news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

introduction:

since 2024, large model manufacturers have started a money-burning war on major platforms. large model advertisements are everywhere, from online to even subway and airport billboards. behind the cost of hundreds of millions of dollars, retention is still a difficult problem for major large model manufacturers.

produced by phoenix technology's "eye of the storm"

author: danny

editor: dong yuqing

since the beginning of this year, advertisements related to large models have suddenly sprung up in people's sight - from information flow advertisements on bilibili, xiaohongshu, and douyin, to billboards in subways, office buildings, and airports, and even in new york's times square.

picture｜from the internet

how crazy is the money-burning war in the big model?

online, taking bilibili as an example, in the second quarter of this year, the number of ai industry customers increased by more than 200% year-on-year from the second quarter of last year. with a high customer base, the second quarter still saw double-digit growth compared to the first quarter.

to some extent, large model manufacturers are also an important increase in advertising for bilibili in the first half of this year. according to disclosures, the number of advertisers at bilibili increased by 50% year-on-year, and the coverage rate of new entrants ai manufacturers exceeded 90%.

leading ai manufacturers such as hoshino, dark side of the moon, kimi, zhipu qingyan, doubao and common ai products on the market have all entered the market.

phoenix technology learned from an insider thatthe total annual investment of these manufacturers is conservatively estimated to be over 100 million yuan.the marketing war in the ai ​​big model industry is becoming more and more intense.

earlier media reports said that the average cost per user of dark side of the moon on bilibili was rmb 30 under the cpa billing model (cpa: cost per action, also known as the cost per conversion).

in addition, a person involved in subway advertising also told phoenix.com technology thatthe advertising costs of zhipu in some subways start from millions.

the battle for big model advertising with lots of money has entered a fever pitch.

kimi is the biggest spender?

according to phoenix technology, in the first half of the year, the delivery of ai big models in various channels was mainly based on effectiveness, focusing on user conversion volume. simply put, it was to attract more new users for the big model products.

in the eyes of ordinary people, these ads often appear with download links. for example, on the homepage and video recommendation area of ​​​​bilibili, it is easy to see the promotion link of "kimi ai smart assistant". the dark side of the moon behind kimi is a large model company with a valuation of us$3 billion. the main feature of its products is the processing of long texts.

picture｜from the internet

"for every registered user kimi gets through bilibili, dark side of the moon has to pay at least 30 yuan." recently, some media reported that after kimi launched bilibili, other ai companies could no longer afford to invest. however, our interview revealed that this was not the case.

"the overall bid price in the industry is not as high as expected." in this regard, a relevant insider told phoenix.com technology that the current average bid price in the industry is between 10 and 15 yuan."currently, the ai ​​customer effect delivery optimization targets of bilibili are mostly activation bids, and the average bid for this category is around 12 yuan."

it is worth mentioning that dark side of the moon is not the ai ​​company with the highest investment amount on bilibili, but in the perception of users and the industry, the one that spends the most money is kimi developed by dark side of the moon.

according to industry insiders, on the one hand, it is because it not only puts out performance ads, but also adopts a strategy of putting out both brand and performance ads. on the other hand, kimi entered the b station earlier and found the right target group, so the user side's mental perception is stronger. "it can be said that kimi's success does not come entirely from the high bids for performance ads, but from finding the right track earlier, and the dual construction of mind and effect is indispensable."

secondly, although both are in the ai ​​industry, kimi is in the long text tool track. there are also different tracks such as ai social, ai video, ai painting, ai office, etc., which means that the target groups and areas of investment are not consistent.

finally, this is also affected by the specific delivery strategies of advertisers. some manufacturers choose softer brand placement and do not strongly display the brand name in the cover title, which makes user perception not intuitive.

in addition to bilibili, xiaohongshu is also a major traffic source that ai manufacturers are trying to develop. as of september this year, kimi, xingye, zhipu qingyan, wps ai and other products have also reached different degrees of cooperation with xiaohongshu.

however, compared with bilibili, the industry generally believes that xiaohongshu's current ai content commercialization atmosphere is relatively poor. the reason is that the front-line sales responsible for the ai ​​industry are not professional enough, cannot impress customers, and lack service capabilities.

the preservation that cannot be burned

although the money-burning war is fierce, a difficult problem facing large-model manufacturers is: low user retention rate. downloading - trying out the new product - gathering dust is the most common dilemma encountered by large-model products.

people often download an app out of curiosity about a certain usage or function, but often quit because it’s not that easy to use or they don’t know how else to use it.

xue su, vice president of baidu and head of ai innovation business, told phoenix technology:"in the early days, many native products neglected insights into user value. i found that some users did not understand how powerful our ai products were after using them. it was like using the old search function. when the new search function was placed in front of them, they did not even know it."

"among the operational indicators of ai products, retention is much more important than conversion. over the past year or so, too many names that have been screen-sweeping have died out after a brief period of popularity. users come quickly, but leave even faster," said lan xi, a well-known commentator in the technology industry.

finding an accurate and economical traffic diversion model is one of the solutions currently considered by major model manufacturers.

the reason why kimi chose bilibili as his main platform is that he saw its coverage of young, educated people in first-tier cities (especially college students), and regarded this group as his seed users. this group is also the customer group that major manufacturers are most optimistic about.

based on foreign experience, the core demands of this user group of highly educated young people are often matched with the most basic applications of large models - completing information aggregation and screening through the form of questions and answers.

different from foreign markets,chinese college students who are accustomed to recommendation algorithms do not have a strong demand for information aggregation. specific application scenarios and application tutorials are the primary requirements when using ai tools, such as ppt production, paper writing, translation, etc.

this can be verified from xiaohongshu, which is known for its search:currently, ai tutorial related content is the best technology content consumed on xiaohongshu.

picture｜screenshot from xiaohongshu

at the same time, ai big model companies are essentially platforms. in addition to the official app, they also need to build momentum for their own big models and attract ai industry practitioners to develop more application scenarios based on their own models. the richness of a big model's application scenarios also determines user retention.

in short, if an ai app wants to fundamentally improve retention, general model capabilities, specific application scenarios and tutorials, and a good developer ecosystem are all indispensable.

therefore, domestic ai marketing seems to have encountered a dilemma of creating an ecosystem from scratch:

it is necessary to demonstrate the power of its own model through general scenarios such as question and answer, and guide users to download it.

we also need to target segmented scenarios and constantly collaborate with kol (kol: industry opinion leaders) to produce segmented scenario tutorials to continuously educate c-end users and tell them their actual application scenarios and uses.

what's more, some companies will launch official teaching materials to popularize the development tutorials of their own large models to developers.

behind the marketing war, what each company can really use to gain retention is still the ecosystem.

the big model manufacturer that can't stop

the upstream has to snatch scarce r&d personnel, and the downstream has to educate and attract users. the large model manufacturers stuck in the middle have to become platforms in the mobile internet era. the war between platforms has always been plain and simple - burning money on marketing.

looking back at the platform "war" in the history of china's mobile internet development, from group buying, e-commerce, online travel, online car-hailing, to food delivery, shared bicycles, new retail, new consumption and "10 billion subsidies", it can be called a history of money-burning marketing.

looking at the winners of every "war" in the past, it is actually difficult to tell who is better in terms of technology or marketing methods. what really determines the winner is capital, who has more money and can burn it longer to truly cultivate users' consumption habits.

however, large-scale model products are fundamentally different from the products and services in the mobile internet era.the penetration rate of ai among current internet users is really too low.

this means that user education cannot adopt the simple and crude model of hard advertising to attract new users.

a practitioner in the large model track told phoenix.com technology,"the vast majority of users will not use ai products at all, so there is little point in investing in hard advertising for your products. they still have to be put into application scenarios."

another person in charge of industry investment told phoenix technology:“some products with high activity levels don’t use hard advertising. instead, they focus more on the features themselves. if a new feature is useful, users will naturally use it.”

large-scale model products urgently need to find a healthy development model to break away from their dependence on capital.

since 2024, the concept of ai in the capital market has become less popular, and even nvidia's stock price has begun to plummet. there are even rumors that alibaba's top management is re-examining its investment in dark side of the moon.

zhu xiaohu said bluntly in a recent interview: "i feel quite pessimistic. in the united states, mergers and acquisitions are essentially about poaching talent and letting investors exit with principal plus interest. to be honest, the situation in china may be even worse than this."

zhu xiaohu said that the best outcome for the "big model six tigers" is to sell them to big companies. the "big model six tigers" refer to the six big model startups: zhipu ai, baichuan intelligence, zero one everything, dark side of the moon, minimax, and step star.

figure｜zhu xiaohu, managing director of jinshajiang venture capital (from the internet)

however, domestic large companies are not very enthusiastic about acquiring large model startups. a cto of a well-known industry company once expressed to phoenix technology,"large companies have no need to buy large-model startups, unless they take over the gpus from the startups."

according to phoenix technology, considering that the capital situation this year is not optimistic, some of the six little tigers have already adjusted their scale, "either laying off some teams or stopping recruitment."