GitHub

Wen Zhu Si Jia

I'm the editor.

In early May, the news that bean buns would launch a fee-paying subscription service triggered extensive market discussion. Its updated fee-buying programme in Apple App Store shows that bean buns will introduce a four-bed fee: base, standard, enhanced, professional. The corresponding monthly rates are: free, 68, 200 and 500; and the annual rates are: free, 688, 2048 and 5088.

An official response follows: “The bean bag always provides free services. On the basis of free services, we are also exploring the introduction of additional value added elements to meet the differentiated needs of different users.”

According to the unique knowledge of 36 kryptonites, these are only preheating actions for the commercialization of bean buns. During the next quarter, bean buns will continue to drive commercialization. According to the source, the bean bag is expected to be officially on-line in late June and to update the functionality at a parallel Force conference. This time node was selected because the PC end and the mobile end still needed about a month to complete the adaptation of the basic function to the fee system.

According to Krypton, if progress is successful, the bean buns will be able to further refine the payment scene in three quarters in conjunction with the electric powerer ' s functionality and lead to a four-quarter run-off of the Mall through subsidies. These actions are intended to provide for commercial returns in 2027 and beyond. Thus, the penetration of paid users will not be used as an indicator for the 2026 bean bag.

In October 2025, the bean buns were reached by the initial and dithered acoustic electric operators. Users can access commodity information and shopping link recommendations through dialogue with bean buns. According to the above-mentioned sources, the byte tested the soybean bag ' s hard-spread capacity on a small scale at that time, but the Quantification was subsequently suspended due to user experience problems. At present, the bean bag has adopted a priority response to user questions, followed by a recommended model of interaction in the answer.

According to the above-mentioned sources, since April 2026, the acceptance of the AI shopping recommendations has increased from the second half of the previous year, with a conversion rate of over 3 per cent by clicking on the card and reading the detailed pages. On the other hand, the current growth of users of bean buns has slowed. According to 39 kryptons, the slowdown is partly a result of the byte — the expensive cost of computational reasoning, which has led to a reduction in the roll-out of bean buns in the new year.

The conventional rhythm of mobile Internet products is the burning of money-ring users before commercialization, but the commercialization of AI products is clearly more urgent. Bytes are no exception.

01 Subscription is a good choice?

According to Bloomberg on 27 May, byte beats are considering capital spending of up to $70 billion to advance AI development. Just less than a month ago, the South China Morning News quoted an informed report on 9 May that the byte had revised its AI infrastructure budget upwards from RMB 160 billion last year to RMB 200 billion (approximately $30 billion). The continuous addition of inputs also shows that bytes are betting AI at unprecedented levels.

The global AI arms race is still warming, but how to make money has become a real issue for all companies.

Previously, 36 kryptons had reported that, although the byte bytes had increased by nearly 50 per cent in 2025, the overall net profit was reduced by more than 70 per cent as a result of a significant investment in AI operations (officially clarified that the decrease was mainly due to IAS). As huge financial resources continue to be invested, it is difficult for global science and technology companies to stay at the stage of talking about technological ideals, not about business returns.

The three routes most commonly mentioned in the current market for the commercialization of AI are subscriptions, advertising realizations and To B model services. Of these, subscriptions are the first commercial scene to be accessed.

Looking back at overseas markets, head AI products actually started to explore payment patterns very early. After coming online in November 2022,ChatGPTIt took just about two months to launch $20 a month in February 2023.GPT Plus subscription service. Then,AnthropicIt was launched in September 2023.Claude. Pro, also priced $20 per month. By contrast, the country's Chatbot products have remained at a free-of-charge stage for a long time and have continued the competitive idea of “burn subsidies to capture the number of users”.

As the country ' s first plan to shift to the subscription of Chatbot products, bean buns do have a deeper user base. According to the global master list of the AICPB product list, the number of monthly active users at the end of the bean bag in April 2026 was about 336 million, an increase of 1.51 per cent in the ring, second only to ChatGPT, the second highest in the world.

The problem is that bean buns are able to reach the head quickly in the Chatbot product, not by the intelligence of the model itself, but by the ease of use at the product level. Added to the strong delivery capacity of the tremors, the bean buns took the lead in taking over user minds in the sink market.

In fact, domestic users are not too loyal to a single AI product. According to LateLatePost, the current daily average of single users of bean bags is 10 minutes. In the absence of significant technical barriers and where competitors remain free, open subscriptions can result in loss of users.

However, leaving out the product itself, subscription itself is not the best path for the commercialization of AI.

Even in Western markets, where there is a greater willingness to pay, the yield of subscriptions remains limited. In order to increase the number of paid users, ChatGPT launched its cheaper ChatGPT in 2026 Go.It's only $8 a month. This increased its number of paid users from 47 million at the end of last year to 55 million in the first quarter. However, according to The Information, the average number of active weekly users in ChatGPT in the first quarter was about 905 million. In other words, even the largest MAU-scale Chatbot product in the world, after three years of subscription, the rate of conversion to pay is only 6.1 per cent.

The difficulty in increasing the number of payments lies in the inherent dilemma of the AI subscription model.

On the one hand, in a context of intense competition and high product homogenization, no model company dared to completely eliminate free services. For the majority of light users, the free version is sufficient to meet their basic needs for day-to-day questions and answers, as well as for information searches. On the other hand, at a time when Token is still expensive, if AI is required to perform more complex tasks, the current maximum price to be paid is $200 per month (approximately RMB 1,400). Those who are willing to pay long-term for this are the few.

Even more crucial is that AI may not make money even at such high prices. In June 2025, Sam Altman publicly stated on social networking platform X that he would not be able to find a solution to the problem.OpenAIWe lost money on a $200/month subscription to ChatGPT Pro. This leads to a paradox — too high a price for users to pay; too low a price to cover the cost of computing.

Therefore, C-end subscriptions are now more than “loss of profits”. For real business returns, major technology companies still need to explore more commercial scenarios. For example, in February of this year, OpenAI began to step up and promote advertising in ChatGPT.

02 How the AI commercialization landed

After long encounters, now it seems that the technology giants have finally found the best path to commercialization — the ability to sell models to businesses. Businesses don't have to train themselves. Just pass.APIAccess allows direct access to AI modelling capabilities developed by technology companies and pays for token usage.

Compared to individual users with limited willingness to pay, enterprises are well positioned and willing to use these AI tools in an attempt to promote business efficiency within the organization. More importantly, in the To B model, enterprises usually pay on a continuous basis, based on Token's call, rather than on a month-by-month subscription (as noted above, monthly subscriptions for ordinary users usually do not cover the cost of energy). For model manufacturers, this is clearly a good deal that can be easily earned.

Anthropic ' s profit growth is amply illustrative of this. According to the Wall Street Journal on 21 May, income from the second quarter of Anthropic is expected to more than double to $10.9 billion, with the potential to achieve $559 million in operating profits. Within the industry, which is still in the phase of large-scale money-burning subsidies, Anthropic is the first large, near-profit model company. The main source of income is the API call of enterprises and start-ups. As of October 2025, Anthropic had over 300,000 business clients, about 80 per cent of total income.

Large Internet factories in the country have been successively organized and active in ToB markets. Unlike other AI start-ups like Anthropic, large Internet factories have cloud-capable computing capabilities: On the one hand, the realization of the model API, which encapsulates the model as a form of external output of cloud services, is the Maas (model or service); on the other hand, the use of the model may lead to the growth of its own cloud operations.

In this new AI cloud competition, the volcanic engines under the byte flag are growing fastest. On 7 May, data published by the IDC of the International Data Corporation (IDC) show that in 2025, the number of large models used by Chinese enterprises on public clouds increased by 16 times over last year. Nearly half of these calls took place on volcanic engine platforms, with a market share of 49.5 per cent, ranked first in the industry.

Ali also launched the "Physical Battle" in late November 2025, in an attempt to catch up with the market share on the AI cloud. According to Ali's Q4 fiscal year 2026 financial data, the number of clients of Ali's “Platform” (MaaS platform in Aliyun) increased eight times over the same period. This also indicates that Ali is rapidly re-establishing its AI To B market capacity.

In addition to the visible gains from the To B market, the byte is also actively using AI capacity to empower its own home-based operations.

In the case of byte-based A.I. model 2.0, the production of video has given further impetus to the creation of short play markets. Short opera industry sources have told 36 kryptons that as the video creation threshold continues to decline, there is an influx of new players in the industry. They come from different backgrounds, from games, tides and even from traditional manufacturing. In addition, the reduction in the cost of single short play production has allowed production companies to produce more dramas with the same overall budget.

The surge in the number of creators and the reduction in production costs have resulted in double push-up of the supply of the short play. At the moment, the way the short play gets to the core of the play is still the flow of the shivering side. According to a short-timer, the current peak of the AI short-time drama within the byte mega-engine system is approximately 120 million yuan.

The founder of the book of short play technology, Wang Xiaomei, told 36 kryptons that his company spent only hundreds of thousands of dollars a day last year, and that this year it has grown by about 10 times. At the same time, the short play product is usually based on the ROI on a large engine, which is only 1.03 to 1.07. This means that for every $1 million in advertising, there is a profit of around $30,000 to $70,000.

AI+ Electrician shopping may be the next byte dynamic scenario. On 30 March 2026, it was reported in the Interface News that a greyscale test was being carried out on the purchase payment function of bean buns. This feature supports the user in the soybean bag, which is directly listed and paid for, and does not need to jump to the tremor mall.

Bean buns are not the first A.I. product to try to get through the A.I. On 11 May this year, Ali thousands asked for official access to the treasure shop and the Al shopping facility. In September of last year, ChatGPT also launched the Instant Checkout feature, which attempts to direct users directly to the AI dialogue box to complete the orders and payments. But in March of this year, OpenAI cancelled this function. The reason for this is that it is easy to access the technology, but it is difficult to access the entire electrical ecosystem behind it.

The advantage of bean buns and questions lies in the fact that they stand behind two large, mature electric power systems, namely, shivering malls and poaching. Both commodity banks, commercial ecology, payment and compliance capabilities already exist. It also makes the Ai shopping more theoretically possible.

But even so, it still has to face two core tortures. First, is the A.A. Shopping itself a real need? During the spring festival, a $3 billion red bag was spent, and users were asked to use the Ai tea to try to develop consumer habits to complete consumption decisions through AI, as reported by 36 kryptons. However, as a result, most users do not have long-term retention. Shopping itself is a more complex and more dependent scene than take-out.

Second, will consumers really trust AI's recommendations? With the increased importance of AI search and referral capabilities, some businesses have begun to influence the capture and sequencing of information by large models, for example, through GEO. The GEO industrial chain, which was exposed at the 315th evening, also generated extensive discussion. It remains to be seen whether the AI recommendations will be seen by consumers as objective recommendations or as new advertising outlets when they begin to be influenced by commercial placements and interests.

Zhou Yu has also contributed to this paper