Recently, there have been rumors in the AI community that, against the backdrop of overall financing difficulties and pushing valuations higher, someLarge ModelStartupsBecome difficult, in and out of the dilemma, which is highly concerned about the "six little tigers" in a two tigers, and even have given up the basic large model research and development, turned to theAI ApplicationsBecause. Because, if not transformed I'm afraid that "six little tigers", will be cool.
If the Tigers really do this, it somehow means that they are starting to take the initiative to walk off the AGI poker table, shifting from the grand narrative of the stars and the sea to the reality of grounded tea, rice, oil and salt. There's no way around it. It's the same old saying.The situation is stronger than people, the current environment, make money to live is the biggest king.
01 Ecological mapping of large model startups
With the rapid development of AI, especially the rise of big models, more and more startups and old-school big manufacturers are squeezing into this track to get stuck, which not only pushes the progress of technology, but also shapes a new industrial pattern.K roughly divides the main force in this field into three categories:
1. Big Model "Six Tigers"
At the beginning, the domestic AI jianghu only "five little tigers" name, namelyZhipu AI,Dark Side of the Moon, minimax, and Baichuan Intelligence,Zero One Everything5 AI unicorn enterprises. Evaluated from various dimensions such as team background, technological advantages, financing ability, and market valuation, they are all in the first echelon of the industry. Until June this year, former Microsoft Global Vice President Jiang Daxin founded theStep StarJoin the fray, become the new darling of the capital market, a round of new financing valued at $ 2 billion, but also sent it to the throne of the "AI unicorn", "big model six small tigers" of the new pattern has been laid since then.
In the "six tigers", the wisdom of the spectrum AI is known as the most "openAI temperament" of the domestic AI companies, products include intelligent customer service and knowledge mapping and other applications, in the B-side of the market has a better application, there are already more than 1,000 enterprise customers.
Dark Side of the Moon, launched Kimi with excellent long text processing capabilities, and grabbed the lead in the fierce battle for users through social media and video platforms.
Minimax focuses on image recognition and computer vision technology, and has launched Conch AI, a productivity tool, and Starfield, an AI companion product, with the dream and ambition of creating a community similar to Xiaohongshu and B Station.
Baichuan Intelligence, playing a more differentiated approach, focusing more on the medical and health field of large model research and development, medical AI as the core business, and launched a number of AI medical products. Recently, it has also begun to make efforts in the C-end, based on the founder's deep background in search technology, launched a "better understanding of search" AI assistant products "hundred small should", worth looking forward to.
Zero One Everything was founded by Lee Kai-Fu, a senior big brother in the technology circle, and also actively lay out the application side, insisting on advocating strength in the To C market, and its subsidiary set up in Singapore has launched a number of consumer-oriented application products.
Although Step Star is a rising star in the tiger, but it launched the "Step Series" general large model, let the industry a bright spot, its Step-2 trillion parameters MoE language large model, is the first domestic release.
2. New AI forces
In addition to the "six tigers", the domestic AI entrepreneurial camp also emerged in the face of the wall of intelligence, the secret tower AI and deepseek and other development momentum of the rookies.
Facade Intelligence was founded by Li Dahai, CTO of Zhihu, and its product "Small Steel Cannon" series has achieved good results in the field of end-side AI, and also summarized the "model of the knowledge density is constantly improving, on average, doubled every 8 months" based on its own development experience. The law of the wall.
Secret Tower AI is also not bad, focusing on AI search, its products "Secret Tower Translation" and "Secret Tower Writing Cat", also have a good market performance, became a rising star backed by the Ant Group.
DeepSeek, on the other hand, is a model dedicated to the study of deep learning algorithms, with 67 billion parameters, is currently one of the largest deep learning models in China. Its performance has been benchmarked against the internationally recognized Llama2, and is fully open source.
3. Veteran technology giants
Of course, veteran tech giants such as Ali, Baidu, Tencent, Byte, Huawei, Xunfei, Shangtang, and Kunlun Wanwei are never absent from any potential new track. Ali's Tongyi Thousand Questions, Baidu's Wenxin Yiyin, Tencent's Mixed Assistant, Byte Jump's Doubao Big Model, Huawei's Pangu Big Model, as well as KU Xunfei's Starfire Big Model, Shangtang Technology's SenseChat Big Model, and Kunlun World Wide's Tiangong Big Model. Not only have they formed a deep accumulation and strong barriers in technology, but they have also further strengthened their competitiveness in the field of big models through their large user base and rich data resources. In addition, the advantages of these veteran big players in terms of talent and capital are also incomparable to those newcomers in technology who are still in the startup stage, which also makes their position in the AI field more difficult to shake.
02 Shift from "basic big models" to "AI applications":
AGI ideals are beautiful, but living is more important
1, basic large model training, very burning money
Yang Shilin, the founder of Dark Side of the Moon, was asked in an interview, "If the money you get today is the last money, how would you spend it?" When asked in an interview, "What would you do with the money if you got it today and it was your last?", he answered frankly:" I hope this never happens because we're going to need a lot more money in the future." Yang Zhilin said the big truth, basic big model training is really too much money.
According to Musk's estimates, GPT-5 training will require roughly 30-50,000 H100s, and the cost of the chip alone will be more than $700 million, or about 5 billion yuan. This is only the cost of training, but does not include other costs such as human salaries, data procurement, infrastructure construction.With such a high cost, only a few giant companies and startups with huge funding will dare to get involved and keep playing.For most of the "mediocre" companies, such financial pressure is undoubtedly unbearable.
This mode of "burning money" has also led to the increasingly obvious Matthew effect in the field of large models. Giant manufacturers with its strong financial strength, can continue to invest in research and development, and constantly improve the performance of the model; while startups are facing enormous financial pressure, it is difficult to have the courage and stamina, and the giants to start a competition with the strength of the obvious asymmetry. In front of the gradual solidification of the competitive pattern of large models and the insoluble "winner takes all", many small and medium-sized AI enterprises are forced to face the reality and turn their attention to the application of large models, seeking lower costs and faster returns. At this stage, living is much sexier than grand ideals.
2、C-end marketing, the cost is high
In addition to the high cost of basic model training, C-end marketing is also very costly. A little earlier, many startup AI companies, taking advantage of the AI trend and the public's curiosity about the freshness of AI products, can get a good conversion effect with a CPM (Cost Per Mille, cost per thousand displays) of about 5 dollars. However, with the entry of those small tigers and new forces who have financed big money, and the disappearance of the public's mystique of AI products, the C-end marketing costs have been continuously pulled up, and according to the industry's estimates, the average bid in the industry has been pushed up to between $10 and $15 at present.
However, this price level is still not the ceiling, the head forces represented by Six Little Tigers, while continuously launching C-end products, have also opened a crazy marketing mode. For example, the dark side of the moon in order to make its Kimi quickly seize the user's mind, the user conversion cost (CPA) on the B station even directly dry to about 30 yuan. Of course, the effect of this kind of costless marketing is also instantly visible, and Kimi's monthly visits also skyrocketed, once surpassing Wenxin Yiyan and becoming the top of the ChatBot track in terms of monthly activity and visits. Dark Side of the Moon's "money power" has also attracted other Tigers to follow suit. According to industry insidersThese big model makers have become the largest group of goldmasters at B, outside of Poundland.
In addition to online stream casting, the big model vendors offline burning money is also equally crazy, according to technology media reports, the subway has become a must-contend place for AI vendors to advertise. The subway advertising investment of companies such as Smart Spectrum has been raised to the level of a million to start. Gradually pushed up the cost of customer acquisition, has become AI enterprises in addition to technological breakthroughs in addition to another problem: others are investing in the flow of promotion, you do not follow the market can not be seized, and with the strength and no strength, the two ends of the problem, very embarrassing.
03 Big Models, Hard to Profit: Make a Friend, It's Really Not Profitable
1. AI Smile Curve, Tear Off the Shade
As regular readers know, K has talked specifically about the "AI Smile Curve".This curve is a good representation of the current status of profit distribution in the AI industry chain.In the entire AI industry value chain, upstream hardware manufacturing and downstream application software occupy the high profit zone at both ends of the smile curve, while the basic big model R&D in the middle is located at the low point of the curve's profit, facing the dilemma of huge investment but difficult to realize.
AI Smile Curve Mr.K@Technology Leadership
Hardware to Nvidia, for example, the entire AI market for GPUs a piece of hard to find, its market value is once pressured Apple, Microsoft. Huang Jen-Hsun and a host of other shareholders cashed out at a high level and made a lot of money. In terms of application, systems based on artificial intelligence, such as automatic driving and assisted medical care, have been more and more widely used, showing great potential for development and business opportunities. On the contrary, the big model, Baidu founder Robin Li has long been snappy.Don't roll big models, roll applications, AI big models without applications are worthless.We agreed to change the world together, why do you take the profits and leave only the big models to pay for the dream? Tear off the cloth of shame, the big model is only left with embarrassment and poverty.
2, price war, come early
At a time when pressure is mounting on big models to survive, the global price war between big factory giants is adding to the already difficult realizations and meager profits of AI companies.
Since May this year, the country has Ali, Baidu, byte, KU Xunfei and many other companies, to join the AI big model price reduction army, at the expense of entering the profit "centi era" to attract more users.Beanbags under the Byte umbrellaThe big model price reduction resulted in an inference input pricing of only $0.0008/thousand tokens. one of Tencent Cloud's main modelHybrid-lite modeling, the price is even from 0.008 yuan / thousand tokens, directly adjusted to free. Just past September, the price war continues to burn, theAliCloud three Tongyi thousand questions main model, with significant price reductions in the form of a jump sale. Among them.Qwen-Plus and Qwen-MaxThe input prices of 80% and 50% were dropped straight down; Qwen-Turbo was even dropped straight down by 85%, with prices as low as $0.3 for a million tokens.
Internationally, OpenAI launched OpenAI o1 mini version, although the performance is far worse than the GPT-4o, but the cost is greatly reduced 80%. other such as Anthropic, Character.AI, etc., are also actively adjusting the price of the model to attract more users. These early price wars, to the already unprofitable "six little tigers", brought more competitive pressure.Originally was living "while looking for money while burning money" hard life, and now just have a little "self-blood" ability, and then caught up with the price war, looking to earn the "three melon two dates "and even more nothing.
04 "Six Tigers", who can come out of the pre-dawn darkness
1. Spell valuation: a skinny camel is bigger than a horse
Although the overall large model is facing various difficulties, but also is not an iron plate, still shows a different degree of differentiation, to the "six little tigers" as an example, in the valuation of the temporary leading wisdom spectrum AI ($ 3 billion), the dark side of the moon ($ 3 billion), minimax ($ 2.5 billion). Temporarily lagging behind is zero one million things, for which a media said "at least 2 give up the basic large model research and development, there is zero one million things". Dr. Kai-Fu Lee hurriedly posted in his circle of friends to clarify that "Zero One Everything has been doing pre-training ......".
After all, business valuation is not simply a numbers game, but also reflects investors' judgment of the company's future prospects.A high valuation means that the company has stronger technical strength, broader market prospects and a better team, and is more likely to obtain subsequent financing.In this stage of striving for financial strength, striving for a long life, even if the days are hard, we all want to be the "thin camel", and no one wants to fall down first.
2. Look at the market share: whoever gets the users gets the world!
In addition to the valuation of the division, the "six tigers" in the number of users and product influence there are also significant differences. According to the "AI product list" ranking in September, "six tigers" in KIMI (the dark side of the moon), conch AI (minimax), Zhi Spectrum Qingyin (Zhi Spectrum AI), is located in the first echelon, respectively, ranked in the domestic list of the fourth, twelfth and fifteenth place. 15th place. And step over the stars (leap ask, bubbling duck), Baichuan intelligence, zero one million things, are not in the list of the top 50.
The user gets the world, this is since the traditional Internet era, the truth is unbreakable.The number of users is always the most critical indicator of product competitiveness. Therefore, whether by improving user experience, enriching application scenarios, or increasing marketing efforts, AI companies should take improving user conversion and retention rates as the top priority for company development. With more users, it is the same as having more chips to attract more investment, better fight against the industry's Matthew effect, achieve commercialization faster, and live with more peace of mind and autonomy.
Famous investor Zhu Xiaohu famously opined that the best outcome for Six Little Tigers would be to sell them to a major manufacturer. Is this really the best place for AI unicorns? Everything is still to be verified by time, but K has a small suggestion that he would like to share with entrepreneurs in the field of AI: last year, we could talk about technology, products, and scenarios; then this year, we'd better talk more about growth and earning money; otherwise, there won't necessarily be a next year.