The price of large models has fallen, and the Internet-style "turf war" has reappeared, will big factories really lose money?

National Business Daily

2024-05-23 21:47Published on the official account of Sichuan Daily Economic News

Every reporter: Yang Hui Ke Yang Every editor: Wei Guanhong

At the beginning of this year, the race in the field of long text made many domestic large-scale model companies show their faces on the "involution" track. Entering May, this war without gunpowder "burned" to the most brutal stage - price reduction. From the two overseas AI (artificial intelligence) giants OpenAI and Google, to domestic companies such as ByteDance, Alibaba, Baidu, Zhipu AI, and iFLYTEK, price reduction players are still "relaying".

Judging from the data announced by all parties alone, this round of price reductions is stronger than the other, and there is a tendency to "smash through the floor price", which once caused heated discussions in the industry.

It is unknown whether to "post money", but for the current large-scale model enterprises, whether it is active customer acquisition or passive adjustment with the market, price reduction has become the general trend. It is worth digging into whether players who take this step will really lose money?

OpenAI, Google, etc. cut half the price, and domestic manufacturers simply "find out"

How hard is this round of price cuts? Judging from the current magnitude announced by all parties, the two overseas AI giants, OpenAI and Google, were the first to start cutting prices, but the magnitude was relatively small. Among them, OpenAI's GPT-4o API call price is halved compared with GPT-4-turbo, at $5/million tokens; Google, which has been slower to cut prices, is cheaper, and the price of Gemini 1.5 Flash has dropped to $0.35 per million tokens, which is much cheaper than GPT-4o.

Image source: Visual China-VCG31N2008743681

Domestically, despite the fact that celebrities in the industry such as Kai-Fu Lee, CEO of 01000 Everything, publicly called for avoiding the "ofo-style" bloodshed and money-burning style, the "tragic" price reduction curtain still came quickly and ruthlessly.

On May 15, ByteDance took the lead in "shooting", and its volcano engine announced that the main model of bean bags was priced at 0.0008 yuan/1000 tokens in the enterprise market, saying that it was 99.3% cheaper than the industry, which has set off a low-price "hand-to-hand battle" of domestic large model manufacturers.

In order to explain the price reduction in detail, ByteDance said at the press conference that users can process the content of 3 copies of "Romance of the Three Kingdoms" for 1 yuan. At that time, GPT-4 (0.42 yuan/1000 Tokens), Ali Tongyi Qianwen and Baidu Wenxin Yiyan (0.12 yuan/1000 Tokens) were also compared.

Ali quickly "shot back".

On the morning of May 21, Alibaba Cloud announced that the API price of Qwen-Long, the main model of Tongyi Qianwen, dropped by 97%, and the API input price dropped from 0.02 yuan/10005 yuan/10005 yuan/1000 tokens, a drop of 97%, which was lower than that of bean bags. After Alibaba Cloud announced a significant price cut, Baidu posted on its official account that the two main models of Wenxin model, ERNIE Speed and ERNIE Lite, are free of charge and will take effect immediately; Prior to this, Zhipu AI also adjusted the glm3-turbo price to 1 yuan per million tokens. On May 22, iFLYTEK announced that iFLYTEK Xinghuo API capabilities are free and open.

Image source: Screenshot of the webpage

At this point, it seems that it doesn't matter who "moves" first.

The marginal cost continues to increase, is the Internet "enclosure play" still applicable?

From "volume technology" to "volume price", what has the artificial intelligence technology represented by large models encountered in the process of change? In the view of economist and new finance expert Yu Fenghui, from domestic to foreign, price reduction may be more of an independent choice made by manufacturers based on market competition and cost considerations, hoping to attract more users and market share. Guo Tao, an angel investor and senior artificial intelligence expert, believes that the increase in open source projects may also put pressure on the price of commercial products.

Cheng Yin, research manager of IDC China, told the "Daily Economic News" reporter that from the perspective of short-term development, the large model capability will converge after the update and iteration, and neither domestic nor foreign technology suppliers can establish a long-term moat. Some technology vendors choose to directly cut the threshold on the cost of large models, not only to promote the application of large models, but also to increase exposure, compete for users, and prevent user loss.

In addition to the idea of "taking the initiative", Shen Yang, director of the Metaverse Culture Laboratory of the School of Journalism of Tsinghua University, also pointed out that manufacturers are fighting a "price war" at this point in time, indicating that the functions and performance of each large model are rapidly approaching, and the stage of imitating GPT-4 has basically come to an end, and AI has moved from innovative experiments to the stage of mass promotion. As the stage of rapid expansion arrived, market share became the primary concern. "If you fall behind at this stage, it will be useless to function well."

In short, whether it is active choice or passive follow-up, the price war of large models has indeed arrived. If this trend continues, will it really fall to the level of "subsidizing money" to attract customers ridiculed in the industry? The reporter of "Daily Economic News" consulted a number of industry insiders, but as of press time, no specific price reduction figures or price reduction space have been obtained.

Zhan Junhao, a well-known strategic positioning expert and founder of Fujian Great Aim Brand Positioning Consulting, said that price reduction is actually a common development model in the Internet industry, and all Internet projects must lose money to do traffic, do user groups, and cultivate users to develop usage habits in the early stage. The advantage of large factories lies in the overall ability, and the funds often do not rely on profits, but from the secondary market. However, Zhan Junhao said frankly that the price reduction will definitely increase all kinds of pressure on enterprises. However, in the absence of a special breakthrough in technology, enterprises can only win the market and users through the "simple and crude" way of price reduction.

However, the changes caused by artificial intelligence, represented by large models, are very different from those in the early Internet era. Ji Yu, the founder of Xingyun Integrated Circuit, said bluntly in a recent event that in the past, the cost of adding a user to the infrastructure of Internet manufacturers was almost negligible. In contrast, the marginal cost of large models is still very high, and the cost of infrastructure is visible to the naked eye for each additional user. More importantly, at present, large models are to be commercialized on a large scale, and there are further demands in terms of model quality and context length, which cannot be ruled out to further increase the marginal cost.

So, can the development models of Internet companies such as ofo, Meituan, and Pinduoduo still be applied to large-scale model enterprises? In Cheng Yin's view, the ability of large models and the experience of using related products will affect whether users choose to retain, but in the end, technical products are king. Some manufacturers with good generation quality/performance, good business model and market expansion trend, and relatively abundant cash flow can amortize R&D costs in the long run, but some start-ups have higher financial risks.

The reporter noticed that it is also a current trend that well-known start-ups do not have all employees. Wang Xiaochuan, founder and CEO of Baichuan Intelligence, revealed that in this round of price cuts, he is more like "eating melons and watching the excitement". In his opinion, this price reduction is a B-end strategy, more like the price reduction method of the "Four Little Dragons", rather than price competition around the C-end like Didi and Meituan. It is not a change in the relationship of production, but a direct supply of productivity, and the direct supply of AI to productivity. "I think for us, it's just not to get involved." Wang Xiaochuan said.

Cloud vendors may move away from the traditional service model, and price reductions are expected to activate API model calls

How should large-scale model enterprises after the price reduction compensate for this loss?

Wang Xiaochuan gave an idea. According to him, this round of price reductions is limited to the actions of cloud service providers. "The core depends on what your business model is, if you are doing To B services, what you sell in the end is not the model itself, but the whole set of cloud services, so cloud vendors are more traditional service models and have entered a new battlefield."

In this regard, NVIDIA, the AI giant that has just released its outstanding results, has given more detailed reference data.

On May 22, local time, Nvidia announced its financial results for the first quarter of fiscal year 2025 as of April 28. In the subsequent earnings call, Nvidia's executive vice president and chief financial officer, Colette Kress, revealed that training and inferencing AI on Nvidia's CUDA can drive growth in cloud leasing revenue, and that every $1 of Nvidia's AI infrastructure spending gives cloud service providers the opportunity to earn $5 in GPU instant hosting revenue over four years.

Nvidia headquarters Image source: Photo by reporter Zheng Yuhang

In addition to attracting traffic to cloud services, it should also be noted that the commercial use of large models not only calls a business path for API models, but also has multiple modes such as paid subscribers and rewards, and the degree of API model calls is not high. Kai-Fu Lee once pointed out at a recent press conference that today's API model calls are still a very low proportion. "I think the industry as a whole can be expected to reduce the cost of reasoning to one-tenth per year, and it must happen. If it is reduced to 1/10 a year so that more people can use it, this is very good news. Kai-Fu Lee said.

Wang Xiaochuan also proposed that large model enterprises have an expectation of reducing the cost of reasoning about the price reduction of the model itself. "I believe that large manufacturers also expect that the cost of the model will be reduced a lot in the future, maybe lose money today, and not lose money in another year, in this case, large manufacturers are not considering short-term losses, but considering whether there is an opportunity in the future, and the model will be cheap enough, not just to attract customers." From the perspective of the industry, a large part of the current internal business of large manufacturers is still calling AI applications and business exploration, and standardized model APIs have not ushered in deterministic growth. It is expected that the price reduction will attract more users, which will further activate the market.

The cost of using large models is reduced, and the application side "takes off"?

Extending the focus to the application side, the price drop is undoubtedly a positive.

Judging from the current situation, the domestic end-side model has certain advantages in terms of application. According to Weng Xi, a professor at the Department of Applied Economics at Peking University's Guanghua School of Management, at a recent event, many fields in China are currently applying artificial intelligence for digital and intelligent decision-making, and have achieved certain results. In his view, the main advantage of China's development of AI lies in its strong ability to apply AI in specific vertical industries and fields, such as finance and marketing. In addition, Chinese society is more inclusive of the development and application of AI, which also contributes to the wider application of AI among the public.

With the start of the price war for large models, is the "iOS-style success" of large models coming? In this regard, Li Mingshun, chairman of Xingxing AI, proposed that the price war may stimulate application companies to intervene in applications faster. "In the past, many businesses did not accept to make an application, because the cost of computing power was too high, and many applications could not be implemented. Li Mingshun said.

Shenyang believes that on the application side, the current price war is conducive to To B users to obtain AI services at a lower cost. For the C-side, with the intervention of large manufacturers, it will make it more difficult to obtain customers for To C applications. "The head Tencent, Alibaba, Baidu, and Douyin, if they fully promote AI, in fact, the Internet will popularize AI applications in an instant. For small and medium-sized factories, it is very difficult to obtain customers. Shenyang believes that small and medium-sized manufacturers are now playing more of a role in experimenting with business models, and once mature, large manufacturers will quickly follow suit.

In Wang Xiaochuan's view, although the price war may give birth to some small applications, this is not the main direction of Baichuan Intelligence. "We don't want to compete with these smaller startups now, each with its own niche, because they are not making super apps, they may want to make small apps in a certain vertical." He believes that the emergence of open source and even low-cost closed-source large models will make each vendor more clear about its own positioning: whether to focus on application development, build models independently, or integrate models with applications.

"At this stage, there will be some companies that have not thought clearly; The rest of the people make models, some people do applications, and the market will become healthier. In particular, large manufacturers will prosper faster if they make contributions, whether it is To B or applications. Wang Xiaochuan said.

In addition, Cheng Yin pointed out that the cost of using large models continues to decrease, which will promote the trial and trial of large models in the short term, but the actual application depends on the actual business value of large models. In this regard, IDC suggests that technology vendors and industry users should first evaluate the application scenarios of large models, and determine priorities based on their value, complexity, and other comprehensive evaluation use cases. Consider starting with scenarios with high value and low implementation complexity.

National Business Daily

View original image 1.19M

• 

• 

•