How long does it take to turn a piece of text into a video?
The answer might be just 10 seconds. With the help of AI large models, you only need to input a description of your requirements, and you might only have to wait 10 seconds to get the video you want, and the cost is very low or even free.
Behind the 10 seconds is not only technological progress but also the result of the "money-making ability" of large models.
For AI large models that have been rapidly developing since the end of 2022, finding landing scenarios and commercialization has become an important task this year.
Under the pressure of landing, large model manufacturers have begun to exchange prices for users and expand the market.
At present, several domestic manufacturers have announced price adjustments, and some people have jokingly said that large models have entered the era of "pricing in cents", and some models are even completely free for users and developers.
Advertisement
But in fact, large models are very expensive and require a lot of money to invest.
In the case where application scenarios and business models are not yet clear, what considerations are behind the price reduction chosen by several manufacturers? How to balance costs, profits, and long-term development? What impact will this have on users?
Several large manufacturers have begun to "roll" prices.
The last time large model manufacturers announced a price reduction was in September this year, and Alibaba announced the price reduction.
Specifically, the three main models of the Alibaba Cloud Bai Lian platform, Tong Yi Qian Wen, have been reduced in price again: both input and output prices have been adjusted, and new users will also be given free quotas.
Taking the input price as an example, the price reduction of Qwen-Turbo (128K) is 85%, and it is 0.3 yuan per million Tokens (the basic data unit processed by the model); the price reduction of Qwen-Plus (128K) and Qwen-Max input is 80% and 50%, respectively.
This is not the first time that domestic large models have reduced prices.
In May this year, Huan Fang Shen Du Qiu Suo (DeepSeek) officially open-sourced the second generation of MoE models, DeepSeek-V2, priced at 1 yuan per million Tokens input and 2 yuan output, which is nearly one percent of the price of the ChatGPT language model GPT-4-Turbo, setting a new low record for large model APIs (referring to Application Programming Interfaces, which can achieve data sharing and exchange through this interface).Subsequently, major AI manufacturers such as ZhiPu AI, ByteDance, Alibaba, Baidu, iFLYTEK, and Tencent have successively followed suit with price reduction actions. Among them, ByteDance announced that the pricing for the main model of DouBao in the enterprise market is 0.0008 yuan/thousand Tokens, and the price of large models has entered the "cent era"; some models of Baidu, iFLYTEK, and Tencent are completely free.
Both B-end and C-end users can use large models at lower prices or even for free.
Why choose a price reduction strategy when the application scenarios of large models are not yet clear, and commercialization is also gradually exploring the road?
GuoSheng Securities research report believes that the price reduction of large model APIs may stem from the progress of large model inference technology and the reduction of inference costs, which objectively gives developers more choices.
Huatai Securities pointed out that the main ones with a large price reduction range are Internet companies, which have more abundant resources. The price reduction is mainly focused on entry-level lightweight APIs, which can be explained from the aspects of technical optimization and ecosystem capture.
"On the one hand, DeepSeek-V2 has proven that it is indeed possible to greatly reduce the amount of cache required through attention mechanism optimization, reducing costs and improving efficiency. On the other hand, even without a technical price reduction logic, attracting developers through price reduction means is still an important means of capturing the ecosystem." Huatai Securities research report pointed out.
Alibaba Cloud CTO Zhou Jingren said in an interview with the media during the Cloud栖 Conference that the main reasons for the price reduction are mainly considered from the aspects of scale effect, high-quality technology, and resource scheduling. In addition, it is hoped that the price reduction will stimulate more industrial innovation.
At present, the price reduction of large models has not yet seen the end.
After the large model manufacturers concentrated on price reduction in May, they continued to "roll" the price, and in the next few months, ZhiPu AI, Baidu, Deep Search, Alibaba, etc. announced price reductions again on the basis of the above adjustments.
"Generally speaking, price reduction is mainly related to two factors. One is that the product has not yet adapted to market demand and needs to win more users through price reduction. The second is to solve the problem of retaining users through price reduction." Ai Media Consulting CEO and Chief Analyst Zhang Yi said in an interview with a reporter from China Economic Weekly, considering the current AI technology capabilities and user usage habits, it is expected that this round of AI price reduction has not yet reached the bottom.Price reduction is just the beginning, with tough competition still ahead.
Regarding large models, the industry is also anxious about "prices even reaching negative gross margins." Does the price reduction of large models signify the start of a price war?
Tan Dai, President of Volcano Engine, stated in an interview with the media at the Volcano Engine AI Innovation Roadshow in September 2024 that this is not a price war, but rather reducing costs to a reasonable level, doing a good job with applications, and unlocking more scenarios.
Zhou Jingren also does not believe this is a price war. In his view, it is akin to mobile internet charges; today, 200 yuan can be used for dozens of gigabytes of data, but 20 years ago, users would likely "go bankrupt" using dozens of gigabytes of data.
"Today's prices are not low enough; relative to the vast applications of the future, they are still too expensive," said Zhou Jingren.
"Domestic manufacturers will not blindly engage in a 'price war' in this wave of price reduction; they will still consider factors such as costs and proceed 'step by step,'" said Huatai Securities.
The price reduction of large models has also brought about more thinking.
"The main impact is on the market. First, the capital market's consideration of the sustainability of large model business models and future large-scale profit models; second, users, whether B-end or C-end users, their game with manufacturers over prices will continue," said Zhang Yi.
For the promotion of large model applications, price reduction is just the beginning; future technology will become an even more hardcore competition.
In Zhang Yi's view, the moat and threshold for large models still lie in the product's problem-solving ability, that is, user satisfaction. Large model manufacturers still need to work hard on solving user needs.Tan Dai stated that the next step is to improve quality and performance on the basis of this price. Quality mainly refers to making the model more capable and diverse.
Since the beginning of this year, AI has gradually moved towards practical application, and manufacturers have been continuously fighting in the hardware field, from AI mobile phones to AI PCs (Artificial Intelligence Personal Computers). For AI, in addition to combining with hardware, it has also been endowed with more imagination: text dialogue, video generation, and even intelligent butlers like Jarvis in "Iron Man".
It is necessary to continuously prosper in applications and promote large-scale commercialization for large models to truly "fly" into ordinary people's homes.
"The road to commercialization of large models is not smooth," said Zhang Yi.
In his view, the problem faced by the commercialization of large models now is to what extent the products can meet user needs. At present, there is still a significant room for improvement in the level of intelligence and the ability to meet needs of large model products from various manufacturers.
In addition, Zhang Yi believes that the fierce competition among large models will also affect commercialization. Under fierce competition, there is a large uncertainty between the early R&D investment and the future output results. This uncertainty will lead to a more cautious attitude towards early R&D investment from relevant entities.