The year 2026 is poised to be a fiercely competitive period in China's AI market. Major technology firms will engage in critical battles for dominance. Alibaba's substantial strategic investment in its "Qianwen" project will become more pronounced, exemplified by a massive campaign offering 3 billion yuan in incentives to users. Similarly, Tencent's distribution of "Yuanbao" red packets is merely the beginning of its significant initiatives.
ByteDance, however, presents a formidable challenge, having already secured a leading position in consumer-facing AI with its vast user traffic. The launch of the "Doubao Phone" is just one of its experiments to expand boundaries. While major tech companies previously had room for hesitation in their AI strategies focused on model development, 2026 represents the final window of opportunity for them to compete effectively in the consumer AI market, especially given Doubao's high daily active user count. The significance of this contest rivals past industry clashes in ride-hailing, digital payments, and food delivery.
For mid-tier companies, the narrative of the "AI Six Tigers" has reached a conclusion with Zhipu AI and MiniMax pursuing IPOs in Hong Kong, Moonshot AI and StepFun securing funding, and a new wave of model releases. The question of whether these firms have the capacity to advance toward Artificial General Intelligence remains uncertain, but achieving self-sufficiency is urgent. Each must find a distinct path, particularly in commercialization. As one executive from these companies noted, rushing into an IPO without a solid foundation can lead to severe scrutiny and immediate stock price repercussions in secondary markets.
Nevertheless, a partner at a prominent U.S. dollar fund stated that no mature business model currently exists, and several years of exploration are still required. The diversification of business models will inevitably lead to variations in organizational structures. A fundraising lead at a large model startup outlined five primary models: consumer subscriptions and ad-based revenue, enterprise API sales, customized B2B and B2G solutions, pay-for-performance, and integrated software-hardware systems. Each model necessitates different strategies; for instance, consumer-focused ventures must look overseas, while enterprise-focused ones require strong sales capabilities.
The most vibrant and promising sector remains AI applications developed by startups. Many aspire to emulate Manus's success, achieving significant annual recurring revenue and attracting lucrative acquisitions by larger firms. However, this area is also the most unpredictable. In 2026, the AI landscape will continue to be monitored with keen interest and an open mind regarding entrepreneurial ventures. Based on interviews with dozens of frontline entrepreneurs, investors, and employees at major tech companies, ten key trends for 2026 have been identified.
ByteDance: Maintaining Advantages and Chasing Global AI Leadership
A former Flow employee suggested that the optimal form for an AI entry point remains undetermined. Doubao's key decision was to democratize AI capabilities by prioritizing multimodal functions early on and capitalizing on growth opportunities in 2025. ByteDance achieved a significant milestone by attracting top-tier talent, including Wu Yonghui, building a robust team that places its foundational models in China's top tier. The core challenge for 2026 will be retaining this talent as competition for AI assistants and models intensifies.
A strategic source at a major tech firm indicated that ByteDance's boldest move in 2025 was launching the Doubao Phone, which disrupted the mobile internet and handset markets, forcing competitors to accelerate their plans. The long-term challenge for Doubao is to maintain leading AI model capabilities while integrating the assistant into users' daily lives, particularly in offline services like e-commerce and food delivery. Although ByteDance has made inroads in these areas and benefits from a nimble organizational structure, effective internal coordination remains difficult.
Alibaba: Creating a New Consumer AI Gateway
A Qianwen employee revealed that the product is not only an external AI entry point but will also serve as an underlying AI platform for various Alibaba businesses. The product strategy focuses on differentiation, targeting professional office scenarios initially, with the long-term goal of becoming a primary AI gateway. The challenge in 2026 lies in understanding the diverse AI needs across Alibaba's complex business ecosystem and improving collaboration between the independent Qianwen team and other business units.
A strategic insider at a large company described 2026 as the year when major players will truly compete in the ChatGPT-style arena. Alibaba's ability to swiftly transition from Quark to Qianwen demonstrates its strong organizational capabilities. Despite existing products with over 100 million DAUs, the market is far from saturated, giving Qianwen room to grow. While current AI assistants are highly homogenous, the winners will be those who iterate rapidly, understand user needs, and excel in long-term operations.
A former Alibaba employee noted that the AI era offers a chance to redefine entry points, moving beyond traditional search, social, and e-commerce gateways. However, the window of opportunity may be brief. The establishment of the Qianwen consumer business group reflects Alibaba's use of organizational change to speed up decision-making, opting to test new products in the market rather than through prolonged internal coordination.
Tencent: Catching Up in AI Applications and Models
A former Yuanbao employee emphasized that the importance of "model as a product" was recognized before integrating DeepSeek and has since been reinforced. Throughout the year, Yuanbao concentrated on enhancing model capabilities and deepening product-model integration. Early on, it targeted highly educated users for differentiation, as their high standards and influence can shape broader adoption trends.
A Yuanbao team member stated that the product aims to reduce its reliance on DeepSeek. Currently, Tencent's Hunyuan model lacks a dominant market position, and while Yuanbao's search service utilizes both Hunyuan and DeepSeek, most users still prefer DeepSeek as the default. The recent merger of TEG's search team with Yuanbao's is expected to improve cooperation efficiency and may lead to a unified "Yuanbao Search" in the future, integrating model-based search functionalities.
A strategic analyst pointed out that Tencent's cautious pace contrasts with faster-moving rivals. Its Hunyuan strategy might avoid direct competition in foundational models, instead focusing on differentiated areas like Agent models. Tencent needs to demonstrate more convincing results in self-developed models. The delayed deep integration of AI into WeChat is due to the immense privacy and security challenges of a national-level app. Initiatives like the "Yuanbao Pai" allow for innovative experimentation without disrupting existing users. In 2026, clarifying the strategic roles of Yuanbao and WeChat will be crucial to leveraging Tencent's product strengths.
Baidu: Building AI-Native Organizations and Overcoming Application Hurdles
Shen Dou, Executive Vice President of Baidu and President of Baidu Intelligent Cloud, stated that AI is initiating a "super cycle" with value surpassing the internet era, potentially restructuring entire industry chains and unlocking a multi-trillion-yuan market. Intelligent agents are key to industrial application, showing breakthrough efficiency in programming optimization, digital employees, and industrial SOPs. Enterprises need to build AI-native organizations, potentially leading to flatter structures with decision-makers and intelligent agents.
Wang Ying, Vice President of Baidu and head of the Personal Super Intelligent Business Group, highlighted three major user pain points: cognitive偏差 and AI hallucinations, implementation gaps where AI cannot fully execute tasks, and fragmented experiences requiring switches between tools. A true super personal intelligent agent must resolve these issues to empower users as "super individuals." Baidu aims to develop Baidu Wenku and Baidu Wangpan into such agents, offering personalized, free, and universal capabilities.
Ping Xiaoli, Vice President of Baidu overseeing e-commerce and digital human services, discussed the evolution of digital humans. The 1.0 era featured basic virtual appearances; 2.0 introduced hyper-realistic digital humans with high-precision cloning and interaction capabilities, now mainstream. Baidu recently launched persuasive digital humans in the 3.0 phase, capable of thinking, decision-making, and coordinating multiple agents. In the future, digital humans with world knowledge and tireless operation will surpass real humans in various scenarios.
Model Commercialization: Earning High-Quality Revenue
An executive at a model startup outlined five business models. Consumer subscription models often target overseas markets due to low domestic payment willingness. Selling APIs is essentially an extension of cloud services, and with cloud providers likely to drive prices down, it can only serve as a short-term supplement. Customization requires not only technical delivery capabilities but also strong relationships. Pay-for-performance and integrated software-hardware offer opportunities for startups, with Physical AI holding great potential for new traffic entrances. However, both demand high capabilities: strong model performance for pay-for-performance and comprehensive multimodal models with delivery experience for integrated solutions.
A member of a model startup observed that since 2023, many firms suffered from "OpenAI disease," aiming to become "China's OpenAI." By 2025, fewer companies made such claims, with more referencing Anthropic, as resources proved insufficient for expansive product matrices. With limited resources, priorities shifted towards reasoning and coding capabilities, which have clear market demand. Future model differentiation will depend on each company's resources, strengths, and downstream client needs.
A partner at a leading U.S. dollar fund commented that no AI business model is mature yet. Product forms are still evolving as model capabilities, like video consistency and understanding, develop. While chatbot-style products commonly use subscriptions, OpenAI is exploring ad-based models for broader reach. Several years of exploration are needed for mature商业模式 to emerge.
Finding Scenes: Monetizing Vertical and Niche Areas
Liao Qian, CEO of极致上下文, argued that universal agents will not dominate. Task definitions and interaction modes vary greatly across contexts, making general agents inefficient and cost-oriented. Vertical scenarios allow clear task definitions and industry standards. Startups should focus on information production scenarios, providing end-to-end services rather than tools, and avoid consumption scenes like entertainment and social media, which are dominated by large firms.
Deng Jiang, former partner at Baichuan Intelligence and founder/CEO of AI healthcare company缘启智慧, noted that startups have two advantages over large companies: deeper technical focus in vertical areas and technological independence. Not all scenarios welcome large tech firms due to potential data security and competition concerns. Each specialized field, such as dermatology with its vast patient base, offers substantial market opportunities if capabilities are deeply developed.
Wang Ming, CEO of攀峰智能, predicted that 2026 will mark the beginning of pay-for-performance for agents. Traditional SaaS models charge subscriptions regardless of user outcomes, whereas the new economic model ties revenue to task success shares, aligning incentives with user ROI. This shift forces startups to prioritize features that directly help users generate income. Once proven, agent adoption can spread rapidly due to low user decision costs.
Financing: IPO Benefits and Pitfalls
A partner at a top U.S. dollar fund viewed Hong Kong IPOs as an opportunity to improve the primary market environment. Historically, China's primary market struggled to support long-term, large-scale R&D not focused on net profit. Without IPOs, fundraising efficiency for large model companies would remain low. Friendly HKEX policies and more listings provide exit channels and fair international pricing, potentially revitalizing the primary market and fostering innovation. Only with such a cycle can the primary market support hundred-billion-dollar tech companies, unlike the early IPO pressures faced by some firms.
Zhang Jinjian, founding partner of绿洲资本, urged companies to pursue global innovation rather than regional efforts. While AI is often seen as a U.S.-China game, overseas investors need clear Chinese AI investment targets. Each listing of an AI or embodied company opens a window for global capital to recognize China's innovation, as evidenced by revenue sources in prospectuses. Increasing foreign investor interest in direct investments in China underscores that capital will flow where innovation and global service capabilities exist. Entrepreneurs should focus on global innovation; despite barriers, opportunities like HK IPOs will arise for promising innovators.
A fundraising lead at a model startup described moving to secondary markets as a "double-edged sword." Recent large funding rounds for Moonshot AI and StepFun demonstrated primary market support, contrary to signals from Zhipu and MiniMax IPOs that suggested primary funding dryness. While IPOs offer better financing channels and visibility, they bring immediate commercialization pressure. Many listed companies quickly expand B2B operations for faster monetization, as secondary markets expect results within one to two years; failure to meet expectations leads to swift stock declines.
AI Organizations: Small Teams and Human Efficiency Are Key
Zou Ling, initiator of Honghub鸿鹄汇, stated that competent founders of tiny teams need three core abilities: opportunity identification based on industry experience, rapid execution using AI to create demos and iterate quickly, and self-promotion skills to attract early users and generate cash flow without traditional sales teams.
Wu Yi, assistant professor at Tsinghua's Cross Information Research Institute and head of the AReaL project, explained that极小组织形式 and full-stack innovation complement each other. Even within large companies, lean AI R&D teams are necessary because human communication bandwidth is limited, and excessive organizational hierarchy creates inefficiencies. Separating algorithm and infrastructure teams leads to a client-contractor dynamic, stifling innovation and底层感知. Integrated teams co-designing and co-evolving are more effective.
Wu Chenglin, founder and CEO of DeepWisdom, cautioned against迷信一人公司, emphasizing that organizational evaluation should focus on "human efficiency." Communication costs constitute 80% of company expenses, not coding or documentation. AI can mitigate these hidden costs. Some leading AI firms already use AI for task allocation based on employee skills and comfort zones. Critical thinkers and generalists are needed to reduce interpersonal communication issues. Experimental structures like "ROOT," with no traditional roles and full-stack responsibilities, have shown severalfold efficiency gains over conventional organizations.
Next Battles in Foundational Models: Understanding, Memory, and Affordability
Song Jiaming, chief scientist at Luma AI, advocated for "unified" multimodal models combining text, image, video, and audio understanding and generation. Such models offer stronger in-context learning and zero-shot capabilities, with higher potential than separate modality models. While拼接不同模态模型 provides short-term benefits, it delays the development of superior unified models.
Jiao Ke, former co-founder of Baichuan Intelligence and founder/CEO of来福电台, identified memory as the true barrier in the AI era. Research on memory increased in 2025 but remains nascent. Human-like memory consolidation, abstraction, and forgetting are complex. AI products compete for user memory; voice interfaces facilitate natural long-context expression. Daily Talk Users and Long-term Memory Users are more valuable metrics than DAU, as they represent deeper engagement and personalized service potential.
Yang Hongxia, former AI lead at Alibaba/ByteDance, founder/CEO of Infix.ai, and professor at HK PolyU, highlighted the落地鸿沟 for models in specialized fields, SMEs, hospitals, and government agencies due to lack of domain data in centralized models. Knowledge injection occurs only during pre-training, so local deployment requires continuous pre-training with private data. Data sharing barriers prevent global or industry-wide model coverage. She envisions every organization having its own locally deployed model, with "decentralization" enabling domain-specific model fusion, such as combining hospital specialty models into a medical foundation model.
Breakthroughs in Embodied World Models: Algorithm Innovation and Scenario Validation
Wang Xiaogang, chairman of大晓机器人 and co-founder of商汤科技, stressed that world models require downstream validation loops. Early models were seen as "data generators" without real-world testing, undermining trust.商汤 addressed this by generating scenarios in world models and validating decisions with上汽智己 vehicles, refining models through feedback. Similarly, embodied intelligence applications need real-task verification, like using quadruped robots for street patrols to iterate world model capabilities.
Huang Guan, founder/CEO of极佳视界, predicted a "ChatGPT moment for the physical world" within 2-3 years, defined as achieving 95% success rates in 90% of 100 common tasks. Current "VLA+reinforcement learning" approaches face data bottlenecks; world models could solve this. VLA handles task complexity, world models address generalization, and reinforcement learning ensures accuracy and reliability.
Zhao Hang, assistant professor at Tsinghua's Cross Information Research Institute and co-founder of星海图, considered world models a前沿方向 driven by algorithmic ingenuity rather than data. Modeling physical world dynamics would enable robots to predict action consequences, moving beyond imitation learning. However, predicting the future is more challenging than planning current actions, making it suitable for cutting-edge lab exploration.
Comments