In early February 2026, China’s two internet giants, Tencent and Alibaba, fired the first shots in the year’s AI war. Tencent’s Yuanbao offered 1 billion yuan in cash incentives and launched a new AI social feature called “Yuanbao Pai,” while Alibaba’s Qianwen put up 3 billion yuan in benefits to encourage users to order food delivery inside its app.

ByteDance had already secured the top sponsorship slot for CCTV’s Spring Festival Gala that year through Volcano Engine, but amid the noise of red packets and subsidies, it was not the center of public attention. That changed a week before Lunar New Year, when its video generation model Seedance 2.0 was released and quickly gained traction among creators and in industry circles, pushing ByteDance back into the conversation. On February 14, ByteDance began rolling out Doubao large model 2.0 and the image creation model Seedream 5.0 Lite.

On Lunar New Year’s Eve, ByteDance brought Volcano Engine, the Doubao app, and the Seed family of large models onto the Spring Festival Gala stage. From stage visuals and robots speaking to broadcast support, AI permeated the country’s biggest mass entertainment event more fully than ever before.

Even the interactive mechanics were redesigned: viewers had to use the large models inside Doubao to generate images or text before they could receive red packet benefits. The red packets contained not only cash, but also tech products such as 3D printers, cars, drones and robots.

This was ByteDance’s Chinese New Year campaign.

A ByteDance employee told us that, compared with many companies that treat Chinese New Year simply as a growth opportunity, ByteDance wanted to use the moment to create a shared memory around AI and technology. It also wanted to signal that doing AI is not about building one or two apps, but about serving multiple industries. The tech products given away during the gala were all from customers of the Doubao large model, we understand.

Official data showed Doubao recorded 1.9 billion AI interactions on Lunar New Year’s Eve. Before the holiday, the closely watched video generation model Seedance 2.0 and the latest Doubao large model 2.0 Pro were also connected to Doubao’s “Expert” mode at launch, bringing advanced model capabilities into a product that users could access immediately.

Whether in its fit with everyday scenarios such as Chinese New Year or in the higher ceiling of its model capabilities, Doubao now has the traits of a national-level AI product.

In the technology industry over the past few years, the most convincing growth stories have never been built by stacking subsidies. After ChatGPT kicked off the global AI boom in November 2022, model iteration turned it into a super product with 800 million weekly active users. DeepSeek, which became popular more than two years later, has steadily maintained more than 30 million daily active users without any advertising spend or multimodal capabilities.

At ByteDance’s annual employee meeting at the start of the year, CEO Liang Rubo set the company’s 2026 keyword as “climbing the peak.” He said the AI era contains many important opportunities, and ByteDance must pursue the most important among them and climb the highest peaks.

“The real AI offensive is a hard decision to turn yourself into a technology company,” the ByteDance employee said. During Chinese New Year 2026, the company is trying to light up its own ChatGPT moment.

From Ink-Wash Stages and Robots to Computing Power Surges: A Different Chinese New Year Campaign

As one of the few truly national gala platforms, the Spring Festival Gala has been an important channel for internet companies to acquire new users over the past decade. Its partnerships with internet products have also been almost inseparable from “red packets.” Since 2015, major internet products including WeChat, Alipay, Taobao, JD.com, Douyin, Pinduoduo and Kuaishou have all handed out red packet benefits during the gala.

But at this year’s Spring Festival Gala, the first task given to title sponsor Volcano Engine was related to stage visuals.

A ByteDance employee told us that the gala’s director team had been exploring ways to combine the latest technology with traditional culture and art. As the exclusive AI cloud partner, Volcano Engine and ByteDance behind it were not only participating as sponsor and title partner, but also embedding AI and cloud computing capabilities into the creation and production process of the gala’s programs.

In the program “Song of Riding the Wind,” a national-treasure ink painting of galloping horses appeared behind the singer. During preparations, the director team was not satisfied with using it as a static backdrop. They wanted the horses to move through the stage space and interact with the singer.

To achieve that effect, they initially tried some overseas models, but most failed to capture the spirit of ink painting. Domestic models faced similar limits: high-quality training data for ink-wash style was scarce, and the generated images easily drifted in style. Motion consistency was even more difficult: the horses had to run smoothly, while every frame still had to match the brushwork and structure of the original painting without deformation. The number of horses and the movements of each horse also had to remain consistent throughout.

The turning point came with Seedance 2.0 from ByteDance’s team. The new video generation model could better follow physical laws, making the horses’ gait closer to reality, with natural joints and coherent motion. It could also follow the director’s requests well, carrying out subtle scheduling instructions with adverbs, such as “run a little slower” or “make the mane flutter more lightly.” It also delivered another important breakthrough: the model was no longer limited to learning the style of a single image, but could learn from large amounts of multimodal material across different styles at the same time.

A member of ByteDance’s Spring Festival Gala project team told us that they fed the model the director team’s sketches, ink paintings and a large amount of horse-related video material, then used reference-based generation to keep the visual style and motion stable under the same standard. They also avoided writing prompts directly, instead using Seedance 2.0 together with the image generation model Seedream: Seedream first generated the key frames, then Seedance generated dynamic video under those key-frame constraints, achieving both stylistic consistency and coherent motion.

Showing the ideal effect was only the first step. Every Spring Festival Gala program iterates at a very high frequency, so the ByteDance team often found that as soon as it delivered a version the directors liked, the director team would immediately ask: “Can we take it one step further?” These programs also had to go through repeated review and acceptance, which meant the team had to produce a better version each time. “The overall cadence basically had to move forward by the week,” the ByteDance project member said.

Soon, a new problem appeared. Although Shanghai Animation Film Studio classics such as “Havoc in Heaven” and “The Tadpoles Search for Their Mother” represent the peak of hand-drawn full animation, in the context of AI digital generation, making ink wash, an art form with “no boundaries,” avoid flickering or breaking down under fast, large-scale motion at modern broadcast standards of high frame rates and high resolution is a recognized industry challenge.

Considering that mainstream video generation models usually output video at 720p or 1080p and 24 frames per second, the ByteDance team focused its breakthrough on the post-production pipeline, using Volcano Engine Video on Demand’s image-quality enhancement capabilities to upgrade the footage. It first used super-resolution technology to enlarge smaller images to 6K or 8K without changing the content, then used frame interpolation to raise the frame rate from 24 FPS to 50 FPS. This processing was not a one-size-fits-all general algorithm, but could optimize each frame based on its image quality and content characteristics. In the end, they solved the problem.

Beyond stage visuals, the ByteDance team also brought AI capabilities into multiple parts of the Spring Festival Gala. The robots that appeared on stage were connected to Doubao’s vision large model, text large model and speech recognition model, while the speech side used Doubao’s synthesis and voice-cloning models. The models gave the robots not only a “brain,” but also a “mouth” and “ears.”

They also collected 3D material of the dancers in advance, then used spatial video technology to turn the dancers into 3D digital avatars, creating 3D human clones on stage that made it hard for viewers to tell what was real. Finally, they added ByteDance’s self-developed 4D Gaussian algorithm and used the Doubao large model to optimize lighting and shadows.

The Doubao app was the core vehicle for audience interaction at this year’s Spring Festival Gala. Unlike the common red-packet shaking and grabbing mechanics on internet products in past years, this year’s interaction required users to first use large models in Doubao to generate avatars or New Year greetings, and only then grab red packets. The change sharply raised costs. In previous red-packet interactions, the main expenses were concentrated in networking, I/O and real-time connections; now, huge computing power requests and computing power spending were added on top of the existing load.

The difficulty of managing computing power resources also rose sharply. A member of ByteDance’s Spring Festival Gala project team conservatively estimated that for a user to generate one New Year greeting or one image, a single request required 10 TOPS, or 10 trillion operations per second. By contrast, the computing requirement for similar interaction requests in the past was only about 1/100000 TOPS. The gap in computing power demand was a full 1 million times. At the same time, on Lunar New Year’s Eve, they also had to handle large-model calls from Chinese New Year campaigns across products such as Douyin and Toutiao.

Because time was tight, ByteDance could not withstand the computing power surge simply by temporarily adding machines. Fortunately, Volcano Ark stepped in. It is ByteDance’s central console for unified computing power scheduling and has built up resource scheduling capabilities over time across various high-concurrency scenarios for ByteDance and external customers. One feature of Volcano Ark is that it places inference, training and offline tasks into the same resource pool for unified management. When Chinese New Year computing power and traffic peaks arrived, the system could shift delay-tolerant workloads to off-peak periods, freeing more resources for various holiday activities.

This was not easy to do. It involved massive numbers of scenarios, multiple server rooms, many machine types and many model categories, while also requiring continuous dynamic resource allocation. A Volcano Engine employee described the process as “bin packing”: on one hand, different traffic streams have different latency requirements and different dependencies on heterogeneous hardware, creating many constraints and a huge solution space; on the other hand, large-model inference is unlike traditional CPU services, where resource types are relatively uniform, and a single inference service may mix multiple types of hardware. More troublesome still, GPU migration is not a single-point action. Supporting resources such as storage, networks and upstream load balancing also have to migrate in coordination; otherwise, the computing power may move, but it still cannot absorb the traffic.

“For ByteDance, sponsoring the Spring Festival Gala is no longer a simple growth campaign. It is more like a real technical trial,” a ByteDance employee said.

In the AI Era, Growth Through Red Packets Is Breaking Down

The shift in how ByteDance worked with the Spring Festival Gala this year seems to show once again that the growth logic of the AI era is being rewritten.

In the mobile internet era, giants did not rise by leading innovation at every moment. They rose by systemically building similar products after someone else had validated a need, acquiring users at higher efficiency and scale, then rapidly iterating on the experience based on user feedback. Better experience brought more revenue, which was then spent on user acquisition, creating a loop. Spending money was the most direct growth tactic.

“The core is a flywheel of extremely fast growth, extremely fast iteration and extremely fast monetization,” said a product manager at an internet company.

For internet products, the faster they acquire new users, the faster the product experience improves. “In most cases, recommendation algorithms discover that you and another user have similar preferences, then show you what that user likes,” said a recommendation product manager. As users increase, the data available for algorithms to learn from and discover correlations also increases. This applies to recommending videos and novels, as well as product cards and livestream shopping rooms. The rise of WeChat Pay is another example: red packets turned payments into part of social interaction. To participate, users had to link a bank card and be able to transfer money. The more people used it, and the more often they used it, the more it gradually became the default payment tool.

Transaction products benefit in the same way. A ride-hailing platform subsidizes both sides at once: drivers to increase supply, and passengers to stimulate demand. As driver density rises, passenger wait times fall. As orders become denser, drivers spend less time driving empty, the platform’s fulfillment costs decline, and the user experience improves.

The costs of internet products also do not rise linearly with user growth. Incremental expenses are mainly in bandwidth, storage and machine resources, and they are diluted as the user base grows.

But AI products are almost the opposite of internet products.

In early 2025, ByteDance CEO Liang Rubo said at an all-hands meeting that Doubao had not shown the internet-product trait of “getting better the more people use it.” This was partly because chatbot products are not social networks or platforms.

User growth also brings limited new data. For a short-video product, as long as users keep swiping up and down, they generate sets of data for recommendation algorithms to optimize. But when a chatbot product generates a response, users usually have the motivation to click feedback buttons only when the answer is extremely bad. Even if data can be collected, there is no guarantee it will make the underlying model smarter. “Most users’ questions are highly repetitive and not very deep, so they can’t improve model capabilities,” said an AI product manager at an internet company. “For example, in coding, companies will look internally for programmers to write cases.”

The professional capabilities of large models have long surpassed those of most humans. They cannot improve themselves by collecting data from ordinary people, just as AlphaGo did not need to play against ordinary people to improve. Its later variant, AlphaGo Zero, did not even use data from world champions; it trained machines against machines and beat everyone. Even though ChatGPT’s monthly active users are more than 100 times those of Claude, that has not made OpenAI’s models even twice as good as Claude.

AI also cannot dilute costs through user scale the way internet products can. A secondary-market investor who has studied AI in depth once calculated that for a mainstream AI product today, serving 100 million daily active users would cost tens of millions of yuan per day in model inference costs. That calculation does not factor in new Agent products. If products like Manus become popular, the computing power required by each user per day could multiply several times again.

The most critical issue is monetization. Overseas, ChatGPT, Gemini and Claude have poured in huge investments to support complex computation, and users must pay: lower-tier plans cost $17 to $20 a month, while higher-tier plans can reach several hundred dollars a month. But few Chinese users are willing to pay that much for software services. Companies such as ByteDance and Alibaba more often recoup their AI spending through cloud services, packaging model capabilities as cloud APIs, hosted services and industry solutions, with enterprises paying by calls, concurrency, storage and computing power duration.

According to official data, in December 2025, the Doubao large model on Volcano Engine processed more than 50 trillion tokens per day on average, up more than 200% in six months. The growth came not only from the rapid development of ByteDance’s own AI apps such as Doubao and Jimeng, but also from a group of external customers using large models deeply: more than 100 customers had cumulatively used over 1 trillion tokens, twice as many as global cloud giant AWS.

At present, improvements in AI product experience come almost entirely from stronger underlying models. The AI code editor Cursor broke out because it connected to the Claude model family after its coding capabilities improved sharply. OpenAI’s Deep Research delivered an impressive experience because the underlying model learned long-chain reasoning and step-by-step problem solving.

If large models can keep improving, many finely tuned product capabilities will become part of massive models, and users will be able to get the results they want simply by saying a few sentences. In that case, the large model itself is the ultimate product. The success of ChatGPT and DeepSeek both shows that the old playbook of moving users through paid traffic and subsidies is becoming increasingly strained. What truly creates distance is still the hard strength of models and product experience.

The Only Answer: Become a Technology Company More Decisively

At a CEO face-to-face meeting in 2017, a ByteDance employee asked Yang Zhenyuan, the company’s algorithm technology lead, where ByteDance lagged behind BAT in artificial intelligence. Yang replied, “Toutiao was originally an artificial intelligence company.” He further explained that information distribution requires artificial intelligence, and Toutiao wanted to use a combination of machines and humans to improve efficiency across creation, distribution, discussion and every other link.

At the time, AI already had some applications in ByteDance’s products, including recommendation, content creation, Douyin’s AR effects, content moderation, advertising systems, automatic cover generation for Toutiao accounts, and image selection features in Time Album and Dongfang IC.

At the end of 2022, ChatGPT’s arrival triggered a new AI boom. ByteDance became one of the Chinese technology giants investing most heavily in this round. It quickly ordered large numbers of GPUs, formed a new AI department, and began catching up with Silicon Valley companies across infrastructure, data, models, products and talent. Two years later, ByteDance’s model count, iteration speed and performance had improved markedly.

Technological leaps rarely happen overnight. They require firm, long-term investment behind the scenes. A ByteDance employee recalled that when Seedance 1.5 was released, outside assessments were not very high: characters tended to break down during motion, and details were not stable enough. Later, Seedance 2.0 showed a clear improvement, and some people guessed that the team had adjusted its benchmarks, putting metrics that affected user experience more heavily further toward the front.

“In fact, everyone overlooked the importance of foundational work,” the person said. Take motion breakdown as an example: the team had to repeatedly validate different approaches and identify more effective paths. This kind of work is time-consuming, but it is a foundational capability that has to be built. “It’s like a very smart child who may be able to solve college-level problems in the future, but if his basic knowledge isn’t complete and you force him to do it now, the result definitely won’t be good.”

Another example is Doubao’s speech synthesis model, Seed TTS. In 2024, when the team was building Seed TTS 1.0, its goal was to enable the model to naturally and smoothly replicate a speaker’s characteristics, especially maintaining similarity and rhythm across languages. After 1.0 reached a usable level, new problems emerged: the model’s tone was too flat. That problem could not be solved quickly at the time, so when developing version 2.0, the team focused on integrating understanding and reasoning capabilities into the end-to-end speech pipeline. It was not until Seed TTS 2.0 was released in mid-2025 that the model gained stronger emotional expressiveness.

ByteDance is also one of the few Chinese technology companies willing to invest resources in basic research. A person close to ByteDance’s senior management once told us that ByteDance’s infrastructure, or engineering capabilities, are already stronger than any other company in China. But compared globally, the biggest issue is the lack of people like those at OpenAI who can set directions and explore the frontier, such as GPT-4o and Sora.

“China did not have real corporate research institutes in the past because private companies had limited profitability. Now they can finally give it a try,” the person said.

In late January 2025, ByteDance formally established a research project code-named “Seed Edge,” with the core goal of conducting longer-term, more fundamental AGI frontier research beyond pretraining and large-model iteration. The plan set a more flexible evaluation mechanism: ByteDance normally reviews performance every six months, but Seed Edge will conduct its final evaluation after the project achieves breakthrough progress.

ROI remains important when ByteDance evaluates AI projects, but the time horizon has grown longer. A member of ByteDance’s AI product team told us that ByteDance now uses the per-user value of an AI product after a certain period as an evaluation coefficient to estimate future returns. Different products have different evaluation cycles, with longer ones extending to as much as five years. But ByteDance has not made this evaluation method mandatory.

Business and technology history have proven countless times that winning users, markets or larger opportunities always depends on technological leaps and hard product strength.

More than 20 years ago, worried that Microsoft would enter the search market, Google created an internal “Finland Plan.” The core of that plan was not to defend its territory, but to force itself to keep building more creative products and push the experience to the extreme. Even on a home field such as Windows, it had to make a browser better than Microsoft’s built-in tools. Ultimately, without an advantage in product iteration or network effects, Google relied on outstanding engineering capabilities to build Chrome and win the war.

In its competition with Microsoft, Google also forged a powerful monetization system, AdSense, tightly binding steady cash flow and an ecosystem network to itself, which over the following years gradually settled into a hard-to-shake moat.

In 2024, OpenAI co-founder and former chief scientist Ilya Sutskever said, “The 2010s were the age of scaling, and now we’re back in the age of miracles and discoveries.” For ByteDance, it took 13 years to become China’s largest internet company. Now it has encountered a new game with a high enough ceiling, one that requires it to prove itself again through real technology and products. It will not miss that opportunity.