DeepSeek's commercial "coming-of-age ceremony": from "not short of money" to "working hard to make money"

DeepSeek is undergoing a strategic transformation from "not short of money" to "learning to make money". Founded nearly three years ago, it has been adhering to the "three no's" principle - not accepting external financing, not diluting equity, and not being kidnapped by anyone's commercialization schedule - all R&D investment is supported by parent company Huanfang's quantified annual revenue of approximately US$700 million. This is an anomaly in the large-scale model industry that burns money like water, and it also contributes to DeepSeek’s unique temperament and past success.

But in 2026, the "three no's" principle was broken one after another. First, it violated the "ancestral precepts" and introduced external shareholders to raise 50 billion yuan. Now the more urgent test is to charge C-end users.

Why must we charge? The computing power bottleneck is forcing

DeepSeek's available funds are not abundant, and the most intuitive manifestation is the downtime frequency - it has the highest number of downtime among the major AI Apps. There have been service outages almost every week since the V4 series went live in May this year. OpenRouter data shows that the Token consumption of DeepSeek V4 Flash in the last week of May reached 3.65 trillion, a month-on-month increase of 32%, ranking first in the industry. On the afternoon of May 29, DeepSeek even began to impose limits on the number of times it can regenerate and modify functions—temporary current restrictions can only be used as an emergency, but cannot cure the root cause.

The deeper contradiction is that the marginal effects of Internet products do not apply to AI products. Quite the contrary - The greater the number of users of the AI ​​App and the higher the usage of Token, the more serious the losses will be. Taking Doubao as an example, the average daily Token consumption soared from 120 billion in May 2024 to 120 trillion in March this year, an increase of approximately 1,000 times. Doubao, which has 345 million MAU and 145 million DAU, cannot escape this reality, so it took the lead in announcing three levels of paid subscriptions on the App Store in early May—the lowest is 68 yuan/month. The professional version includes software development, data analysis, professional design and other capabilities.

Doubao relies on the cash cow of Douyin and the financial support of Byte Group. The full-year revenue target of the Volcano Engine MaaS business has been raised to 15 billion yuan. DeepSeek's B-side strategy is to reduce prices four times a month, permanently reduce prices by 75%, and cache hits to only 0.025 yuan/million tokens - the signal released is very clear: the B-side will grab the user base first and is not in a hurry to make profits. This means that C-end revenue needs to carry the banner of commercialization.

The B-side is racing to win the market, while the C-side is in urgent need of breaking through.

On the B-side, DeepSeek already has the conditions for charging - the V4 series is extremely cost-effective and attracts a large number of professional users and enterprises. But it chose the "price butcher" route, hoping to gather as many users as possible in productivity scenarios. According to reports, DeepSeek has formed an Agent Harness team with the goal of programming intelligent agents, directly targeting Anthropic’s Claude Code.

But on the C-end product side, DeepSeek has obvious shortcomings. While competing products such as Doubao and Qianwen are trying their best to incorporate various office functions into the app, DeepSeek still remains in the simple form of a chatbot and does not even support multi-modality. In AI programming, DeepSeek's strongest area, developers mostly call its models through third-party tools. There is no official independent programming product like Codex or Claude Code.

The results are already showing: DeepSeek topped the list of domestic AI apps early last year, but now it has been overtaken by Doubao, and even fell behind Qianwen. The overlay of a single function often crashes, which is an important reason for the decline in popularity. But DeepSeek’s loyal fans have not left – in early May this year, an open source project called DeepSeek-TUI received 16,000 stars in one day on GitHub. It is a terminal native programming agent based on DeepSeek V4, which is called "DeepSeek version of Claude Code" by developers. If the official productivity cannot be provided, fans will create it themselves.

Commercialized double "coming-of-age ceremony"

DeepSeek is facing an increasingly clear fact: it is difficult for it to continue to lead, and it may even become a second-tier player, relying solely on the magic square to quantify annual revenue of about US$700 million. The impact of weak funding is showing up in multiple dimensions - the number of App users has been surpassed. Although the much-anticipated V4 is still an open source model SOTA, it has no absolute advantage over the closed source flagship model. Key personnel such as Luo Fuli (who switched to Xiaomi) and Guo Daya (who joined Byte) have defected to giants, which is an irreparable loss.

In the capital market, DeepSeek's highest post-investment valuation of 400 billion yuan is equivalent to two MiniMax, but 250 billion lower than Zhipu; compared with the trillion-dollar OpenAI and Anthropic, the gap is two orders of magnitude.

Charging the C-side is one of the two "coming-of-age rites" that DeepSeek must complete - the first is to accept external financing, and the second is to shift from "no loss, no profit" to "strive to make money." In the past, Liang Wenfeng was cautious about profitability, and profit was not the company's primary goal. But in a new stage where AI is shifting from a SOTA model duel to a comprehensive competition of "product + technology + business closed loop", Anthropic not only surpasses OpenAI in revenue but is also expected to turn a profit for the first time in Q2 this year - the capital market is rewarding AI companies that make money. DeepSeek cannot change this trend, it can only change itself.

By imitating Doubao and charging consumers, not only can they gain immediate revenue, but they can also force the product to be launched through commercialization, forcing DeepSeek to complete its capability matrix for productivity scenarios such as AI programming. This will be another leap in Liang Wenfeng's strategic thinking after breaking the ancestral motto of "no financing".