Four of China's biggest spenders are sitting at the same poker table.
Today, let's analyze the comparison of the absolute new forces among China's young AI teams:
Zhipu is the most expensive, with a market value exceeding 700 billion yuan. Its revenue is 724 million yuan, a year-on-year increase of 131.9%. It has the largest revenue scale among the four. However, its annual loss is as high as 4.718 billion yuan, a year-on-year expansion of 59.5%. The adjusted net loss is 3.182 billion yuan, and the gross profit margin has dropped sharply from 56.3% to 41.0%.
MiniMax was the first to achieve self-sufficiency. Its revenue is 79.038 million US dollars, a year-on-year increase of 158.9%. Overseas revenue accounts for more than 70%. Its gross profit margin has increased significantly from 12.2% to 25.4%, and it is the only one among the four with a continuous improvement in gross profit margin.
Dark Side of the Moon has the fastest growth rate. Its annual revenue in 2025 was average, but it experienced explosive growth at the beginning of 2026. Less than a month after the release of the K2.5 model, its ARR (Annual Recurring Revenue, the same below) exceeded 100 million US dollars, and it increased to 200 million US dollars in April 2026.
Recently, Dark Side of the Moon has initiated consultations for a Hong Kong IPO, and its valuation has soared from 4.3 billion US dollars at the end of 2025 to over 20 billion US dollars.
And DeepSeek has the lowest cost structure and the highest ideals. It has had almost no revenue in more than three years since its establishment and has been completely supported by the annual profits of tens of billions of yuan from its parent company, Magic Square Quant. However, it launched its first round of external financing in April 2026, with a post-investment valuation of about 45 billion US dollars.
Today, we will take you to analyze the completely different business logics and decision-making cards of these four companies: DeepSeek, MiniMax, Zhipu, and Dark Side of the Moon, starting from these financial report figures.
The origin of everything:
Four backgrounds and four obsessions of the founders
1. Liang Wenfeng of DeepSeek: An AGI believer from a quantitative laboratory
Liang Wenfeng, born in 1985, holds a master's degree in Information and Communication Engineering from Zhejiang University. He co-founded the quantitative investment institution Magic Square Quant during his postgraduate studies in 2008. The management scale of this quantitative private equity fund once exceeded 100 billion yuan in 2021 and still exceeded 70 billion yuan in 2025, with an annual yield of 56.55%. It is one of the leading quantitative institutions in China.
Quantitative investment and AGI may seem unrelated, but in fact, they share the same core: an extreme belief in data, algorithms, and computing power.
In July 2023, Liang Wenfeng incubated DeepSeek with Magic Square Quant as the main body, with an initial positioning of "basic research on large language models for AGI".
He clearly stated that developing large models is not for short-term commercial realization but "to solve the most important problems" and attract top technical talents.
DeepSeek became well-known in January 2025. At that time, DeepSeek released the large inference model R1 and the open-source model V3. While approaching the performance of OpenAI's o1 and GPT - 4o, its extremely low training cost shocked the entire industry.
One of DeepSeek's core moats is its cost structure - it has reduced the price to one-thirtieth of others, relying on architectural innovation and domestic computing power.
2. Yan Junjie of MiniMax: An industry veteran betting on "economies of scale"
Yan Junjie, born in 1989, represents another path with his work experience: giving priority to industry experience.
He studied for his bachelor's, master's, and doctoral degrees at the Institute of Automation, Chinese Academy of Sciences. He joined SenseTime in 2014 and was promoted from an early - stage employee to vice - president, responsible for core R & D work such as computer vision. He left SenseTime in 2019 and founded MiniMax in December 2021.
Yan Junjie's judgment is that the competition of large models is ultimately a competition of "economies of scale". The more users, the more data, the better the model, the lower the price, and then more users. This is an idea closer to the Internet business logic.
MiniMax has placed commercialization at the core from the beginning: it has three revenue streams, namely API revenue, subscription fees, and enterprise - level solutions, and gives priority to the overseas market because overseas users have a stronger willingness to pay and a higher lifetime value.
It was the first among the four to cross the break - even line. With its C - end overseas product Talkie, it achieved break - even in 2024. It is advancing in five full - modality areas, covering text, images, videos, audio, and music.
3. Zhang Peng of Zhipu: An academic with technical awareness and the patience to face the capital market
In 2019, the Knowledge Engineering Laboratory (KEG) of Tsinghua University, in cooperation with multiple in - school laboratories, officially incubated Zhipu AI, with Zhang Peng, then an associate researcher at KEG, serving as the CEO. This is the company with the most academic background among the "Four Little Dragons of Chinese AI".
Different from the pure commercialization route, Zhipu has regarded "cognitive intelligence" and "safety and controllability" as its core labels from the beginning.
Its GLM series of large models have been continuously iterated in indicators such as Chinese semantic understanding, logical reasoning, and multi - modality understanding. At the same time, it actively participates in the standard formulation of domestic large models and government cooperation projects.
Localized deployment is Zhipu's core business model. Customers (mainly governments and large enterprises) deploy the model on their own servers, and Zhipu charges authorization fees and technical service fees. The advantage of this model is a high unit price and stable cash flow; the risk is that the growth ceiling is relatively limited, and it faces the pressure of technological substitution from open - source models.
Zhipu has the highest market value, once exceeding 700 billion yuan. It has raised prices three times by 83%, but the call volume has increased by 400%.
According to Sina Finance, 9 out of the top ten Internet companies call its GLM model every day.
4. Yang Zhilin of Dark Side of the Moon: The youngest CTO and the fastest financing machine co - exist.
Yang Zhilin, born in 1992, is the youngest among the founders of the "Four Young Stars" and also the one with the fastest pace and the most controversy.
He graduated from Tsinghua University with a bachelor's degree and then went to Carnegie Mellon University for a doctorate, studying under the head of Apple's AI research team. He founded Dark Side of the Moon in June 2023 and released the AI assistant Kimi, which became one of the earliest large - model applications in China with a lossless context window of 2 million words.
Yang Zhilin has a distinct style: he extremely pursues technical indicators and growth data and is willing to exchange high valuations and high financing density for a time window. "Long - termism" was a term he repeatedly used in the early days, but the preparation for the Hong Kong IPO in May 2026 marks that this narrative is being rewritten by reality.
Four sets of answers in business, reflecting four different paths
1. Liang Wenfeng: The essence of AI is the infinite approximation of the basic model
In Liang Wenfeng's view, all commercialization in the AI industry is based on the capabilities of the model. If the foundation is gone, how can the superstructure exist?
Therefore, DeepSeek's only strategy is to make the model perfect and must make it open - source. The extremely low API pricing and free access on the web are all in service of the goal of "maximizing the technical influence".
However, this purity is also facing real - world pressure. In April 2026, DeepSeek officially launched its first round of external financing, which is due to the real - world pressure of the talent war.
When giants like ByteDance and Xiaomi offer attractive compensation packages, DeepSeek, which has no equity realization channel, is at a significant disadvantage. Even someone as pure as Liang Wenfeng has to find a new balance between ideals and reality.
2. Yan Junjie: The essence of AI is the optimal solution of "cost - experience"
Yan Junjie's first - principle is game theory. He believes that the essence of the competition of large models is an optimization game of "cost - experience".
Users will not pay for the "strongest model" but for the "model with the best experience within an acceptable cost". Therefore, MiniMax's strategic focus is not to pursue absolute leadership in model parameters but to pursue the optimal balance among model efficiency, inference cost, and user experience.
This thinking runs through all of MiniMax's decisions: firmly going overseas (to avoid the domestic price war), self - developing multi - modality (to reduce external dependence), and going all - in on the C - end (user growth drives model iteration).
In 2025, MiniMax's gross profit margin increased from 12.2% to 25.4%, and it is the only one among the four with a continuous improvement in gross profit margin. The R & D expenditure only increased by 33.8%, far lower than the 158.9% revenue growth rate, and the economies of scale began to be realized - the R & D investment required for each additional 1 yuan of revenue decreased from 6.19 yuan in 2024 to 3.20 yuan in 2025.
At the same time, the marketing expenses decreased by about 40% year - on - year, which means that the growth comes more from user word - of - mouth and natural traffic rather than paid promotion.
Although the annual book loss is as high as 1.87 billion US dollars, about 1.6 billion US dollars is the non - cash revaluation of the fair value of preferred stocks. After adjustment, the actual operating loss is only 250 million US dollars.
3. Zhang Peng: The essence of AI is the reliability of productivity tools
Zhipu's academic background determines that its understanding of AI is closer to "infrastructure": to become a reliable productivity tool for enterprises.
Therefore, Zhipu's product design is known for its stability. It provides full - stack services: from the model layer to the platform layer and then to the application layer. This "one - stop" strategy allows customers to solve their needs in one go, but it also faces higher R & D investment and heavier delivery pressure.
In 2025, Zhipu's R & D expenditure reached