With the rapid development of AI technology, the global large model competition has entered the final stage. Internationally, five companies such as Microsoft/OpenAI, Amazon/Anthropic, Google, Meta, and xAI are the main contenders. In China, with the rise of DeepSeek, the original “Six Little Tigers” pattern has been broken, forming a “new top five” pattern represented by ByteDance, Alibaba, Step Stars, Zhipu AI and DeepSeek. This article will explore the advantages of these companies in terms of technology, resources, and talent, as well as their strategic layout and future prospects on the road to AGI.
After 2 years of development, the competitive pattern of large models is rapidly converging.
Abroad, the finals of the large model competition were gathered to five companies: OpenAI, Anthropic, Google, Meta and xAI. In China, DeepSeek’s success has greatly impacted the existing competitive landscape, making the industry and the market re-realize that the level of intelligence is more essential than traffic, and it is the real confidence to enter the international arena. At the same time, the adjustment of Baichuan, Zero One and other companies has naturally disintegrated the narrative of “AI Six Tigers”. The powerful players at the table have quietly formed a new top five pattern, namely Byte, Alibaba, Step Star, Zhipu AI and DeepSeek, and the common characteristics of these five companies are their deep accumulation and continuous innovation ability in basic models.
Behind all this evolution points to the fact that intelligent lifting is still the only thread on the AGI roadmap.
Since the beginning of this year, DeepSeek has verified one thing: model is application. In the face of steep intelligent technology, it is difficult for phased products to form absolute barriers. No product or business model innovation is as good as smart innovation.
Today, Crow Jun will take you to take stock of the new pattern of China’s large model industry.
01 The “100 Model War” is over, and the “Top Five” is beginning to appear
Let’s first look at the situation in the West. At the end of last year, Sequoia Capital put forward three major predictions for the development trend of artificial intelligence in the United States in 2025 in the report “AI in 2025: Building Blocks Firmly in Place”, and judged that the large model competition has entered the final stage, and there are only 5 finalists, namely Microsoft/OpenAI, Amazon/Anthropic, Google, Meta and xAI.
To achieve these three challenges, product managers will only continue to appreciate
Good product managers are very scarce, and product managers who understand users, business, and data are still in demand when they go out of the Internet. On the contrary, if you only do simple communication, inefficient execution, and shallow thinking, I am afraid that you will not be able to go through the torrent of the next 3-5 years.
View details >
In Sequoia’s view, these five companies have their own advantages:
- Google has strong vertical integration capabilities, because Google is the only manufacturer with independent research and development of first-class chips, and has its own data center and a very strong research team;
- OpenAI’s biggest barrier is branding. Reflected in revenue, OpenAI’s revenue has reportedly exceeded $3.6 billion;
- Anthropic’s strength lies in its talent, and it is one of the companies that has recruited the largest number of artificial intelligence scientists in the world;
- xAI has built a Colossus cluster with 100,000 GPUs at a record speed, and has now become a leader in data center expansion.
- Meta, on the other hand, focuses on open source and is the only tech giant in the industry to adopt an open source strategy. Once cutting-edge technology developments begin to slow down, Meta will be in a good position to use its open-source model to quickly spread these features.
This trend is not accidental, and the domestic market is showing a similar trend. With the emergence of DeepSeek, it has not only completely changed the competitive landscape of Chinese and American large models, but also broken the original industrial map of domestic large models in one fell swoop.
Among the domestic large models “Six Little Tigers”, many have begun to adjust their strategies and begin to transform. For example, Baichuan and Zero One Things abandoned pre-trained large models and bet on medical AI and comprehensive industry implementation respectively; MiniMax has shrunk its To B business and focused on overseas markets for C-end video generation and other scenarios.
But on the other end, we still see more resources betting on the large model industry, technology giants have increased investment, and some leading large model companies have also received investment from investors.
According to the Zheshang Securities report, ByteDance’s capital expenditure is expected to reach 160 billion yuan in 2025, a full double of last year’s 80 billion. In February this year, Alibaba announced that in the next three years, Alibaba will invest more than 380 billion yuan in building cloud and AI hardware infrastructure, more than the past decade combined.
In terms of startups, we observed that the leading companies still have the ability to attract money in the past year, but there has been a clear difference in the development path, which is not only reflected in the differentiation of the aforementioned base model (Zhiyuan, Step Star) and vertical applications (such as the dark side of the moon, Minimax, Baichuan, Zero One Thousand Things), but also reflected in the bets of the capital market.
Crow Jun believes that the most noteworthy thing in the start-up camp is the “two strong north and south” that insist on basic model research and development and are blessed by state-owned assets – Shanghai team Jie Xingchen and Beijing team Zhipu, which have completed a round of large financing at the end of last year and the beginning of this year.
Benchmarking the “top five” in the United States, the powerful players who really remain at the table in the domestic large model market have evolved from the “six small small powers of large models” to the “new top five” pattern with basic models as the core capabilities, namely Byte, Alibaba, Step Star, Zhipu AI and DeepSeek.
02 The two main lines of the “Warring States” era of large models
When the competition of large models enters the finals, the technical decisive point is also undergoing some changes. From the current point of view, the technical decisive point of the next stage of base model competition will focus on the upper limit of intelligence and multi-modal capabilities.
Let’s talk about the first point, on the roadmap to AGI, intelligent improvement is still the most important main line.
The reason is simple, the prosperity of large model applications cannot be separated from the capabilities of basic models. To use the well-known “barrel principle” as an analogy, assuming that AGI is a barrel, the level of the basic model is like the comprehensive integrity of the barrel, which determines the upper limit of the capacity of this barrel. If the bottom of the barrel is very thin, or if there are only one or two long plates on all sides, then it seems that it does not constitute a barrel at all. Therefore, only by continuously breaking through the intelligence level of the basic model can we improve the overall capacity accumulation source of the barrel and provide a stable ecological environment for the application of large models in the barrel.
Especially in the process of Agent implementation, the gap in model capabilities will be further magnified. Because the agent needs to do long-distance multi-step reasoning, if the model is 5-10% better, the leading effect will be magnified many times.
On the other hand, the model itself is also the biggest application. For model companies, the capabilities of basic models are directly related to the competitiveness of the user side. This is why ChatGPT is still the most trafficked application in the AI field.
At the beginning of this year, DeepSeek suddenly emerged, and in a short period of time, it smoothed out the traffic advantage that Byte Bean spent a lot of money to smash without any promotion investment, relying on stronger model capabilities. In March, DeepSeek’s chatbot has surpassed 16.5 million monthly visits.
In addition to model intelligence, multimodal technology is also another decisive point in this large model final.
Zhang Hongjiang, founding chairman of Beijing Zhiyuan Institute of Artificial Intelligence and foreign academician of the National Academy of Engineering, once said that the multimodal large model is the ultimate model of AGI.
In Academician Zhang’s view, a unified multimodal large model can solve all problems of understanding. To put it simply, 80% of people’s information is obtained through vision, which includes multiple high-dimensional information such as objects, light and dark colors, and spatial relationships, while speech is single-dimensional information, so the amount of visual information is much more than other modalities. The value of multimodal data is that it can raise the upper limit of intelligence.
Especially in the process of AI implementation, multimodal technology plays a crucial role. For example, with multimodal large models, robots that can truly be universal, self-planned, and make autonomous decisions can be developed.
In fact, the model of traditional robots is completely designed by humans, decomposing complex tasks into “simple tasks”, and then manually designing each simple task, and the rules drive the planning of movement, which is a very primitive model.
With the language model and the world model, we can divide complex tasks into “simple tasks” through multimodal large models, understand human commands, and then further decompose them into machine actions through visual models, so as to complete the process of complex-simple-machine actions.
In general, the “breakthrough” of the unified multimodal large model is expected to accelerate the singularity of AGI. Under these two technical decisive points, let’s take a look at what advantages these players who have entered the large model finals have?
03 Where is the strength of the “Top 5 Basic Models”?
In the final analysis, the competition of artificial intelligence, especially large models, is a “big gamble”, testing the comprehensive strength of each family. It not only requires excellent technical strength, huge resources, and high-density layout of talents, but also requires strong determination and determination. In Crow Jun’s view, China’s top five AI can be divided into three camps: the “big factory faction” represented by Byte and Alibaba; The “national team” represented by Step Leap Star and Zhipu, and the “special forces” represented by DeepSeek.
Let’s talk about Byte first, which may be the most resolute company in China to invest in AI in the past year.
Byte’s investment in AI is not only reflected in the scale of investment, but also in the posture of investment. Byte founder Zhang Yiming invested most of his energy in AI like an entrepreneur, and Byte prepared new AI organizations Flow and Seed, independent of the original organizational structure, and formulated a new assessment method for the AI R&D team.
Today, Byte has invested in the whole chain of computing power chip layer, cloud computing layer (Volcano Engine), model research and development layer, and application layer. Among them, the main product, bean buns, will become the most AI application in China by the end of 2024.
Alibaba’s strategy is somewhat similar to Meta, focusing more on open source. Alibaba is not only the first company to choose an open source strategy, but also the company with the most comprehensive open source model in China. At the application level, there are also star applications such as Quark and DingTalk.
Let’s take a look at the stepping stars in the top two of the national team. Established in Xuhui, Shanghai in April 2023, Step Star is undoubtedly a star enterprise in Shanghai. At the end of last year, Leap Star announced the completion of Series B financing, with a total financing amount of hundreds of millions of dollars, and the core investors included Shanghai State-owned Capital Investment Co., Ltd. and its funds;
From the perspective of capabilities, Step Leap Xingchen has outstanding performance in model performance and commercialization. In terms of models, in the 2 years since its establishment, Step Star has released a total of 22 self-developed base models, of which 16 are multi-modal models, accounting for more than 70%, and have won the first place in many authoritative lists at home and abroad, known as the “multi-modal volume king” in the industry.
In terms of commercialization exploration, Step Star has a clear strategy and path. According to Jiang Daxin, founder and CEO of Jie Leap Xingchen, Jie Yue has previously mentioned in an interview that Jie Leap regards intelligent terminal Agent as the key force for the implementation of large model technology, and has won head customers in four key tracks such as smart cars, smart phones, IoT, and embodied intelligence, including Geely, Oppo, etc.
Similar to Jieyue, Zhipu is also a well-deserved large model national team, with a comprehensive model layout and a clear business path. According to the company’s disclosure, the Zhipu business will maintain rapid growth in 2024, with commercial revenue increasing by more than 100% and the average daily token consumption of the platform increasing by 150 times. Not long ago, the company announced the launch of an IPO, becoming the first company among large model startups to officially launch the IPO process.
Compared with the above four, the success path of Special Forces DeepSeek is relatively unique, focusing on language models, firmly opening up the route, and taking the path of ultimate engineering optimization.
Looking back, the above model companies shortlisted for the finals all have one thing in common, that is, they are rich + have people.
In terms of money, Byte and Alibaba naturally needless to say, DeepSeek also has the support of Magic Square, the largest quantitative fund company in China. Zhipu and Jieyue Xingchen have also maintained a good financing rhythm respectively. But more important than money is the layout of these companies in talent construction. For example, Wu Yonghui, head of Seed, the R&D department of Byte AI, has led projects such as Neural Machine Translation (GNMT) and RankBrainn at Google.
In addition to technology giants, the talent density of large model companies should not be underestimated. Take Step Star as an example, its chief scientist Zhang Xiangyu is a post-90s scientist who co-authored the ResNet (Residual Network) paper, which is known as one of the key papers in the field of artificial intelligence in recent years, with more than 250,000 citations.
In the long run, the road to AGI is a competition of comprehensive ability, and capital and talent will be the most critical factors in this competition. The reason is simple, money means the resources you can mobilize, and talent will determine the height of the company’s reach.
Now, the curtain of the big model finals has begun. Who will come out on top in the end, we’ll see.
