Charging is DeepSeek's "rite of passage"

The list of "financial backers" for DeepSeek's first financing is gradually coming to light.

On June 3rd, media reports stated that DeepSeek plans to raise approximately 50 billion yuan, with a post - investment valuation of 350 to 400 billion yuan. Compared with the rumored 70 billion yuan at the end of last month, the latest reported financing scale has slightly shrunk.

In addition to DeepSeek founder Liang Wenfeng's commitment to invest 20 billion yuan, Tencent is considering investing 10 billion yuan, and CATL may also invest 5 billion yuan. Based on multiple sources of information, other potential investors include the National Integrated Circuit Industry Investment Fund, NetEase, JD.com, Lisi Capital, IDG Capital, and Cornerstone Capital.

DeepSeek is the only major domestic AI company that has not yet raised funds. Once this capital is in place, its financial pressure will be significantly relieved. However, compared with Liang Wenfeng's ideals and the public's expectations for DeepSeek, this amount of money is obviously far from enough.

No matter how large a company is, it lacks funds for AI development. Doubao has already decided to charge fees, and DeepSeek might learn from it.

In early May, Doubao updated its App Store page, announcing that it would start charging fees and setting three levels of subscription prices, with the lowest being 68 yuan per month.

On the evening of June 3rd, Doubao announced that it plans to launch a professional version, which will include multiple professional services such as software development. The functions that users use daily, including search and Q&A, writing and image generation, as well as voice and video conversations, will remain free of charge. The professional - version services will also be free within a certain limit.

It is not surprising that Doubao App has started charging fees.

AI apps burn a large amount of money but generate very little income, and it is a long way off for them to achieve positive cash flow. This is a problem faced by all AI companies. Doubao, with 345 million MAU and 145 million DAU, cannot escape this industry reality either.

The deeper problem is that the marginal effect of Internet products does not apply to AI products. They cannot, like Internet products, turn the corner and achieve profitability immediately after the user volume crosses a certain threshold. On the contrary, the larger the user volume of an AI app, the higher the token usage, and the more severe the losses.

Making money has become the top priority for AI apps. Directly charging membership fees is one of the simplest, most direct, and most effective means.

Abroad, AI apps such as ChatGPT, Gemini, Claude, and Grok have already established a complete set of paid membership systems, and at the same time, they have imposed more and more restrictions on "free - riding" users, which is similar to "forcing users to pay". In China, major AI apps are more decent in their approach, but the underlying gameplay is the same.

Now, Doubao, the most successful domestic AI app, has also put membership fee collection on the agenda.

However, compared with its peers, Doubao has the support of ByteDance, a cash - cow company. It can obtain financial support from the group. At the same time, ByteDance has repeatedly raised the annual revenue target for its Volcengine MaaS business, reaching 15 billion yuan in April this year, an increase of 5 billion yuan compared with the end of last year, which reflects ByteDance's strong overall profitability in the AI field.

Perhaps it is DeepSeek, which has not been in a hurry to commercialize, that really needs to quickly start charging C - end users.

On the surface, DeepSeek is not short of money under the protection of Magic Square Quantitative, and hundreds of billions of yuan in large - scale financing are on the way. However, since Doubao still needs to "subsidize its household" by selling memberships, DeepSeek, with weaker financial strength, also needs to expand its sources of income.

Another benefit of charging fees is that DeepSeek, which has always had smooth sailing, can push itself forward. In addition to researching and developing new models, it can also delve into the productivity scenarios centered on AI programming and truly have the ability to "do the work". Compared with technological and engineering innovation breakthroughs, making up for this shortcoming is of no less strategic value.

Previously, DeepSeek broke its "ancestral precepts" and introduced external shareholders, achieving its first self - breakthrough. Now, a DeepSeek that dares to charge C - end users will have the opportunity to complete its "coming - of - age ceremony" again.

01

DeepSeek's available funds are not abundant. One manifestation is the obvious shortage of computing power resources.

Among major AI apps, DeepSeek has almost the most frequent downtime. Especially after the launch of the V4 series models in May this year, DeepSeek experienced multiple service interruptions, which were more frequent than in previous months.

Part of the reason for the downtime is the soaring token consumption. According to data from the AI model aggregation platform OpenRouter, in the last week of May, the token consumption of DeepSeek V4 Flash reached 36.5 trillion, a month - on - month increase of 32%, ranking first in the industry.

Solving the computing power bottleneck is not difficult: just add servers and purchase more cloud computing power. However, this also means higher daily operating costs. If DeepSeek wants to avoid frequent downtime, it needs to make more money.

In the B - end market, DeepSeek already has the conditions to increase its charging intensity.

Although the newly released V4 series of large models are not SOTA in all aspects, they have excellent cost - performance, attracting a large number of professional users and enterprises. Coupled with the previously established reputation, DeepSeek is fully qualified to make a lot of money through price increases.

But DeepSeek obviously doesn't want to "reap the harvest" immediately.

While its competitors are raising prices, DeepSeek has lowered the price of its V4 model four times in a month. At the end of May, it even permanently reduced the price by 75%. The price is 0.025 yuan per million token inputs (cache hit), 3 yuan per million token inputs (cache miss), and 6 yuan per million token outputs. It can be called the "price butcher" in the AI circle this year.

The signal released by DeepSeek is: It hopes to gather as many B - end users as possible. Even if it can't make money for the time being, it will go all out to expand the user scale.

This is in line with DeepSeek's next plan - to be implemented in productivity scenarios.

Previously, there were reports that DeepSeek has established an Agent Harness team focusing on programming agents, targeting Claude Code under Anthropic. At the same time, DeepSeek has started recruiting for relevant positions.

It is not difficult to see that DeepSeek has high expectations for the B - end market. It hopes to capture market share through ultra - low prices and is not in a hurry to make a profit. This also means that at this stage, C - end revenue needs to shoulder the banner of DeepSeek's commercialization.

C - end users consume tokens every day but rarely contribute to revenue. They have long been a black hole of losses for AI apps.

Taking Doubao as an example, its daily token consumption soared from 120 billion in May 2024 to 12 trillion in March this year, an increase of about 1000 times. A large part of this comes from C - end users. However, since the chatbot functions of AI apps are all free and unlimited, the huge token consumption cannot be directly converted into revenue, but only brings losses.

During peak periods, this contradiction is even more prominent. DeepSeek was a bit "unable to hold on" before. On the afternoon of May 29th, many netizens found that there were limits on the number of times DeepSeek could regenerate and modify content. It is reported that DeepSeek was under too much computing power pressure and took temporary restriction measures.

However, temporary traffic restrictions alone cannot really solve the problem. Moreover, DeepSeek is adding multimodal capabilities. The token consumption of pictures, videos, and audios is hundreds or thousands of times that of text, requiring more computing power and causing more severe losses.

DeepSeek is already looking for funds. However, the 50 - billion - yuan financing scale is far less than that of Zhipu and MiniMax, which are booming in the Hong Kong stock market, and even more so compared with giants like OpenAI and Anthropic, which raise tens of billions of dollars in financing. Relying only on external financing, DeepSeek will sooner or later be unable to meet the still - rapidly rising C - end usage.

The way to break the deadlock may only be to directly charge users and set up a paywall for high - consumption advanced functions.

This is actually transplanting the B - end API business model to the C - end. Only in this way can DeepSeek truly link token usage with revenue scale, and there will be a chance to truly resolve the computing power bottleneck.

02

Charging C - end users can not only immediately generate revenue but also force product implementation through commercialization and help DeepSeek make up for the shortcoming in productivity scenarios.

Compared with domestic AI companies, DeepSeek's strengths lie in its technological concepts, model capabilities, engineering implementation, and cost - performance. Every time it releases a new model, it can always refresh the industry's understanding in these aspects and set new benchmarks.

Although DeepSeek has high - performance and high - cost - performance models, it has not fully explored AI productivity. There are still many shortcomings in its product implementation.

Take AI programming as an example. DeepSeek's latest model ranks among the top in the industry. According to DeepSeek, the Agentic programming ability of V4 is the strongest among open - source models, and it has also been specially optimized for Claude Code. After the release of V4, the call volume quickly ranked among the top in the industry, which also confirms from the side that programmers like this new model.

The problem is that DeepSeek lacks independent AI programming products like Codex and Claude Code. Most developers call models such as DeepSeek V4 through third - party tools. This limits DeepSeek's business prospects to a certain extent and also hinders the development of the DeepSeek App's function matrix.

While Doubao, Qianwen, etc. are trying their best to add various office functions to their apps and connect with e - commerce, local life, learning and education modules, DeepSeek still remains in the simple form of a chatbot and does not even support multimodality.

Facing productivity scenarios, DeepSeek holds good cards but lags behind by several steps. This slowness has led to limited commercialization progress of the app. The two aspects drag each other down, forming a vicious circle.

Now, as AI apps are starting to charge membership fees, it provides an opportunity for DeepSeek: Taking C - end charging as an entry point, let commercialization take the lead and drive up the speed on the product side.

When AI apps sell memberships, the basic chatbot function will definitely remain free. The selling points can only be advanced functions. Whether these functions are useful and sufficient largely determines whether users are willing to pay.

Taking Doubao as an example, the professional version will include professional services such as software development, data analysis, professional design, process automation, financial analysis, and scientific research.

AI apps with more mature membership systems often feature AI programming as the main offering.

For example, Kimi has divided its paid packages into four levels. The entry - level version is a continuous monthly subscription at 49 yuan per month, and the highest - level version is 699 yuan. The differences between different levels mainly lie in the Agent quota, whether Agent multi - task parallelism is supported, whether AI programming can be called, whether a professional database is supported, and whether "shrimp - raising" is allowed.

If DeepSeek wants to sell memberships, it has to emulate other AI apps and actively make up for the ability matrix related to productivity. The current DeepSeek App, which is almost blank, is bound to undergo a major transformation before it can be presented to users.

The technical difficulty of this transformation is not high, but it highly meets the needs of high - value users and is in line with the trend of emphasizing work ability in the Agent era. DeepSeek should have done this a long time ago, but for various reasons, it has never taken the first step.

The consequences are already obvious: DeepSeek defeated many competitors at the beginning of last year and topped the list of domestic AI apps; now it has been overtaken by Doubao and even lags behind Qianwen. Although there are factors such as other apps' large - scale promotion with heavy investment, the single function of DeepSeek and its frequent downtime are also important reasons for its declining popularity.

Fortunately, DeepSeek still has a large number of loyal fans.

In early May this year, an open - source project called DeepSeek - TUI attracted attention on GitHub and received 16,000 stars in one day. It is a terminal - native programming agent based on DeepSeek V4, and many developers call it the "DeepSeek version of Claude Code".

If the official cannot provide productivity, fans will create it themselves. The enthusiastic fans not only help DeepSeek make up for the shortcoming in productivity but also provide a better foundation for it to start charging membership fees.

03

DeepSeek's charging users will be another "coming - of - age ceremony" for Liang Wenfeng.

The previous "coming - of - age ceremony" was when it stopped saying "no" to external capital.

In the AI industry driven by venture capital, DeepSeek is a rather unique company: it does not accept external financing, does not dilute its equity, and is not bound by anyone's commercialization schedule. The "three noes" principle has created DeepSeek's unique temperament and past success.

However, in 2026, DeepSeek unexpectedly abandoned these principles and started to contact many giants and venture capital funds.

What is becoming increasingly clear to Liang Wenfeng and DeepSeek is that it is difficult for DeepSeek to continue leading the pack or even avoid becoming a second - tier player with only the annual revenue of 700 million US dollars from Magic Square Quantitative.

The impact of weak funds is becoming apparent. The number of App users has been overtaken; although the highly anticipated DeepSeek V4 is still the SOTA among open - source models, it has no advantage compared with a number of closed - source flagship models; the defection of core personnel such as Luo Fuli and Guo Daya to giants is an irreparable loss for DeepSeek.

In the capital market, DeepSeek's post - investment valuation of up to 400 billion yuan is equivalent to two MiniMax, but 250 billion yuan lower than Zhipu; compared with OpenAI and Anthropic, which are valued at trillions of dollars, there is a gap of two orders of magnitude.

So, DeepSeek changed its course, and Tencent, CATL, etc. are about to become new shareholders. In addition to bringing abundant funds, external investors can also create broader user entrances and implementation scenarios for DeepSeek and "graft" their own productivity tool matrix onto DeepSeek.

However, in addition to changing its view on capital, DeepSeek also needs a "coming - of - age ceremony" in terms of commercialization - from "breaking even" to "striving to make money".

In the past,