China's largest technology companies turned the 2026 Lunar New Year holiday into the country's first large-scale, premeditated AI application war, collectively pouring close to RMB 10 billion yuan (approximately US$1.38 billion) into subsidies, advertising and infrastructure to drive adoption of their artificial intelligence products — yet the campaign exposed as many vulnerabilities as it demonstrated strengths.

The mobilization was unprecedented in scale. ByteDance cancelled employee holidays at its Beijing headquarters, where office towers remained lit through Lunar New Year's Eve. Alibaba Group, Tencent and Baidu did the same at their respective campuses. For the first time in two decades of Chinese internet competition, the industry's defining battles were no longer being fought over e-commerce, payments or short video — but over who controls the AI interface to daily life.

The campaign produced dramatic, if fleeting, results. Tencent's AI assistant Yuanbao saw daily active users surpass 50 million, with monthly actives reaching 114 million. Alibaba's Qwen app recorded approximately 75 million daily active users at peak on February 7, up from around 10 million prior to the holiday. ByteDance said its Doubao AI recorded 1.9 billion interactions on Lunar New Year's Eve alone. But the central question facing all three companies remains unanswered: how many of those users will return once the subsidies are gone.

The Battle for the Spring Gala Stage

The strategic logic behind the holiday offensive traces back to a single precedent: in 2015, WeChat used its partnership with China Central Television's Spring Festival Gala to popularize digital red packets, adding approximately 200 million linked bank cards in a matter of weeks. Every major Chinese internet platform has since sought to replicate that moment.

For the 2026 Gala, CCTV's production team positioned the event as the "inaugural digital-intelligence Spring Festival Gala," seeking partners who could integrate AI into the broadcast experience. Tencent was approached but declined, wary of large-scale marketing spend following years of cost discipline. Alibaba hesitated — its Qwen app had only launched a month earlier and was still iterating rapidly. ByteDance moved decisively, finalizing the partnership in under two weeks from initial bid to signed agreement, choosing to brand the sponsorship under its cloud computing unit Volcano Engine to reinforce its identity as a technology company rather than a consumer app maker.

The speed of ByteDance's move alarmed its rivals. By late 2025, Doubao's daily active users had already surpassed 100 million; the team was targeting 200 million by the end of the holiday season. A gap of that magnitude, against Yuanbao and Qwen still operating in the tens of millions of daily actives, would represent a potentially decisive lead.

Alibaba responded by securing exclusive sponsorships of four regional satellite broadcasters — Dragon Television, Zhejiang Satellite TV, Jiangsu Satellite TV and Henan Satellite TV — at a total cost of less than one-tenth of ByteDance's CCTV deal. It also secured product placement in director Zhang Yimou's Spring Festival theatrical release. On February 2, Alibaba announced RMB 3 billion yuan (approximately US$414 million) in subsidies to encourage consumers to use Qwen for food delivery, movie tickets and travel bookings during the holiday. Tencent, regretting its initial withdrawal, negotiated a premium advertising slot on CCTV immediately before the Lunar New Year's Eve broadcast of the flagship news program, along with a partnership with Bilibili, one of CCTV's official streaming partners for the Gala.

Model Readiness Shaped Tactical Choices

Beneath the marketing spectacle lay a more consequential divergence: the uneven state of each company's underlying AI models directly determined how each company could compete.

ByteDance's research unit Seed had spent much of 2025 building toward a flagship trillion-parameter model, Doubao 2.0, under the leadership of Wu Yonghui, formerly a research vice president at Google DeepMind. The effort hit technical setbacks — accumulated infrastructure debt in training data and model architecture required resolution before the model could scale reliably. The team worked through the problems in time for a pre-holiday release, and was rewarded with an unexpected viral breakthrough: its video generation model Seedance 2.0 spread internationally, with tutorials circulating on overseas social media instructing non-Chinese users on how to access the product, prompting industry observers to describe the phenomenon as a reversal of the usual technology diffusion dynamic.

Alibaba's model development also encountered complications. The company released a preview of Qwen 3-Max-Thinking on January 26 and open-sourced Qwen 3.5-Plus on Lunar New Year's Eve, but the higher-performing Max version was delayed. DeepSeek made only incremental updates during the period; its next major release, V4, is expected around March.

Tencent faced the most constrained position. Its model team had only completed a reorganization in the fourth quarter of 2025, and newly recruited researcher Yao Shunyu, who joined in November, could not deliver a competitive flagship model within the holiday window. Instead, Tencent concentrated resources on image generation — an area with lower training complexity — releasing its Hunyuan Image 3.0 image-to-image model on January 26. According to a person familiar with the matter, user engagement and retention improved measurably following the release. Tencent CEO Pony Ma personally checked with the Yuanbao team to ensure GPU resources were sufficient to prevent performance degradation during peak traffic.

The video generation ambitions of the Spring Festival Gala partnership itself tested ByteDance's technical limits in real time. The team had planned to embed stylistically consistent animated horse characters across multiple broadcast segments, but Seedance's consistency in transferring a character's texture and appearance across different scenes remained unstable. A workflow built on open-source models was deployed as a stopgap — but required managing more than 30 process steps and could only reproduce roughly half of the intended visual fidelity. The problem was resolved four days later as Seedance's training advanced to the next stage.

Subsidies, Crackdowns and the Limits of Manufactured Growth

The competitive intensity produced friction almost immediately. Tencent fired the opening shot on January 25, when Yuanbao announced RMB 1 billion yuan in cash red packets distributed through task-based interactions — generating images, writing holiday greetings — with results shared via WeChat groups. The campaign's viral spread exceeded internal projections, causing service outages that required emergency server reallocation.

Three days later, WeChat's official account issued a notice stating that Yuanbao's campaign had induced users to share links at a frequency that constituted harassment. WeChat restricted direct in-app access to Yuanbao links. Yuanbao pivoted to a password-based redemption system to sustain virality within WeChat groups, but with diminished effect. WeChat's public relations director posted a response on Weibo stating that user experience takes priority and that rules are applied consistently — accompanied by a meme suggesting the platform was prepared to enforce them aggressively.

Alibaba's RMB 3 billion yuan subsidy campaign, offering RMB 25 per user for food and lifestyle purchases made through Qwen, launched on February 6 and was blocked on WeChat within one hour. Despite the restriction, the campaign generated 15 million orders on its first day — 15 times internal projections — overwhelming Alibaba's logistics infrastructure and requiring manual traffic throttling. In subsequent days, Qwen was adding more than 20 million incremental orders per day to Alibaba's flash-commerce platform Taobao Flash Shopping, straining fulfillment capacity across the system.

The regulatory environment intervened before the Lunar New Year's Eve climax. On February 13, China's State Administration for Market Regulation summoned representatives from Alibaba, ByteDance, Baidu, Tencent, JD.com, Meituan and Taobao Flash Shopping, warning against what it characterized as "involution-style" competition. According to a person with knowledge of the meeting, the Spring Festival AI campaign was a primary focus of the discussion. Alibaba was subsequently called in for a separate bilateral conversation, during which the company characterized its RMB 3 billion yuan outlay as a customer acquisition cost equivalent to RMB 25 per new user. The day after the meeting, Qwen launched a second round of subsidies.

Following the regulatory intervention, the tone of the campaign moderated. By Lunar New Year's Eve, the dominant mechanic had simplified to direct cash red packets tied to app opens at specified times. ByteDance, which had held its primary Doubao incentives in reserve for the Gala broadcast, distributed prizes including drones, robots and automobiles — all products from Doubao enterprise clients — alongside cash. Tencent's Yuanbao offered users the ability to generate gold-highlighted WeChat Moments posts, a social differentiation mechanic designed to sustain organic sharing without triggering another platform enforcement action.

The Economics of Scale Remain Unsolved

The holiday campaign illustrated a structural tension that no amount of short-term subsidy spending can resolve: AI inference costs do not scale the way traditional internet services do.

One ByteDance engineer estimated that generating a single holiday greeting or AI image requires approximately 10 TOPS (trillion operations per second) of compute — roughly one million times the computational load of a comparable pre-AI interactive request. According to a secondary market analyst who has studied the sector in depth, serving 100 million daily active users on a mainstream AI product currently costs tens of millions of renminbi per day in inference alone — before accounting for more compute-intensive agentic applications of the kind gaining traction internationally.

The contrast with global peers is instructive. OpenAI's ChatGPT, Google's Gemini and Anthropic's Claude have all raised substantial capital to fund inference infrastructure, and charge users between US$8 and several hundred dollars per month for premium tiers. Chinese consumers have historically shown limited willingness to pay comparable amounts for software subscriptions, leaving the country's AI developers without a clear path to cost recovery at scale.

The hardware constraints were equally concrete. Across the companies that participated in the campaign, GPU capacity was broadly sufficient, but available data center rack space was not — a bottleneck that constrained deployment flexibility during peak demand periods.

A War That Has Only Just Begun

The Spring Festival campaign resolved nothing strategically, and the companies involved know it. "We know very clearly that after the holiday, the daily active numbers that spiked will start to come down," said one person at Tencent.

Each company enters the next phase of competition carrying both assets and liabilities. ByteDance's Seedance 2.0 has established a credible position in video generation, but its text-based large language model has yet to fully prove itself against the field. Tencent's strategic bet is on the first generation of large models to be trained under Yao Shunyu's leadership, while WeChat — the platform that made Tencent's 2015 red packet moment possible — continues to search for its own moment of AI leverage. Alibaba is building an AI agent team centered on Qwen and deepening integration with its broader ecosystem of commerce, payments and logistics services.

The historical analogies the industry reaches for are instructive. In 1995, Microsoft recognized that the browser threatened to render its operating system irrelevant and responded by embedding Internet Explorer directly into Windows. Eight years later, Google, facing Microsoft's entry into search, launched what it called the "Finland Plan" — committing to build products so superior that users would choose them even within a competitor's ecosystem, ultimately producing Chrome. In both cases, the decisive factor was not marketing spend but a product that was genuinely better.

The more recent precedents are equally clear. ChatGPT reached 800 million weekly active users through successive model improvements, not subsidies. DeepSeek maintains more than 30 million daily active users without advertising or multimodal features. The Spring Festival campaign demonstrated that China's AI giants can manufacture attention at scale. Whether they can build the underlying technical capability to retain it is the question that will define the next phase of the competition — and the answer will not be known until the subsidies have fully run out.