As the year 2025 draws to a close, the focus has shifted significantly: “The year 2025 was about what AI models can do; 2026 should be about how AI can actually make money, and do so on an industrial scale.” The year-end was marked by the acquisition of intelligent agent startup Manus by Meta, while KNOWLEDGE ATLAS and MiniMax filed for listings in Hong Kong back-to-back, with KNOWLEDGE ATLAS potentially debuting on the Hong Kong Exchange as early as January 8 next year. In the domestic market, the concept of the "Six Dragons" has become a thing of the past; the foundational model startup landscape has undergone a clear divergence, and internet giants have officially begun to exert their strength. In the international market, leading players including OpenAI, Google, and Anthropic have taken turns leading in foundational models while simultaneously seeking opportunities on the application front. Top researchers, facing a growth plateau in Scaling Law, are exploring new technical paradigms. Zheng Qingsheng, a partner at Sequoia Capital China Fund, believes the AI sector in 2025 carries a sense of inflection. Although the AI industry was already quite hot in 2024, the pace wasn't exceptionally fast. However, following the emergence of DeepSeek, the overall sentiment in the industry received a massive boost. Swept up in this fervor, participants in the large model arena are undergoing a fierce "survival of the fittest" selection process: startups are either rushing to list to capture capital market opportunities, retrenching to focus deeply on vertical scenarios, or quietly exiting the foundational model race. Meanwhile, major tech companies are leveraging their advantages in computing power, data, and ecosystems to press their advantage fully, accelerating their harvest of market share through technological iteration and scenario implementation. Looking ahead, the old scale-based competition has peaked, and the exploration of new paradigms has just begun. Whoever breaks through the Scaling Law bottleneck first will gain a first-mover advantage in the next phase. Another industry insider noted that if 2025 focused on what AI models can do, then 2026 will be more concerned with how AI can actually make money, and do so industrially. The startup track has seen a clear "life-or-death divergence." The once capital-darling "Large Model Six Dragons" have completely lost their halo in 2025, with their paths clearly diverging: KNOWLEDGE ATLAS and MiniMax were the first to pass hearings with the Hong Kong Exchange, initiating their IPO processes; Moonshot AI and StepFun continue to update their model technologies and application products, striving to meet listing revenue requirements; Baichuan AI and 01.AI have pivoted to vertical applications, focusing on healthcare and enterprise-level (To B) commercialization scenarios, respectively. Behind this transformation lies the rapid evolution of the entire AI sector in 2025. According to QuestMobile's Q3 data, the landscape of leading native apps is relatively stable: Doubao surpassed DeepSeek, with monthly active users (MAU) of 172 million and 145 million, respectively; Tencent Yuanbao, Jimeng AI, and Kimi had MAUs of 32.86 million, 10.12 million, and 9.67 million, respectively. Kimi's data situation once sparked widespread discussion. Regarding Kimi's current development status, a source close to Moonshot AI indicated that Kimi has a high proportion of professional users, leading to a greater share of web-based users than app users. After launching the K2 Thinking model in November, web traffic surged by 48.6% to over 43 million visits. Kimi's current business model primarily relies on C-end subscription fees for Agent tasks and B-end API calls. The team is currently investing in developing the K2 series models with visual capabilities. Separately, an investor revealed that following KNOWLEDGE ATLAS and MiniMax, Kimi is also vigorously "pushing for revenue," expecting to attempt a listing potentially under Hong Kong's Chapter 18C framework. Chapter 18C refers to the "Specialist Technology Companies" listing regime under the Main Board Listing Rules, effective March 31, 2023, providing a listing channel for hard tech companies that haven't met traditional profit/revenue thresholds but possess high R&D intensity and barriers. KNOWLEDGE ATLAS and MiniMax both belong to this 18C cohort. According to the latest QuestMobile statistics (covering Dec 08-14), the rankings of active users for AI-native apps have shifted. Doubao, DeepSeek, and Yuanbao hold the top three spots, with weekly active users (WAU) of 155 million, 81.56 million, and 20.84 million, respectively; Ant Group's A-Fu and Qwen form the second tier, with WAUs of 10.25 million and 8.72 million, respectively. A senior executive from a foundational model provider believes that the core of next year's competition will still be the models themselves. Whether it's infrastructure construction or data processing, the ultimate goal is to define the experience of the next-generation model. "The various AI tools available now are just the beginning; after interacting with AI for twenty or thirty rounds, their limitations become apparent." A key aspect of the next-generation model is deeper multimodal integration; even Google's Gemini is just starting out in this area, which will be the core direction for the entire industry next year. However, it's important to note that as the core support for capability iteration, the Scaling Law for foundational models has noticeably slowed. Sebastian Borgeaud, one of the leads for Gemini 3's pre-training, stated that scale remains important, but the weight of architectural innovation and data innovation has significantly increased, even becoming more critical. Multiple industry insiders indicated that Silicon Valley is exploring new technical paradigms, but regardless of the approach, training foundational models still requires substantial capital investment, directly leading to fewer players remaining at the foundational model table, while the stakes required to play keep rising. KNOWLEDGE ATLAS's global offering document shows that 70% of the net proceeds from the listing (approximately HK$2.9 billion) will be used for R&D investment in general-purpose AI large models, further consolidating its competitiveness in general foundational models. MiniMax's prospectus indicates the company's monthly cash burn is close to 200 million yuan. Major tech companies have officially entered the competition with force. By the end of 2025, aside from the major news of Manus's acquisition by Meta, the industry's heat was almost entirely generated by the major tech firms. Specific developments include Doubao's daily active users (DAU) surpassing 100 million, with its large model handling over 50 trillion tokens daily; Tencent announced the establishment of AI Infra/AI Data / Data Computing Platform departments, appointing Yao Shunyu as Chief AI Scientist; Alibaba's AI-to-C initiatives are in full swing, with Qwen's MAU exceeding 40 million. Data disclosed by DataEye Research Institute's ADX industry edition shows that in November 2025, among top native AI apps in the mainland market, Tencent Yuanbao's advertising spending remained far ahead, while the new product Qwen showed significant growth, surpassing Yuanbao in daily ad material volume by the end of the month. A source close to Alibaba stated that Qwen's current prominence in the industry is primarily driven by support from the Alibaba Group side, including comprehensive backing in funding, resources, and talent. Business executives communicate frequently, multiple times a week, to discuss development progress. It is understood that a large number of product and technology staff from Qwen's C-end business unit are collectively "stationed" in Hangzhou to accelerate product iteration and internal resource coordination. Amap is the first Alibaba ecosystem scenario announced to be integrated with Qwen. It is further understood that Alibaba plans to integrate various life scenarios—such as maps, food delivery, ticketing, office work, and shopping—into the Qwen app, enhancing its capability to handle tasks. Regarding the specific motivations behind the Group's push on Qwen, industry observers speculate it may reflect a strategic desire to catch up—Alibaba is already a leader in the B-end sector and sees room for improvement in C-end consumer mindshare. This all-hands-on-deck effort might be an attempt to see if concentrating technical沉淀 and resources onto the Qwen app will create a synergistic effect. ByteDance, compared to Alibaba, started earlier in laying out MaaS (Model as a Service) from a cloud service perspective, pushing from both the cloud and foundational model ends. Tencent is making a formal push through organizational restructuring. Zhu Xiaohu, Managing Partner at GSR Ventures, commented that Tencent has historically avoided being an early-stage, cash-burning pioneer, preferring to wait for a赛道 to become clear before exerting force to catch up, gradually moving from the back to the forefront. Tencent possesses strong advantages in traffic, scenarios, and data, allowing it to avoid trial-and-error risks at a lower cost, making it suited for a marathon-like long race, but it remains relatively restrained in early-stage innovation and defining scenarios first. Regarding this wave of collective effort by major firms, KNOWLEDGE ATLAS CEO Zhang Peng does not believe it can be simply characterized as a "harvesting" phase. He stated that, on one hand, the industry is still in its early stages, and the entry of large companies can help cultivate the market with their resources, encouraging more people to use AI. The current state of the AI market means participants are competing mainly on strategic vision, execution capability, and timing. Reflecting on the year, Zhang Peng noted that the large model industry, both internationally and domestically, has shown a clear trend of major companies exerting force. Of course, there is differentiation even among the giants, including success stories like Google, which was backed into a corner but fought back, and others like Meta, which started strong but lost momentum. However, compared to the relatively clear格局 of the US AI market, the competitive landscape in China's AI sector is more complex. Reasons include the sheer size of the domestic market; furthermore, existing domestic players each have their own "survival strategies," leading to asymmetric competition among them. The landscape among startups is relatively clearer: some have explicitly exited the赛道 competition, some have found their niche during development, and others persist through difficulties in advancing towards AGI. 2026 is unlikely to be the final chapter. "Truly, it feels like a year on Earth is just a day in AI." Looking back at the end of the year, recalling the impact of the emergence of DeepSeek and Manus at the beginning of the year, Jia Anya, head of SenseTime's Little Raccoon, expressed his amazement. "I still clearly remember spending several all-nighters reading DeepSeek's paper, going days without sleep, and it was the same when Manus was released, staying up all night because Manus was doing many things so similar to us." But even such significant changes at the start of the year seem somewhat mundane in retrospect, Jia Anya mentioned, "There have been so many changes in AI this year; the threshold for surprise among practitioners has been constantly raised." The most palpable feeling in the industry this year is the sheer speed of AI development. Leading companies release new models on average every one or two months, continuously pushing the upper limits of benchmark scores, making competition fiercer than ever. The emergence of DeepSeek at the beginning of the year stirred the global AI landscape, while Google's Gemini 3 disrupted the OpenAI ecosystem at year's end; the AI industry never lacks for new narratives. Reflecting on the year, Jia Anya believes that people are no longer surprised by single-point technological breakthroughs in AI, but rather look at the real value brought by technological progress. This year has been a process of moving from the virtual to the real. Chen Kai, a Young Leading Scientist at the Shanghai AI Laboratory, believes that AI's coding capabilities and intelligent agents were the key breakthrough directions for various models this year. Particularly, the commercial value associated with coding capabilities is high, as it can directly generate revenue upon implementation, making it a key area for strategic focus. From another perspective, Jia Anya considers coding a very fundamental capability that has brought significant changes and important progress to both the industrial upstream/downstream and the reasoning capabilities of the models themselves. Looking back, the pace of AI iteration remains rapid. On the other hand, discussions about technological plateaus have persisted. Since late last year, some have argued that "Scaling Law has peaked." Just last month, OpenAI's former Chief Scientist, Ilya Sutskever, again voiced his opinion that the mainstream development path for AI has encountered a bottleneck, and the era of scaling is over. Will large models face a technological ceiling? Will the pace of development slow next year? Chen Kai analyzed that the industry格局 is not yet fully consolidated and remains in a phase of competitive iteration. Even if a specific technology or direction hits a bottleneck, new directions may emerge to solve existing problems and bring new growth. "For example, model training also encountered bottlenecks at certain points in the past, with slowing development pace and limited improvements, but then new technologies like OpenAI's o1 emerged, helping models advance further." Chen Kai believes that a bottleneck in one technology does not mean the entire large model industry has hit its ceiling. In Chen Kai's view, the industry will certainly encounter瓶颈 periods or plateaus in the future, but at least next year will likely maintain a high-speed development trend. "The AI industry has a strong talent attraction effect, gathering a group of exceptionally talented practitioners. Whether in China or the US, competition among large model companies is intense. In such an atmosphere, the industry will naturally maintain rapid development." He predicts that companies may increase investment in new architectures and new learning paradigms next year. Industry players are gradually realizing that the continuous progress of large models离不开 more research-driven innovation. While gradual optimization and iteration on existing technical paths can achieve decent results, achieving further breakthroughs requires more fundamental research breakthroughs. Zhang Peng believes that discussions about new technical routes stem from the industry's continued expectation for fundamental innovation and breakthroughs, holding that there are still many possibilities for AI's future development. In his view, 2026 will also most likely not be the year when everything is settled. Looking towards 2026, Jia Anya believes that the commercialization of AI might become a major focus. "2025 focused on what AI models can do; 2026 should be about how AI can actually make money, and do so on an industrial scale." It has been over three years since ChatGPT's emergence; next year will be the fourth year of this large model wave. Model providers will increasingly consider how to commercialize on a larger scale, while enterprise customers on the demand side will evaluate the value and return on investment from spending on AI. In the large model industry of 2025, startups sought survival paths amidst divergence, while major tech firms reshaped competitive rules through resource dominance. As Zhang Peng stated, 2026 may still not be the "final chapter"; this long-term contest involving technology, capital, and ecosystems is still in its early stages.