Written by / Cao Shuangtao
Editor / Yang Bocheng
On July 18, local time in the United States, Microsoft announced at the annual global partner conference that it will launch a new artificial intelligence subscription service. For the AI assistant Microsoft 365 Copilot, subscribers to Office 365 E3, E5, Business Standard and Business Premium will be charged $30 per person per month.
According to CNBC, if you add the subscription fee to copilot, the monthly price paid by TOB customers to Microsoft 365 may increase by more than 50%, or even as high as 83%. After the news was released, Microsoft's stock price has reached $366.78 per share, the highest price of Microsoft this year.
Source: Snowball
In fact, investors are optimistic about Microsoft large model charges, on the one hand, Microsoft office software for TOB customers is just needed, according to Microsoft's official website, Microsoft 365 copilot can provide similar to Excel data analysis, provide Word creativity, etc., which changes the way of working in the global "workers", there will also be many "workers" willing to pay, and then provide Microsoft revenue.
On the other hand, in response to the current enterprise use of a large model to worry about the risk of data breach, Microsoft emphasized that Microsoft 365 copilot is based on enterprise-level security, privacy, identity, compliance policies, enterprise data will be logically isolated and protected, and always within the control of the enterprise.
In addition to Microsoft 365 Copilot opening charges, according to Microsoft, Copilot's artificial intelligence chat tool, Bing Chat Enterprise, will also be built into a standalone product at $5 per person per month.
At the same time, Microsoft's Azure cloud computing platform and Windows system will support a large language model called Llama 2 owned by Facebook parent company Meta. One of the biggest highlights of LLAMA 2 is that it allows commercialization, which can be used commercially by any organization or individual developer.
In fact, whether it is Microsoft's large model to increase the company's revenue with subscription services, or "split" Open AI to choose Meta, this shows that Microsoft, like many large model companies at home and abroad, is accelerating the process of commercialization of large models. However, from the actual situation, the commercial exploration of large models on the TOB side and TOC side is still bumpy. The delay in commercialization also determines that many large models will be submerged in the long river of history in the future.
- TOC side: How to achieve a balance between daily activity and commercialization?
According to Similarweb data, the number of ChatGPT global visits and unique visitors fell by 9.7% and 5.7% respectively in June this year. It should be noted that Chat GPT user visits have entered an era of negative growth, and there have long been traces to follow.
From January to May this year, ChatGPT user visits increased by 131.6%, 62.5%, 55.8%, 12.6% and 2.8% respectively. The reason for the decline in ChatGPT user data is that other manufacturers have launched large models to divert ChatGPT users.
Source: Similarweb
But from a deeper perspective, whether it is the offline Internet celebrity store from a short "explosion" to a rapid desertion, or the previous online explosive APP new users to skyrocketing to plummeting, it shows that the user's momentary freshness comes and goes quickly.
According to Morgan Stanley, 19% of respondents have used ChatGPT, but only 4% said they still rely on it. However, this situation of ChatGPT is not an isolated case, and how to ensure the continuous addition of users and achieve high retention on the basis of continuously improving its own large model capabilities will be a common problem that needs to be solved for TOC large model enterprises.
Source: ChatGPT
However, the problems that need to be solved for the ToC large model do not stop there, and with the gradual enrichment of the capabilities of the large models of subsequent manufacturers, it will inevitably launch a large model of the mobile terminal like ChatGPT. After all, only by taking the user into their own hands can they lay the foundation for the commercialization of their own large models. At the same time, from the perspective of ChatGPT, users still define the large model as a tool APP similar to the browser.
From the current domestic mobile tool APP monetization channels, generally include advertising, membership value-added services, around the tool APP launched scene services to obtain commission income, e-commerce and so on. Although the monetization methods are diversified, the vast majority of domestic tool apps are either loss-making or making a small profit. Simply put, the difficulty of monetizing tool apps is also an indisputable fact in the industry.
Taking Ink Weather as an example, according to the prospectus previously released by Ink Weather, from January to September 2017, the operating income of Ink Weather was 44.73 million yuan, 127 million yuan, 210 million yuan and 223 million yuan, and the net profit was 1.929 million yuan, 24.99 million yuan, 20.53 million yuan and 47.29 million yuan respectively. Taking BOSS Zhiping as an example, its net profit attributable to the parent from 2019 to 2022 was -502 million yuan, -942 million yuan, -1.071 billion yuan and 107 million yuan, respectively.
Moreover, these monetization methods, put on the APP of large model enterprises, must also be subject to targeted adjustments. On the one hand, similar to the advertising business, it is necessary to ensure the conversion rate of customer advertisements without affecting the user experience, and how to place advertisements has naturally become the first problem to be solved by large model enterprises. If a large-model enterprise mobile app is like other apps, users need to watch 15 seconds of ads to unlock the output content, or increase the number of visits, which will inevitably be disgusted by users, resulting in continuous loss of users.
On the other hand, if you insert ads in content, in addition to controlling risks and reducing the impact on yourself, you must also ensure that the quality of Q&A is not affected in the process of implanting ads. But from Baidu's experience with mobile advertising, it seems that it is difficult to achieve a balance between the two. Therefore, how much revenue the advertising business can bring to the big model in the future remains to be seen.
From the perspective of member value-added services, the reason why iQIYI has been losing money for many years can achieve profits, in addition to iQIYI's internal initiative to reduce costs and increase efficiency, the platform launched "The World", "Police Honor", "Canglan Tips", "Qingqing Daily", "Crazy" and other popular films and televisions, while meeting users' pursuit of high-quality film and television content, but also driving iQIYI advertising, membership value-added services at the same time to improve, which while gaining market recognition, it also attracted capital participation and helped iQIYI complete several rounds of financing. This solved iQiyi's debt crisis.
Source: iQiyi
In addition to iQiyi, Tencent Music and NetEase Cloud Music both rely on differentiated music content to drive the growth of platform member revenue.
The enlightenment brought to large model enterprises is that if large model companies also want to rely on membership subscription revenue to drive the company's revenue growth like Microsoft, they must launch core differentiated functions that competitors cannot imitate, and continuously upgrade and improve functions to ensure the continuous payment of member users. But the question is that with the gradual convergence of the capabilities of the current manufacturers' large models, how many large model manufacturers can do this in the future?
- ToB side: The complexity is much higher than that of the ToC side
Perhaps some companies realized the difficulty of surrounding the ToC side large model at the beginning, and when their large model was launched, they focused on the ToB market.
Huawei's Pangu model shouted the slogan is to reinvent thousands of industries. Based on this, Huawei's large model has also launched large models similar to government affairs, finance, railway, meteorology, and other industry segments.
Source: Huawei Big Model official website
The main landing scenarios of the Tencent Cloud model are finance, customer service, and education. Open AI partnered with Khan Academy to develop custom chatbots to help students prepare for exam review materials and prevent students from using AI bots to cheat directly. Saleforce is already offering GPT-4-based custom chatbot businesses to industry customers.
In fact, the big model around the TOB end is essentially a competition for the delivery ability of the manufacturer. Whoever can serve the final delivery will be the first to win the advanced in the TOB large model market. However, from the current stage, manufacturers are also facing many thorny problems around the TOB large model to be solved.
First, there is a serious lack of talents within large model manufacturers. In order to solve the shortage of talents in large-model product managers, manufacturers have given higher salaries. The reason for this situation is that when manufacturers customize large-model products for downstream customers, they need to have a clearer insight into the pain points and needs of customers in order to ensure that the final products delivered to customers meet customer requirements.
Source: BOSS Zhipin
However, because the problems shown by different industries and different enterprises themselves have strong differences, which requires the product managers of manufacturers not only to have the awareness of large-model products, but also to have the expertise of the customer's industry.
Taking agriculture as an example, if manufacturers want to develop large-scale model products that meet the needs of agricultural customers, in addition to mastering basic agricultural knowledge, they must also go deep into the field to understand the growth habits of different crops, the psychological state of farmers, and so on. Obviously, product managers who can achieve the above requirements at the same time are estimated to be rare in the industry.
The second is the lack of training data. When customizing large model services for ToB customers, it is necessary to provide customers with internal data in addition to industry data. However, from the customer's point of view, the privatization deployment model of handing over all data/computing power/services to one company is equivalent to monopolizing the property rights/source code of user data and models, which may lead to too centralized artificial intelligence development.
At the same time, once the manufacturer has a data leak during the training process, the negative impact on the enterprise can naturally be imagined. However, the problem is that if the customer does not provide too much training data, this will inevitably lead to the large model training quality not meeting the customer's expectations, affecting the final customer acceptance and subsequent payment.
Third, customer development is difficult, conversion costs are high, and it is difficult for large model enterprises to form scale effects. On the one hand, the biggest feature of industries similar to finance, medical care, and high-precision industry is that the requirements for data and information are extremely high, and there can be no data deviation. However, due to the relatively poor ability of ChatGPT in mathematical calculation, some information cannot be updated in real time, it is easy to dissuade customers.
On the other hand, the overall profitability of domestic companies is relatively weak compared to that of the United States. Therefore, the willingness of domestic enterprises to pay for IT software is already low, which can also be confirmed by the proportion of our Sino-US software industry revenue to GDP. Under the background of the low willingness of domestic enterprises to pay, cost reduction and efficiency improvement have become the consensus of many domestic enterprises.
Source: Wind
In December 2022, Ma Huateng bluntly said that many of Tencent's internal businesses should be cut, don't linger, how big can you do this (non-core business)? What about getting bigger? Li Yanhong also pointed out in the open letter that the future development of business "should not only look at income, but also look at profits, but also look at the input-output ratio." In the "1+6+N" organizational change in March this year, Alibaba officially put forward the operational requirements that each business did not come from negative profit and loss.
Obviously, if a company wants to cooperate with a large model enterprise to develop a large internal model, its relatively high cost and how much real value it can bring to the enterprise, whether it can recover the cost and when to achieve profitability, this has naturally become the concern of many business owners. Under the concern, it will naturally cause large model companies to encounter the problem that the speed of customer expansion is not as fast as the market expects.
In terms of benchmarking the SaaS industry, the domestic SaaS industry started at the beginning of this century and has developed for more than 20 years. However, under the lack of willingness of customers to pay and the preference of customers to use customized SaaS, it is difficult for SaaS companies to form scale effects. Therefore, at present, domestic SaaS enterprises are generally based on small profits or losses.
Source: Financial reports of various companies
Compared with the SaaS industry, the delivery time of the customized version of the large model is longer, and the human and financial costs that enterprises need to invest are higher. Moreover, the penetration rate of large models on the enterprise side is still in its infancy, and it will take a long time to explore in the future. But I don't know, do the current large model companies have enough financial capacity to wait for the gradual increase in market penetration?
- Will a price war break out in the big model later?
In addition to the commercial exploration of ToB and ToC, ChatGPT is also exploring large-model app stores similar to the App Store. In May, Open AI opened the ChatGPT third-party plug-in store, allowing users to download third-party plug-ins directly from it and integrate them into their own ChatGPT to expand to more use cases.
According to The Information, Open AI CEO Sam Altman said in an internal meeting in June that he was planning to create a large model app store as a new attempt in the commercial field of large models. According to people familiar with the matter, the app store launched by Open AI allows customers to list custom large models on their own. Combined with the actual needs of other enterprises, a platform for customized sales.
However, from the actual situation, the effect of the app store created by Open AI has not met Open AI's expectations. At present, the download volume of some large plug-ins is only in the hundreds of thousands. Obviously, Open AI still has a long way to go in the future if it wants to rely on app stores for commercialization.
Source: Open AI
Although at present, domestic large-model manufacturers have not yet opened the charging model like Microsoft, but with the gradual retreat of the popularity of large models, the gradual decline in the stock prices of related companies, the gradual calming of investors and professional scholars and other practical factors, it is not excluded that subsequent large-model enterprises in order to recover the early R&D investment, stabilize the cash flow of enterprises, and gradually open the charging model.
At the same time, in order to complete the blockade of competitors and expand their market share, large-model companies may fight in the price on the ToC side and the ToB side. In addition, from the perspective of many domestic industries such as the coffee industry and the two-wheeled electric vehicle industry, price wars often lead to enterprises lacking financial capacity being liquidated, and the market share will gradually be concentrated on the head enterprises in the future.
Perhaps, as Professor Huang Tiejun said before, today's large models are an intermediate product of technology iteration, and with the development of the subsequent domestic large model industry, the number of large models that can survive in the future is about 3.
Epilogue:
It is difficult to implement commercialization in the short term, the high investment in research and development in the model training stage, and the possible price war in the subsequent market. In addition to this, how to solve the copyright/data issues involved between different large models, and more practical security/ethics issuesAnd follow-up policy supervision, where is it going? When the large model is concentrated on a few manufacturers, how to face the monopoly problem and so on.
The existence of the above problems makes it easy for none of the current large model manufacturers to be said. But thinking the other way around, with the gradual enrichment of subsequent large model capabilities, perhaps many industries will also be reconstructed. But I don't know how many big model companies are finally drowned in the ruthless competition in the market?