As of the first half of the year, Doubao, used by more than 200 million people a day, was generating less than 1 million yuan in daily revenue, mainly from e-commerce commissions.

By May this year, however, Doubao was consuming tens of millions of yuan a day, according to estimates based on Volcano Engine’s public API pricing, Doubao’s large-model gross margin and user behavior. Text chat is not expensive: at 15 to 20 minutes of average daily use per person, it costs only a few cents. But reasoning, as well as multimodal features such as image recognition, voice chat and video chat, require several times, or even dozens of times, more computing power.

That does not include the cost of building computing power centers to train Doubao’s models. A large AI data center often requires tens of thousands of AI chips, along with supporting infrastructure for the data center, power supply, networking, cooling and operations. According to the South China Morning Post, ByteDance plans to raise its 2026 capital expenditure by more than 200 billion yuan, equivalent to about 60% of its 2025 profit. If ByteDance can obtain more NVIDIA cards needed for training and secure more capacity for domestic inference cards, Doubao’s resource consumption would be even higher.

Doubao follows the mobile internet playbook: attract a large user base with a free app, then work out monetization later. But because AI is so costly, simply keeping Doubao running costs more than Bilibili’s entire operating expenses, even though Doubao’s total daily user time is less than one-eighth of Bilibili’s.

According to our information, two months ago, a senior ByteDance executive visited Anthropic. Soon after returning, ByteDance began adjusting its AI resource allocation, shifting focus from mass-market products such as Doubao to enterprise-facing products.

Over the past six months, Anthropic has shown that AI coding can turn huge infrastructure spending into returns: Claude Code launched in May 2025 and reached $1 billion in annualized revenue within half a year, then rose to $2.5 billion in February this year. By selling paid services to enterprises, Anthropic, which has only 30 million daily users, has seen its valuation climb to $965 billion, overtaking ChatGPT, which is used by nearly 900 million consumers each week.

ByteDance’s Seedance has also shown that the enterprise-services path can work. LatePost has learned that the video generation model has reached $2 billion in annual recurring revenue, or about 14.3 billion yuan, with monthly revenue exceeding 1 billion yuan, roughly enough to offset Doubao’s computing power costs. The vast majority of Seedance’s revenue comes from enterprise customers.

People familiar with the matter told us that ByteDance’s large-model data review team has expanded this year from about 1,500 people to more than 3,000, tasked with cleaning training data for coding models. Volcano Engine’s MaaS business has also been given greater priority, with ByteDance’s top leadership setting a goal of increasing revenue tenfold and accelerating overseas expansion.

ByteDance and Kuaishou have already learned this lesson themselves. Toutiao scaled advertising and growth in tandem in less than two years. Kuaishou moved half a step slower, waiting until the short-video boom was nearing its end before systematically selling ads, and paid far more to catch up. AI is moving faster, and companies will have even less time to figure out how to make money.

The First AI Business to Make Money: Seedance 2.0

Seedance currently has a gross margin of 70%, meaning that for every 10 yuan of API calls sold, server and inference costs account for about 3 yuan. Based on the model parameters of Seedance 2.0, one industry source estimated that a full training run costs several hundred million yuan.

By contrast, language models usually have more versions, more intensive iteration and longer training cycles, with total training costs that may be three to five times those of video models. Seedance is currently focused mainly on a single video generation model, making its training investment easier to amortize through future revenue. With this cost structure, “profitability is not difficult to achieve,” the person said.

Still, the gross margin of large models also depends on how ByteDance calculates depreciation. ByteDance does not primarily rely on long-term rentals of external computing power; its main costs come from self-built data centers and self-purchased chips. Whether chips are depreciated over several years has a significant impact on reported costs. Similar debates are playing out in Silicon Valley: AI chips iterate so quickly that there is still no consensus on whether depreciation should follow the traditional server cycle or a shorter technology cycle.

All of this has come from much larger investment. In 2025, the mainstream view in the industry was that continuing to train along the traditional DiT architecture, or diffusion transformer, left little room for model improvement, and most teams shifted toward post-training. Leading teams in the industry, with more capital and resources, have continued to invest in pushing pre-training models to the limit.

After nearly a year of optimization, Seedance 2.0 became the first video generation model to fully adopt MoE. With 200 billion parameters, it quickly became the world’s strongest video model after launching in February 2026.

Today, training both video and language large models requires expensive and hard-to-buy NVIDIA GPUs or Google TPUs. But once training is complete and the model is deployed, video models have lower requirements for inference chips. Language models generate one word after another and must look back at context at every step, placing heavy demands on memory and inter-card communication. The DiT architecture used by video models can process batches of data together and output in batches. It still consumes computing power, but relies less on inter-card communication, meaning many domestic chips can run it. That is the foundation of its high gross margin.

Seedance has now become an important production tool for AI live-action dramas and animated short dramas. Production companies use it to automatically break down scripts, generate storyboards and prompts, and then upload pre-set character and scene images to generate videos. Human work is reduced mainly to checking results and selecting usable shots.

According to our information, leading animated-drama companies spend about 50,000 yuan on Seedance computing power.

The Challenge of Charging Consumers

ByteDance’s earliest hopes for AI commercialization were pinned on mass consumers, using the old mobile internet formula: scale first, monetize later. By 2026, Doubao already had more than 200 million daily users. It was integrating e-commerce and local services to let users place orders through conversation, while also pushing paid subscriptions. But progress on both fronts has been limited.

The problem is that most shopping is about pleasure, not saving time. AI is good at filtering and summarizing on users’ behalf, but Pinduoduo founder Huang Zheng said as early as 2017 that shopping is mostly “non-purpose-driven”: like strolling through a mall with no clear goal, buying something when it catches your eye. Browsing, comparing, being persuaded by recommendations and taking time to choose are themselves part of the pleasure. He compared shopping to Disney, saying most people would not want AI to visit Disneyland for them.

Most people are also not as busy as AI engineers, willing to pay money to save time. Otherwise, more than 800 million people would not spend two hours a day on Douyin, 400 million on Kuaishou and more than 100 million on Xiaohongshu.

According to our information, Doubao generates only about 10 million yuan in e-commerce transaction volume per day. For a platform with 200 million daily active users, that is not much.

Doubao is also preparing a paid version, but getting mass-market users to pay for digital services is harder.

QQ Music, Qidian and iQiyi spent years getting hundreds of millions of people used to paying for licensed music, online literature and long-form video. But ByteDance then used its traffic advantage and stronger recommendation technology to disrupt the markets for online novels, short dramas and music with the free Tomato Novel, Hongguo Short Drama and Soda Music, which is free after users watch two ads, rapidly eating into the market share of companies such as Tencent.

Even productivity tools in China compete on free access. While Zoom was earning $1 billion in annual net profit from paid video conferencing, DingTalk, Feishu and Tencent Meeting kept core functions free for a long time, so companies did not have to pay for basic tools. Even highly paid financial analysts joke that when a Tencent Meeting hits its 40-minute limit, the standard practice in finance is to have someone else start a new meeting.

Competition among large models makes subscriptions even harder. DeepSeek is free and open-source, and is already good enough in many scenarios. Later entrants hoping to charge because they are “a little smarter” must prove that they are clearly better.

Seedance has given the industry hope that AI can make money, but it has not yet proved that this is a path of sustainable growth.

According to our information, Seedance’s revenue growth has slowed. Kuaishou’s Kling also saw ARR growth slow after reaching $500 million in early May. Both companies now have new customer groups beyond short dramas, but bringing more industries into AI video generation will require further improvements in model performance. NVIDIA processors capable of training large models remain scarce.

Jimeng, which has integrated Seedance, has reached about 2.5 million daily active users at peak. Its goal is to become a next-generation mass content consumption platform. For now, however, its users are still mainly professional content creators.

ByteDance is facing the same unavoidable question confronting every major company investing in AI: AI computing power centers have already burned through hundreds of billions to trillions of yuan in investment. How do they turn that into money? Seedance has proved that video generation can have high gross margins, but it still needs to move beyond the short-drama industry. Doubao has massive user scale, but struggles to produce meaningful revenue. In AI coding, its rivals are the world-leading Claude Code and OpenAI Codex.

In China, there is one more challenge: how to get users accustomed again to paying for services in a market trained on free products.