alibaba cloud cuts prices again, why is big model still not in a hurry to “settle accounts”?
2024-09-19
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
"free" and "price reduction"... on september 19, at the 2024 yunqi conference, alibaba cloud released a number of products and announced a new round of price cuts. "alibaba cloud will work hard to continue to reduce costs," said wu yongming, ceo of alibaba group and chairman and ceo of alibaba cloud intelligence group.
behind this, for cloud vendors, the industry is still in the early stages of the agi transformation, and the "price reduction trend" for large models shows no signs of ending.
"over the past period of time, the cost of model inference has dropped exponentially, far exceeding moore's law. over the past year, the price of calling tongyi qianwen api on alibaba cloud bailian has dropped by 97%, and the lowest cost of calling one million tokens has dropped to 50 cents." wu yongming revealed in his speech on the morning of september 19.
in the afternoon of the same day, the minimum call cost of alibaba cloud qwen-turbo's million tokens was once again refreshed to three cents. zhou jingren, chief technology officer of alibaba cloud intelligent group, announced that the price of qwen-turbo would drop by 85%, and qwen-plus and qwen-max would be reduced by 80% and 50% respectively. alibaba cloud has started a new round of price cuts.
at the same time, alibaba cloud also released a new generation of open source model qwen2.5, and its visual language model qwen2-vl-72b was also officially open source, which can recognize pictures of different resolutions and aspect ratios, and understand long videos of more than 20 minutes. alibaba cloud tongyi's flagship model qwen-max has also ushered in an all-round upgrade. zhou jingren said that its performance is close to gpt-4o. the background models of tongyi's official website and tongyi app have been switched to qwen-max, "and continue to provide services to all users for free." in addition, zhou jingren also announced that tongyi wanxiang has been fully upgraded and released a new video generation model. this ai video production tool is completely free, and "the app end is open for unlimited use every day."
behind alibaba cloud's series of actions, one phenomenon is that spending money to attract traffic and new customers is still the norm in the large-scale model industry.
"open source and price reduction are all based on the same logic, how to make the ecosystem develop." zhou jingren said in an interview with reporters. he said that since alibaba cloud firmly started to open source last year, alibaba cloud has seen the model ecosystem construction exceed expectations in the past year, and large models are no longer high and mighty. as of mid-september 2024, the cumulative downloads of tongyi qianwen open source models have exceeded 40 million.
zhou jingren said that every price reduction of alibaba cloud has gone through very serious internal discussions. in addition to costs, cloud vendors must weigh the development of the entire industry, developers, enterprise users' feedback and other aspects, and further reduce costs in the future through economies of scale, technological progress and resource scheduling. from a long-term development perspective, the capabilities of large models need to be affordable to everyone, thereby stimulating more industrial-level innovation. wu yongming also said in his speech that the cost of model inference is a key issue affecting the explosion of applications.
"there is a saying that today's ai is equivalent to the internet around 1996, that is, in the bbs era. at that time, internet access charges were very expensive. later, with the development of the internet, including the development of mobile internet, operators vigorously invested in infrastructure construction, and traffic charges came down." zhang qi, vice president of alibaba cloud, said that alibaba cloud is now also frantically developing large-scale ai infrastructure. only by lowering charges can it be possible to talk about the explosion of future applications. this long-term goal is what alibaba cloud considers more, "rather than having to do the math today and how much money can be made right away."
as the price reduction of large models erodes gross profit or even turns into negative gross profit, where is the bottom line of the price reduction of large models? in this regard, zhou jingren said that alibaba cloud's price reduction is mainly through technical optimization, not only the rapid iteration of the model itself, but also the model's reasoning efficiency and structural optimization are also being carried out simultaneously to promote the further reduction of model reasoning costs. alibaba cloud wants to pass on the dividends of technology to corporate customers to promote the development of the entire industry.
"we also recognize that the application of models today, including various innovations in models, is still in the early stages. if we put the reasoning of models into an expensive stage at this time, a large number of developers will not be able to use them effectively, in batches or on a large scale, which will to some extent affect everyone's attention (to agi changes)." zhou jingren said.
(this article comes from china business network)
