In the business of large models, the focus of competition is rapidly shifting from “whose model is stronger” to “who can make a profit back.”

However, it is not easy to recoup the money that has been invested. Even companies like Zhipu and MiniMax, although their revenue has doubled in the first half of the year, still suffer losses of over 2 billion yuan and nearly 300 million US dollars respectively.

In order to achieve "hematopoiesis" (i.e., generating revenue through their own operations), large model companies have been focusing on the most rigid Coding (programming) scenarios. However, whether it's Zhipu, DeepSeek, or MiniMax which relied on price cuts to gain momentum, none of them were able to break free from the financial shackles of shrinking gross margins and research and development expenses that far exceed revenue. In the end, they still had to rely on new financing amounts in the tens of billions to solve their computing power issues.

Unicorns are burning money, while small and medium-sized startups struggling in the niche market face numerous challenges.

On the commercial side, B-end projects have a high degree of customization, and many companies find it difficult to escape the situation of being "purely custom-made workers." On the funding side, hot money is accelerating its concentration in the hands of a few giants—according to IT Juzi's statistics, in the third quarter of this year alone, 43 financings of over 1 billion yuan swept up nearly half of the total market funds, while the average financing amount in the enterprise services sector has shrunk to around 33 million yuan per deal.

The winter of this AI gold rush has arrived. How on earth can one survive?

Unicorns are fighting on the Coding battlefield.

This year, companies such as Zhipu, Dark Side of the Moon, and MiniMax are all competing for the Coding market. In an industry where it is generally difficult to find scenarios with a high willingness to pay, Coding has become one of the few places where "customers are willing to pay based on results."

Why is it Coding? Zhu Xiaohu believes that Coding is one of the fastest-growing scenarios this year, due to "rigid demand and easy measurement of output." Developers can quickly see whether the code works properly and how much man-hours are saved, which gives model companies more room for pricing.

Also competing for Coding, the pricing strategies of the four companies have gone in different directions. Zhipu started raising prices in February this year. Zhipu made a structural adjustment to the pricing system of GLM Coding Plan, with an overall increase of 30% or more. The reason given was that the user base and the amount of usage were growing rapidly, necessitating increased investment in computing power and model optimization.

However, price increases did not suppress demand. As of the release of the interim report, the number of calls for Zhipu Coding Plan has increased by more than 23 times compared to the beginning of the year, and the average selling price of API has increased by about 101%. Revenue from the open platform and API has increased by over 27 times year-on-year, reaching 825 million yuan, and its proportion of total revenue has risen from 15.2% a year ago to 86.5%, replacing localized deployment as the company's main source of income.

Unlike Zhipu's continuous price increases, the pricing strategy of "The Dark Side of the Moon" involves back-and-forth adjustments. When Kimi K2 .6 was launched in April, the prices of some API products increased by about 58%; after the release of K3.1 and Pro in October, the price of the new model API decreased by about 30% compared to K3 from July, and it further enhanced programming and agent capabilities.

If the price adjustments of Zhipu and the Dark Side of the Moon are still in the range of several dozen percentage points, then the price increase of DeepSeek is on a multiple order of magnitude. It is reported that DeepSeek will raise the prices of different models of API to 2.3 to 4.5 times the original this year. People familiar with the matter revealed that Leung Man-fung told investors that the company's annual revenue has exceeded 1 billion US dollars.

The first three adjusted prices, while MiniMax focused on reducing costs. MiniMax chose to lower costs first. Yan Junjie revealed during the mid-term performance conference that in the past two months or so, the throughput per unit of computing power for text models has increased by three times. By July this year, the consumption of Token on the MiniMax platform had reached 20 times that of the beginning of the year; in the first half of the year, its revenue from open platforms and enterprise services was approximately 73.9 million US dollars, a year-on-year increase of 703.1%, and its proportion of total revenue rose from 30.3% to 63.4%, surpassing for the first time consumer-oriented products such as Hailuo and Talkie.

When looking at the accounts of these companies together, the revenue structure has changed, but this business has not become more profitable as a result.

In the first half of this year, the gross profit margin of Zhipu's API business turned from negative to positive, reaching 24.6%. However, as the proportion of local deployments, which have higher gross profits, decreased, the company's overall gross profit margin fell from 50% to 26.4%. The gross profit margin of MiniMax increased from 12.1% to 17.9%, with gross profits in the first half of the year amounting to approximately 20.8 million US dollars.

The greater expense is incurred in research and development. In the first half of this year, Zhipu's R&D expenditure was 2.2 times its revenue for the same period; MiniMax's R&D expenditure was 297 million US dollars, which is about 2.5 times its revenue, mainly used for cloud service expenses related to training.

The gap between research and development and revenue is currently being filled through financing. In July, Zhipu completed a placement of approximately HK$31.4 billion, with the funds mainly directed towards computing power; MiniMax raised about HK$15.9 billion net through rights issues and convertible bonds in the same month.

Financing can fill the gap, but for how long? Xie Siyuan, the managing director of Yijing Capital, believes that large-scale models are inherently capital-intensive industries, and the continuous investment in computing power is raising the competitive barrier. Even some mid-tier model companies that have already generated revenue may not be able to keep up with the investment pace of leading manufacturers, and there is still uncertainty as to whether they will be able to remain competitive in the future.

Surviving in the cracks: Less than 50% of standard products are delivered on time

Leading model companies compete for customers, and small and medium-sized startups with limited funds and resources face greater competitive pressures. Turning to specialized niche markets has become a more realistic option.

Some of Lanzhou Technology's orders come from large companies. Their products are included in the procurement systems of cloud service providers, and some financial projects are also contracted out by large companies, which are then handed over to Lanzhou Technology for development. “They sign contracts with customers, and we sign contracts with them,” said Li Jingmei, co-founder of Lanzhou Technology.

Li Jingmei believes that some large companies ‘have platforms and computing power, but lack applications.’ However, as they gradually enter the general application market from cloud and model services, the opportunities for small and medium-sized companies are also narrowing. ‘If what I do is the same as what they do, in a parallel manner without any competitive advantage, then we will have no chance left when they move forward.’

The company, which once developed the "Mencius" pre-trained model, ceased basic model training in 2024 and has since largely withdrawn from general platforms and scenarios, focusing its business on intelligent agents for specialized fields such as finance.

Li Jingmei noticed that this year, financial institutions and central state-owned enterprises have started to shift from small-scale pilots to widespread use of AI within their organizations. Some customers are even willing to pay separately for consulting services. However, these customers usually set their budgets a year in advance, and many projects require public bidding with low-price winners. The expenditure on computing power is often drawn from the same fund. “If the cost of computing power is too high, then there will be less money available for other expenses allocated to us.”

In a legal compliance knowledge Q&A project at the headquarters of a certain bank, Lanzhou initially used standard products. However, it was later found that these products were insufficient in terms of professional terminology, versions of legal provisions, and the formulation of questions in colloquial language, necessitating additional development. After the system was launched, trial operations were required, along with one year of free operation and maintenance services. "To achieve over 50% effectiveness with standard products is already quite impressive in China's B2B market," said Li Jingmei.

Of course, Lanzhou is not the only one that relies on projects to sustain its operations. Since the beginning of this year, some AI development teams that have not been able to secure financing have turned to providing customized deployments, training, and after-sales services for enterprises. A person from the AI industry described their situation as follows: The applications are still used to showcase creativity on official websites, but the actual revenue comes from providing AI upgrades for enterprises.

According to statistics from Zero One Think Tank, in September of this year, the average winning bid for intelligent agents of financial institutions was about 2.11 million yuan per project. Intelligent agents are increasingly being packaged as a module into larger system construction projects.

Lan Zhou has gradually raised its requirements for the gross profit margin of projects over the years, filtering out unprofitable business opportunities, and has transformed the engineering capabilities from these projects into standard products to reduce duplicate development. After completing the Pre-A series of financings in 2023, it has not sought public financing since then, operating on the funds from previous rounds of financing and project payments, which has helped to narrow its losses year by year. Li Jingmei also admitted that "purely customized To B projects are definitely not sustainable."

Independent model companies still need to rely on financing to fill the gap in research and development; whether small and medium-sized AI startups can achieve what Li Jingmei mentioned, namely "a thicker platform and more tailored scenarios," is the question they need to answer next.

Source: Shenwang Tencent News