidy-driven user acquisition. Henceforth, every service will carry a price tag, and each price must justify its value. Doubao's 68–500 yuan pricing tiers mark both a necessary step for ByteDance and a turning point for China's AI industry.

The era of free AI is over. But the era of good AI may just be beginning.