On November 29, after the Zhipu Agent OpenDay meeting, a reporter asked CEO Zhang Peng a question: "How is the progress of To B?"

"It's okay. "Yeah," Zhang Peng didn't say another word until now.

This year, this star AI company has been deeply involved in the battle of large models. Its competitors on the same stage are Baidu, Alibaba, Tencent and Byte. The fate of Zhipu and other AI startups is exactly the same. Being shortlisted is the "consolation prize", and in most cases it ends up as "participation is the most important thing".

The anti-giant encirclement and suppression campaign was quite difficult. As of December 3, a search on the China Bidding and Bidding Public Service Platform using the keyword "large model" found that there were approximately 200 successful bids for large models. According to incomplete statistics from Photon Planet, the number of bids won by the above-mentioned four major cloud vendors reached 98, accounting for nearly 50% of the total orders. Only the amount of orders related to large models and AI, the four companies took away a total of 1.112 billion yuan, dividing the market share. Some state-owned enterprises, government and enterprise markets.

AI companies seem to have fallen into a self-certifying trap and need to rely on press conferences and new releases to continuously generate buzz to attract B-end customers. But these are not worth the innate advantages of cloud giants: computing power order prices are crushing in the market; pure large model functions cannot be priced high, and they are packaged into solutions together with cloud, database and SaaS, which can not only level the cloud Infrastructure costs can also increase negotiation weight; customers are path dependent and based on their usage habits, they directly send an offer to a cloud vendor they have cooperated with.

A more severe reality is unfolding. In the To B market, cloud vendors also took the lead in waging a price war, increasing volume without increasing prices. At one point, they won orders at less than half the price of their competitors. In addition, the first-mover advantage of AI startups is gradually fading, and services among competitors are becoming homogeneous. The gap between them and large manufacturers is not obvious or has been surpassed.

At this point, the overall situation of the large model To B market in 2024 has been determined: Yun Dachang eats the meat, and the AI ​​company eats the soup.

Since last year, Major cloud companies have successively elevated AI and large models to a strategic level.

Each company has different ideas. Alibaba Cloud has implemented the MaaS concept, and its core is to use large models and AI products to drive cloud revenue. At present, Alibaba Cloud-related revenue consists of five parts, including the revenue generated by calling Tongyi Qianwen for Alibaba's internal business, the revenue generated by Alibaba products such as DingTalk and its customers purchasing and consuming the cloud, the AI ​​companies invested by Alibaba and The income generated by its customers, the income from developers calling large model tokens, and the income from To B market orders, etc.

Alibaba Cloud is currently the largest vendor with the most complete cloud service form. Both open source and closed source are undergoing market tests. Driven by the above-mentioned revenue modules, Alibaba Cloud has achieved both revenue and profit.From Q1 to Q3 in 2024, its revenue increased from 25.595 billion yuan to 29.61 billion yuan; its adjusted profit increased from 1.432 billion yuan to 2.661 billion yuan, a year-on-year increase of 89%. Unfortunately, Alibaba Cloud has not yet reached its year-on-year double-digit growth target.

Baidu is a radical group in AI and large models, and is at the forefront from releasing large models to AI transformation. Baidu's accumulation in the IaaS and PaaS layers is not as deep as Alibaba's. It realized that infrastructure construction was not the way out, so it focused on large models and AI applications. Baidu Cloud's revenue mainly consists of three parts: token revenue generated by developers calling the Wenxin model, revenue generated from authorized API interfaces, and To B market order revenue.

Due to the inability to generate economies of scale at the infrastructure level, Baidu is particularly active in developing B-side business. Many entrepreneurs told Photon Planet that Baidu’s marketing department will proactively contact them and promote cooperation with other startups in order to obtain orders as soon as possible and increase the number of token calls for the Wenxin model.

Catching up with this wave of large model craze, the growth of Baidu Cloud is the most obvious. The Q3 financial report of 2024 shows that the proportion of AI revenue has increased to over 11%. The growth is mainly driven by the high demand for model training and inference in the Internet, education, finance and other industries. Among them, incremental revenue from waist enterprise customers increased by 170% month-on-month.

However, as similar products continue to emerge on the market, Baidu's irreplaceability has been reduced. Although large model and AI revenue are still rising, the growth rate has slowed down significantly. Baidu Cloud’s Q3 Q3 growth rate dropped from 14% to 11%, and the Q3 growth rate of generative AI cloud revenue plummeted from 95% to 17%.

Tencent tends to be conservative and prefers to enter when the industry situation becomes clearer. This is true for large models, AI assistants, AI drawings and AI videos. Compared with Alibaba and Baidu, which are studying how to win new customers, Tencent Cloud is thinking about how to use AI to meet existing business needs. Therefore, integrating AI into the ecosystem has become a must-answer question for Tencent.

Tencent Cloud’s AI revenue is scattered across various businesses, like pebbles invested. The ripples are gradual and difficult to include in revenue in a short period of time.

Tencent revealed in its 2024 Q3 financial report that the growth of the marketing services segment was driven by video accounts, mini programs, WeChat search advertising and AI technology. Currently, AI-related revenue accounts for about 10% of Tencent Cloud services, and the AI ​​business is expected to generate considerable free cash flow next year.

The large model market in China is not yet fully developed and is full of uncertainties. Customers are hesitant, and the foundation for adapting large models and AI functions is incomplete. Developers are "shooting one shot and changing the place." At this stage, in China's general environment, real, sustained and stable demand for large models comes from government and enterprises. Various manufacturers say it is To B, but thisIt's essentially To G.

For cloud vendors, the channels, experience and customers in the G-end market are readily available. Just follow the process of the previous stage again, and the protagonist is changed from SaaS and cloud to large model and Agent. Therefore, taking orders has become the way for Yundachang to increase points crazily this year.

Based on public data such as the China Government Procurement Network, China Tendering and Bidding Public Service Platform, and Caizhao.com, Photon Planet has tracked Baidu Cloud, Alibaba Cloud, Tencent Cloud and Huoshan since this year as of December 3, 2024. Bid winning status related to the cloud model. In particular, only related orders generated by large models and generative AI are counted. Large model application services usually refer to large model application platforms, tools, software, etc., and large model deployment training usually refers to model training, inference, tuning, etc. There may be omissions in the statistical results, and undisclosed bid winning amounts are not included in the statistics.

Baidu Cloud won 34 bids, ranking first among the four companies in terms of number and industries involved. Among them, the financial industry has the largest number of winning bids, followed by communications, electric power, education and scientific research, environment and public facilities management. Baidu Cloud has won a total of 446 million yuan in orders, the majority of which are computing power (216 million yuan) and large model application services (170 million yuan), accounting for 49% and 38% of the total amount respectively. Perhaps because of its talent in the field of AI search, only Baidu Cloud has received orders, but the unit price is not too high, with a single transaction not exceeding 2 million yuan.

(Photon Planet Mapping)

Alibaba Cloud won 18 bids, with finance, education, scientific research and government affairs being its main fields. Although it is less than Baidu in terms of quantity, Alibaba Cloud won a total of 426 million yuan in intelligent computing orders with 396 million yuan, almost tying Baidu Cloud in terms of total winning bid amount. In addition to computing power, RMB 12.96 million was deployed for large model training and RMB 7.69 million for large model application services.

After sorting it out, Alibaba Cloud’s strong areas are actually the old cloud business, such as cloud expansion, public cloud, database, middleware, hardware and other procurement. New businesses such as digital people, intelligent customer service, and AI programming account for a low proportion, and Alibaba Cloud does not have outstanding advantages in large model training and deployment. The procurement source of Tongyi Qianwen's large models comes from the scientific research institution "Beijing" that Alibaba participated in and funded. Jiang Laboratory".

(Photon Planet Mapping)

Tencent Cloud has won a total of 24 bids, with a total amount of 180 million yuan, concentrated in the media and communications fields. By sorting out the orders, Photon Planet found that governments and enterprises have the clearest understanding of Tencent Cloud and Hunyuan series products, including several keywords such as digital people, multi-modal, and large models, which may be "MoE labeled" with Tencent Hunyuan. It has something to do with taking the open source route. This brings its share of the winning bid in large model training and deployment to 72%. The Shenzhen Baoan District Government directly adopted Tencent Cloud and Hunyuan large models, and radio and television and news media also proposed fine-tuning vertical models and upgrading platforms. demands.

(Photon Planet Mapping)

A total of 22 volcanic clouds were awarded, but the total amount was only 61.59 million yuan. Basically get all the volcanic cloudsFor orders placed in the smart agent segment, due to the complexity and customization difficulty of the smart agent, its unit price will fluctuate depending on the scale of research and development.

Among the five smart body orders, the lowest was 480,000 yuan and the highest was 4.2 million yuan, with a lower unit price. The winning bid of Huoshan Cloud reflects its marketing strategy. Judging from the results, Kouzi is accelerating its shift from the C-side to the B-side, assuming the function of connecting large models and To B business. In addition, the audio and video capabilities of the bean bag model have also received attention.

(Photon Planet Mapping)

In most cases, the four cloud vendors are competing on the same stage, so the distribution of winning bids by industry and amount is relatively similar. At this stage, except for large orders of computing power, it is difficult to say which manufacturer has an overwhelming advantage. A common problem that everyone encounters is that there is more money than meat. If you exclude computing power, you will find that the unit price of large models and AI-related customers is very low.

After excluding computing power, the average customer price of Baidu Cloud is about 7.42 million, Alibaba Cloud is about 2 million, Tencent Cloud is about 7.5 million, and Huoshan Cloud is about 2.8 million. There may also be an issue of profit distribution involved behind the scenes. According to the usual practice of large manufacturers, an order will be split into N parts and subcontracted to different companies.

In many winning bid orders, the procurement announcement reads "single procurement source" , in most cases it is the cloud vendors that ultimately succeed in bidding. This means that from the beginning, many AI manufacturers lost in the qualifying rounds.

What is reflected behind this is the long-term resource monopoly of cloud vendors in this track, as well as the characteristics of the cloud itself. In order to pursue security, continuity and stability, and reduce migration and development costs, many state-owned enterprises have chosen to Continue to use large models and AI product services on the basis of using the vendor's cloud.

To B or To G is ultimately about solving a problem. It is a project-based target, and its requirements are complex. B-side customers may need not only hardware and cloud services, but also large model inference training and Agent application development services. What they want is a complete set of customized solutions that integrate software and hardware.

For example, in a certain education digitalization order, the purchase list includes requirements for cloud expansion, customized large model calling services, document tools, office software, and data analysis. For start-up companies that are new to the industry, their own products and services simply cannot meet the above order demands, and part of the market is automatically excluded.

The opportunity for AI manufacturers to compete with Yun Dachang on the same stage comes from their outstanding individual capabilities. For example, one bid is split into multiple components such as Vincent Picture, Tush Picture, Vincent Music, and Vincent Video. Sometimes, purchasers will compare the same functional effects of startups and large manufacturers. Zhipu did not perform well in the comprehensive list, but it was still under siege by big manufacturers and got some orders for multi-modality, large model training and tuning, large model pre-training and AI video.

This invisibly leads toLeading AI manufacturers towards the path of "complementation". At present, almost all first-line AI startups are involved in several mainstream directions such as large model training, reasoning, AI search, AI mapping, and AI video. As they become more and more homogeneous with each other, the gap between the big manufacturers is getting smaller and smaller. Some of them are also found in big factories, and their service output is more standardized and stable. This year, Yun Dachang won nearly half of the orders in the large model and AI markets. Zhipu barely got the soup, and the rest just smelled of meat. It will be more difficult in the future.

To G has a long delivery cycle and slow payment collection. These two points alone are enough to kill a number of AI companies struggling on the line of life and death. For a large factory, large model orders are the icing on the cake; for a start-up company, they are a source of income. The nature of the two is completely different. Yun Dachang, which has strong financial resources, can afford to consume and fight a price war.

The unit prices for large models and AI-related products are not affordable, and these big cloud companies have to bear a certain responsibility. In the competition among candidates during the same period, they obtained orders by lowering prices. In this year's competition for "domestic intelligent auxiliary programming services", iFlytek quoted a price of 980,000 yuan, and Flush quoted a price of 560,000 yuan. Alibaba Cloud came forward and cut the price in half to a transaction price of 350,000 yuan.

Startup companies have only two choices to face the competition, either leave the market or offer a lower price. In another AI programming R&D procurement, Baidu Cloud offered 660,000 yuan and Tencent Cloud offered 330,000 yuan. A software startup offered a low price of 280,000 yuan and finally got the order.

Thus, the vicious cycle of the large model To B market began. As the strong become stronger, the cards left in the hands of AI companies will be at their lowest.