(Beijing) –– When Time magazine included ByteDance, Alibaba, and Zhipu AI in its list of the “10 Most Influential AI Companies of 2026,” a clear signal was sent: China is no longer a passive follower in the AI race, but one of the rule-makers. Meanwhile, China’s large model landscape has hardened from a “war of a hundred models” into what the industry calls the “Big Five”—Moonshot AI, StepFun, DeepSeek, Zhipu, and MiniMax—which within a single week took turns shattering valuation records. Their combined valuation surged more than tenfold in just a matter of months. This is not a bubble narrative, but a silent revolution about technological sovereignty.
ByteDance may be the only AI company in the world that simultaneously possesses billions of consumer users and a computing budget in the tens of billions of dollars. Time defines it as an “AI-first” technology giant, not merely a short-video company. Its AI assistant, Doubao, broke through 100 million daily active users during the 2026 Spring Festival, with weekly active users reaching 155 million, making China one of the first markets in the world to achieve mass adoption of AI assistants. What sustains this scale is ByteDance’s capital expenditure budget exceeding $20 billion per year, as well as a reported Nvidia chip purchase plan valued at up to $14 billion. Its strategic logic is cold and clear: use cash flow from consumer products to feed infrastructure, then let that infrastructure feed the next generation of products.
If ByteDance excels in scale, Zhipu AI’s core competency is technological self-reliance. Time specifically praised Zhipu’s ability to train GLM-5 on a domestic computing platform—a model with 744 billion parameters trained entirely on Huawei Ascend processors, and in some benchmarks outperforming Google Gemini 3 Pro. Zhipu’s symbolic significance far exceeds its commercial numbers: it proves that Chinese AI developers can compete at the frontier without any Nvidia ecosystem at all. In January 2026, Zhipu listed on the Hong Kong Stock Exchange as “the world’s first large-model stock,” with market capitalization soaring from HK$5.7 billion to more than HK$1.27 trillion in four months. Founder Tang Jie, a Tsinghua professor, launched the “Summit Plan,” which explicitly states: the next two years will not pursue short-term monetization, but focus on basic AGI research. This is a bet on the upper limit of technology.
Alibaba’s position in AI is neither consumer products nor pure technological research, but ecosystem infrastructure. The Qwen (Tongyi Qianwen) model series has surpassed 1 billion cumulative downloads, spawned more than 200,000 derivative models, and been adopted by international companies such as Airbnb and Pinterest. Time assessed that Alibaba is “reshaping more fields beyond retail,” attracting global developers through open-source models, then monetizing through cloud services. Alibaba Cloud Intelligence CEO Wu Yongming set a target: combined external revenue from cloud computing and AI exceeding $100 billion within five years. This is the classic platform strategy of “open-source attracts traffic, cloud harvests value,” and Alibaba is replicating it into the AI era.
Moonshot AI’s K3 model has 2.8 trillion parameters, the largest open-source model ever released to date. This strategy allowed Kimi to rapidly build reputation in the developer community, but also exacted a heavy price: three days after K3’s launch, due to insufficient server resources, the company was forced to suspend new user registrations. Kimi completed approximately $2 billion in funding in May 2026, with post-investment valuation breaking through $20 billion, led by Meituan Longzhu, with participation from China Mobile and CPE. More noteworthy is its commercial progress: as of June 2026, Kimi’s Annual Recurring Revenue (ARR) broke through $300 million, with API revenue accounting for more than 70%, and overseas paying users growing fourfold. Founder Yang Zhilin holds 51.83% of shares, with personal wealth exceeding 110 billion yuan. Kimi’s dilemma and advantage are equally clear: it has the strongest open-source model, but needs continuous funding to feed its computing hunger.
DeepSeek shook the world with its R1 model in January 2025, proving the ability to compete with American giants at extremely low cost. The Flash version of its V4 series became the most-used model on the OpenRouter platform globally in September 2026. However, the era of “free and cheap” is ending. After completing $7.4 billion in funding in June 2026, DeepSeek significantly raised peak-hour rates. More controversial is its valuation logic: DeepSeek’s ARR is only $500 million, yet its valuation exceeds $50 billion, with a valuation multiple of 163 times ARR—by comparison, OpenAI’s valuation multiple is 34 times, Anthropic’s 21 times. Founder Liang Wenfeng personally injected up to 20 billion yuan in funding, controlling 84.29% of the company’s shares through direct and indirect ownership. The capital market is not betting on DeepSeek’s current revenue, but on its potential as a technology “definer.”
Among the Big Five, MiniMax has the lowest profile, but its commercial data is the most solid. Full-year 2025 revenue reached $79.03 million, growing 158.9% year-on-year, with more than 70% of revenue coming from international markets. The company was founded in Shanghai in early 2022 by Yan Junjie, former vice president of SenseTime, and listed on the Hong Kong Stock Exchange in January 2026, with its share price surging approximately 109% on the first day. MiniMax’s strategy is full multimodal coverage: text, voice, image, video, music generation—none left behind. However, this “big and complete” strategy also faces doubts: its shares closed down 15% on the day of the M3 model launch, and independent evaluation agencies ranked it in the middle of open-source models. MiniMax’s challenge is: among the Big Five, it has neither ByteDance’s scale, nor Zhipu’s technological narrative, nor DeepSeek’s disruptive label.
The formation of the “Big Five” is not the result of natural evolution, but a product of the resonance of three forces: capital, policy, and geopolitics. Chapter 18C implemented by the Hong Kong Stock Exchange in 2023 opened a listing channel for specialized technology companies, and Zhipu and MiniMax used it to become the first two large-model companies to list globally. The National Integrated Circuit Industry Investment Fund (Big Fund) is reported to invest in DeepSeek, providing state capital backing for a company that has not yet achieved substantial commercialization. But concerns are also clear: the valuations of most of the Big Five are severely disconnected from actual revenue, with DeepSeek’s valuation multiple nearly five times OpenAI’s, while its revenue is only one-hundredth of the latter’s.
The deeper structural problem is the scissors gap between computing costs and commercial returns. Zhipu’s 2025 revenue was 724 million yuan, growing 132% year-on-year, but it is still losing money. Although Moonshot’s ARR broke through $300 million, K3’s computing consumption makes it difficult to sustain operations with its own cash flow. Although MiniMax’s revenue growth is strong, the training and inference costs of multimodal models are far higher than pure text models. Each of the Big Five is transfusing funding to exchange for technological leadership, and the patience of the capital market is being tested.
CLSA, in a research report published in September 2026, pointed out that China’s AI cycle is shifting from “technological breakthrough” to the stage of “large-scale commercial returns.” Behind this assessment is the cruel reality pressure: IDC predicts that by the end of 2026, half of AI pilot projects will struggle to achieve expected investment returns, and the logic of industry competition has shifted from “parameter competition” to “full-scale implementation and sustainable commercial closed loop.” The Big Five must answer an unavoidable question: can technological leadership be converted into sustainable profits?
Some signals are positive. Zhipu’s MaaS API platform reached ARR of 1.7 billion yuan, growing 60 times year-on-year, with gross margin rising from nearly zero to 18.9%. Kimi’s paying users in finance, law, and scientific research account for more than 60%, indicating that B-end scenarios are generating real willingness to pay. Alibaba Cloud Intelligence’s AI-related external revenue target is set at $100 billion in five years, demonstrating the scale potential of the platform business model. But overall, the total ARR of Chinese AI developers is approximately $10.5 billion, only 10% of the combined revenue of OpenAI and Anthropic, which reaches $105 billion.
Four of the Big Five (Zhipu, Alibaba, Kimi, DeepSeek) make open-source their core strategy—this is not coincidence, but a rational choice under geopolitical technology competition. Amid continuously tightening US chip export controls, open-source models have unique strategic value: once model weights are published, they cannot be “taken down” or “cut off.” The timing of Zhipu’s GLM-5.2 launch was chosen the day after US export restrictions temporarily cut off Claude access—this timing is no accident. Alibaba Qwen’s 1 billion downloads and 200,000 derivative models have formed a global developer ecosystem that cannot be controlled by a single entity. Open-source is not just a technology strategy, but also a decentralized survival strategy.
Zhipu’s stock price trajectory is the best sample for understanding the “Big Five premium.” In January 2026, it listed at an offering price of HK$116.2, reached a peak of HK$1,229 in May, with market capitalization briefly breaking through HK$500 billion—while the company’s full-year 2025 revenue was only 724 million yuan. The valuation premium granted by the capital market “is already very difficult to measure with traditional corporate valuation logic.” This premium is built on three narratives: the potential to become AI infrastructure, the possibility of becoming a technology definer, and the policy support of domestic substitution. But each layer of these three narratives requires time for verification, and stock prices will not wait forever for verification results.
The Big Five landscape also maps the geographic concentration of China’s AI industry. Zhipu is rooted in the Tsinghua clan in Beijing, ByteDance with Volcano Engine is in Beijing, and Moonshot AI is also in Beijing. DeepSeek comes from Hangzhou, and Alibaba Qwen’s headquarters is also in Hangzhou. MiniMax chose Shanghai. Beijing, with 21 companies in the Forbes China AI 50, accounts for 24.4% of the national total, has more than 2,500 core AI companies and 225 registered large models, about 30% of the national total. This geographic concentration brings convenience in talent mobility, but also means highly concentrated risk—if Beijing’s policy or capital environment changes, the entire Big Five landscape may be forced to restructure.
Using “China’s Big Five AI” to describe these five companies easily creates a misleading sense of symmetry. The reality is: OpenAI and Anthropic’s combined ARR exceeds $100 billion, while the Big Five China’s combined ARR is approximately $10.5 billion. At the level of model capability, Zhipu’s GLM-5 approaches Claude Opus 4.5 and GPT-5.2 in some benchmarks, but “approaching” does not equal “surpassing.” At the level of computing infrastructure, American companies have more abundant Nvidia GPU supply, while Chinese companies must achieve equivalent-scale training efficiency on domestic chips. The true significance of the Big Five landscape is not that they have “conquered the world,” but that they are defining a path of AI development that does not depend on the Western technology stack.
The “Big Five landscape” may be more fragile than most people expect. MiniMax’s shares fell 15% after the M3 model launch, and independent evaluations ranked it in the middle of open-source models. Moonshot’s K3 is too large for users to even deploy themselves. Although StepFun is active on the funding side, its model products’ presence in public evaluations is far lower than the other four. History provides a warning: the “Four Little AI Dragons” (SenseTime, Megvii, CloudWalk, Yitu) were once regarded as the absolute rulers of Chinese computer vision, but in the end only two successfully listed, and to this day have not achieved profitability. Each of the Big Five must answer the same question: when the funding window closes, can your technological leadership be converted into cash flow?
The deeper significance of the Big Five landscape transcends commercial competition. Zhipu trained GLM-5 with Huawei chips, proving that Chinese AI can operate entirely without the Nvidia ecosystem. Alibaba Qwen’s global open-source ecosystem is building an AI supply chain that is not subject to the jurisdiction of any single government. ByteDance’s Doubao has 155 million weekly active users, making China one of the markets with the highest AI assistant penetration in the world. These facts point to the same conclusion: the real bet of China’s Big Five AI is not market share, but technological sovereignty. In a world where chip export controls are increasingly tightening and technological decoupling is accelerating, the ability to independently train frontier models, independently deploy large-scale AI services, and independently build developer ecosystems is itself a strategic asset.
The patience of the capital market will eventually run out. IDC’s prediction—that half of AI pilot projects cannot achieve expected returns—is not alarmism, but a signal of industrial reshuffling that is about to arrive. Each of the Big Five is transfusing funding to exchange for time, but time is not unlimited. Zhipu’s “Summit Plan” promises not to pursue short-term monetization within two years—this requires sustained trust from the capital market. Although Moonshot’s ARR growth is astonishing, K3’s inference costs may devour most of its revenue. DeepSeek’s 163-times valuation multiple requires exponential revenue growth to become rational. Ultimately, the stability of the Big Five landscape does not depend on their total valuation, but on how many of them can achieve self-sustaining profitability before 2028.
If the Big Five landscape eventually stabilizes, what it represents is not a replication of the American model, but a unique Chinese AI development paradigm. The core characteristics of this paradigm are: open-source prioritized over closed-source, domestic computing prioritized over import dependence, B-end scenarios prioritized over C-end subscriptions, policy capital prioritized over pure market financing. This forms a sharp contrast with the path of American AI companies: OpenAI and Anthropic make closed-source models and API subscriptions their core, rely on Nvidia GPUs and venture capital, and mainly serve the C-end and developer markets. China’s Big Five are proving that under three constraints—limited computing, capital controls, and geopolitical pressure—it is still possible to build a globally competitive AI technology stack. Whether this paradigm is sustainable will determine whether the global AI landscape of the next decade moves toward “bipolar” or “multipolar.”
