Source | Bohu Finance (bohuFN)

Author | Kaikai

On the road to commercialization of large AI models, "ToB or ToC" has always been a dilemma. However, there is a consensus in the AI industry that startups are more likely to find opportunities on the C-side, while Internet giants are more likely to gain scale advantages through the B-side.

But for now, that consensus may be about to be broken. The dark side of the month, which has always been regarded as a "To Cist", recently announced that it will release an enterprise-level API, and the context cache storage fee of the Kimi open platform will be reduced by 50%, accelerating the development of the B-end market.

Coincidentally, OpenAI, the originator of large models, also recently announced that it will allow enterprises to customize its flagship AI model GPT-4o using their own data, which means that enterprises can customize and optimize AI models with stronger performance.

From the "100 model war" to the "application war", the large model has reached a critical period of commercialization, not only to consider whether the product is prominent, but also to comprehensively consider the cost, application, monetization and other issues, each company is looking for its own answer, for them, "ToB or ToC" may not be a choice question, but a mandatory question.

01 Monetization, the "roadblock" of the C-end large model

As early as before Kimi began to focus on the B-side, it had tested the C-end reward model on a small scale in May this year, and users could get different priority usage hours during peak periods by purchasing gifts ranging from 5.20 yuan to 399 yuan.

Regarding the launch of the tipping function, the dark side of the moon has said that the business is in the testing stage, and the company remains open to the exploration of the commercialization model.

From this point of view, Kimi's tipping model is more like a test of users' willingness to pay, rather than rushing to make a profit, after all, most of Kimi's user groups are converted from Douyin and Bilibili, and it is necessary for the company to explore the attitude of young people.

But this does not mean that Kimi is ready to "generate electricity for love" all the time, because the "speed of burning money" of large-scale model startups is too fast. First of all, if the large model wants to enter the C-end market, it must pay a lot of marketing costs.

After the Spring Festival this year, large model companies have started a marketing war, and the most common online streaming mode is CPA, that is, after the user browses the website to trigger the advertisement, as long as the registration or download the App, the large model company will pay the advertising fee.

However, at present, the platforms to choose from are nothing more than platforms with a concentration of young users such as Bilibili and Douyin, but there are many large model companies that need to invest in streaming, and the bidding model of CPA has also pushed up the cost of streaming in disguise.

Some industry insiders said that at the beginning of 2023, the CPA price of station B is generally less than 10 yuan/person, but the CPA cost of the dark side of the moon at station B may be as high as 30 yuan.

According to Sina Technology's estimates, since February this year, it has invested at least more than 30 million for the stream. The effect of streaming is significant, and Kimi's visits have increased by more than 4 times, but the excessive marketing costs have burned investors' money after all.

The second is the cost of training. OpenAI initially planned to spend about $800 million on training costs this year, but could double the cost as OpenAI ramps up training of its latest flagship model.

Dario, CEO of OpenAI's number one competitor Anthropic, also said that the training cost of the AI model currently under development by the company is as high as $1 billion, but the training cost could rise to $10 billion, or even $100 billion, by 2027.

Finally, there is the cost of computing power. Guosheng Securities has estimated that to build a large model that benchmarks ChatGPT, the price of a single A100 chip is 100,000 yuan, and an investment of 1 billion yuan is the admission ticket.

As costs continue to rise, the "AI Five" may not be short of money after receiving new financing, but it cannot always ignore commercialization, let alone other large model companies that have no surplus in their hands.

It's just that if you want to monetize in the C-end market, it's not so easy. One is that most of the current general large models, such as Wenxin Yiyan and ChatGPT, are free models, and it will obviously take a long period to cultivate users' payment habits, and homogeneous AI applications are far from reaching the stage of rigid demand.

The second is that the revenue model of the large model in the C-end market is relatively simple, and in addition to the subscription fee, there are still many difficulties in other charging models, such as relying on advertising to generate revenue, which may affect the user experience and involve privacy issues; Previously, the office software WPS charged for AI functions, and it rushed to the hot search because of "nesting dolls".

Third, C-end consumers recognize that products are more based on brands, which is also the reason why the dark side of the moon has to spend a lot of money to invest in the stream, in this context, in case Ali and Tencent smash an AI application, it will be difficult for other startups to parry.

Even OpenAI, whose revenue in the C-end market has reached $1.9 billion, mostly from 7.7 million subscribers worldwide, who pay $20 per month for ChatGPT Plus each, is still far from enough to cover the cost of building and running the model.

Foreign media quoted OpenAI's undisclosed internal financial data, saying that OpenAI will face a loss of up to $5 billion this year, of which the annual revenue is estimated to be between $3.5 billion ~ $4.5 billion, but the operating costs may reach $8.5 billion, of which the inference cost is $4 billion.

OpenAI offers a free version of ChatGPT to C-end users, which allows the company to increase the cost of reasoning without bringing in any additional revenue, which is one of the reasons why it can't make ends meet.

02 Bet, the end of the C-end is the B-side?

To put it simply, in order for the large model to develop rapidly in the C-end market, it is necessary to achieve better comprehensive capabilities and use effects, and through low-cost or free use models, it can better attract users.

This model needs to continue to "burn money", and the top players of the large model have to set their sights on the B-end market. Taking Baidu as an example, it recently released its second quarter report for 2024, in which Baidu's intelligent cloud business revenue reached 5.1 billion yuan, a year-on-year increase of 14%, and continued to achieve profitability, and the proportion of revenue contributed by AI has increased from 6.9% in the previous quarter to 9%.

However, judging from Baidu's 2023 annual report, although Baidu has achieved hundreds of millions of yuan in incremental revenue through the advertising system reconstructed by the Wenxin model, its online marketing revenue is still declining compared with the previous two quarters, which shows that the activity and realization rate of the large model in the C-end application have not met expectations.

Therefore, whether the large model should be ToB or ToC, the big guys also have different opinions. Wang Xiaochuan, CEO of Baichuan Intelligence, has made it clear that To C is ten times that of the To B market, and large manufacturers will roll to B, (Baichuan Intelligence) must do differentiation.

Zhu Xiaohu of GSR Ventures believes that the business model of To B is far more suitable than the model of To C in China.

Although each has its own point of view, judging from the actual actions of large model companies, they do not have too much entanglement about To C or To B, not only "all wants", but empowering each other.

Zhang Yaqin, dean of the Intelligent Industry Research Institute of Tsinghua University, once said that at the application and service level, the cycle of To B is relatively long, while the application products of To C can be launched quickly, which is basically consistent with the development path of the mobile Internet.

Therefore, most large-scale model startups adopt the strategy of To B and To C in parallel, and even Baichuan Intelligence, which frankly says "we do the C-side", has also launched an API interface business.

This "C+B" business model is also the mainstream business model of large model companies. For example, OpenAI not only collects ChatGPT membership fees on the C side, but also charges large model API calls through the "public cloud + API" method on the B side.

In addition to "monetization" through fees, large model companies will also use AI to support their existing mature businesses. For example, Alibaba's Quark browser recently released the PC side, upgrading a series of "all-scene AI" functions such as AI search, AI writing, etc., to further increase the attractiveness and stickiness of customers; SenseTime, which has always been mainly on the B-side, will also be introduced to the C-end market this year and released the AIGC product "Second Painting Fun Shot", which can generate creative photos.

Therefore, in the choice of the commercialization path of the large model, betting on ToC and ToB at the same time is not only to make up for the "uncertainty" of the C-end with the "stable income" of the B-side, but also to empower each other in terms of technology and brand effect of the two businesses in the long run.

On the one hand, through the C-end product, the large model company can continue to collect user feedback, accumulate the application practice of the model, and finally feed back to the large model to achieve iterative upgrades.

On the other hand, as Kai-Fu Lee, CEO of 010000 Things, said, there are more opportunities in the short term in China's To C, C-end products are more likely to explode and gain word of mouth, and their traffic and potential energy can also be fed back to the B-end business.

Finally, large model companies that "do C-end" are also actively launching API reduction services, and even continuously reducing call costs, hoping that developers can develop easy-to-use AI applications in their own ecosystems.

Just like the Internet industry in the past, when the degree of differentiation between products is not large, the fight is whose ecology can be the first to "blossom", and more creators can participate in the opportunity to develop easy-to-use AI applications.

Therefore, the current debate about whether ToC or ToB is better is actually not of great significance, because the main contradiction in the current large-scale model industry is not just financial pressure, but the need for more people to apply, so as to create an ecology, and this goal cannot be easily achieved by relying on the B-end or C-end market alone.

According to the Cyberspace Administration of China, as of August, there are more than 190 generative artificial intelligence service models that can provide services in China, but Kai-Fu Lee has predicted that when the competition for large models is nearing the end, there may be only 30 large model companies that can survive.

At present, there is no recognized "leader" in the large model industry, including Internet manufacturers, and they may not be able to take advantage, for start-ups, no matter what path they take, they must compete for brands, products and ecology, otherwise there will be no chance in the future.

03 ToB, a fierce battle in the era of customization

It can be seen that the B-end is an indispensable part of the closed loop of commercialization of large model companies, but it is not so easy to kill the way in the B-end market.

First of all, there is the endless price war in the B-end market. In May this year, Byte officially released the bean bag model, and the pricing of the main model in the enterprise market is only 0.0008 yuan/1000Tokens, which is more than 9% cheaper than the industry, so that the token price of the domestic large model has developed from "cent-based" to "cent-based", which shocked the industry, and many competitors have followed up to reduce prices.

It is true that most large model companies believe that the development of the industry should avoid price wars, but when it comes to the battlefield, they will be squeezed out by their opponents if they do not reduce prices. Even though Baidu Robin Li once called on entrepreneurs to "roll" AI applications, rather than volume prices, after the price of bytes and Alibaba Cloud was greatly reduced, Baidu's two main large models were soon announced to be fully free.

Industry insiders believe that there are not many companies willing to pay for software, the profit margin of the B-end of the large model is shrinking rapidly, and the large model project that can be sold to 10 million yuan last year may only be sold for 1 million yuan this year, and there are too many open source large models on the market that can be shelled, and the competition is very fierce.

Secondly, the B-end and G-end businesses are sometimes difficult to bite. First, every B-end customization case is a "lonely book", and non-standardized customized services mean higher cost investment, especially for the working environment with complex scenarios and late digital transformation, data security, information islands and other problems are not easy to break through.

Second, the sales cycle and accounts receivable period of the B-end business are often long, which requires the company to show more patience and continuous efforts. Recently, Liu Qingfeng, chairman of iFLYTEK, said that the company will take the initiative to slow down or even give up part of the G-end business, because of the problem of the payment cycle.

Third, for the cooperation of some vertical industries such as finance, medical care, and law, there is a very high demand for the intervention of large models, and it is also a big cost to reserve high-quality talents in related fields.

Finally, even if large-scale start-ups overcome these difficulties, enterprise development infrastructure models often favor established cloud vendors. The data shows that in the large model market in 2023, the market share of Baidu Intelligent Cloud, SenseTime, and Zhipu AI will rank among the top three, accounting for half of the market together.

However, this does not mean that start-ups cannot break through, data shows that in 2023, the enterprise adoption rate of generative AI in China has reached 15%, and the market size is as high as 14.4 trillion yuan, and this number is expected to continue to increase.

In the second half of the B-end market, the large model may need to be upgraded from corporate vision to industrial vision, rooted in office, production, education, manufacturing and other links, and become the driving force for new growth of enterprises, which is the key for enterprises to accept this new technology.

Some business people believe that they don't care about how much work the large model has, the key is how much money can be saved for the enterprise, and what they really need is a complete business solution based on the overall technology improvement.

With the continuous development of the large-scale model industry, the question of how to find a sustainable business model in user growth and model capability progress will still plague every large-scale model enterprise, but the answer may be different for different enterprises.

This means that in the field of large models, there is no set of models that can be used by all enterprises, whether it is ToB or ToC, it is just the path to the future of large models, and what really determines the future of enterprises is whether they can bring more innovative AI applications and services in the process of serving customers.

If in the end, only 30 large models can be left, then the surviving large models are not necessarily the most famous, but they must be the most practical.

*The copyright of the first image and accompanying picture of the article belongs to the copyright owner. If the copyright owner believes that his work is not suitable for everyone to browse or should not be used for free, please contact us in time, and the platform will correct it immediately.

This article is from Xinzhi self-media and does not represent the views and positions of Business Xinzhi.If there is any suspicion of infringement, please contact the administrator of the Business News Platform.Contact: system@shangyexinzhi.com