DeepSeek limit: Free users become a "stumbling block" to AGI?
Farsight
2026.11.20 11:39
· From Beijing
Full text 3669 words
Text | Farsight
"Go somewhere else to play, don't delay AGI training." This ridicule of DeepSeek is now becoming a reality.
Recently, DeepSeek has tightened the usage quota for free users of the web version and App. Long conversations that have been inactive for more than 7 days cannot be used anymore. The upper limit of single input text in a new conversation has been compressed to about 40,000 words. If used continuously for about two hours, the prompt "The server is busy and the usage quota has reached the upper limit" will be triggered.
The core strategic appeal of Internet products has always been to meet the needs of users in every detail, thereby capturing more traffic. DeepSeek’s quota strategy deviates from the above trends and shows that AI products face completely different challenges.
According to the "Southern Metropolis Daily" report, in July 2026, at a closed-door meeting with investors, DeepSeek founder Liang Wenfeng once said, "We did not spend a lot of effort on the C-side. At one time, we even did not want to maintain those users, but the users could not be driven away." The recent quota strategy continues DeepSeek's indifferent attitude towards free users.
When a star AI company with over 100 million monthly active users began to take the initiative to give quotas to C-side users, a very contrasting question surfaced - in 2026, when AI is gradually becoming more popular, why did the massive free C-side users that were once regarded as the standard become the first baggage to be abandoned by large model companies on their way to AGI?
The scale effect fails, the more free users lose money, the more miserable they will be.
In the mobile Internet era, although the functions of social networking, games, entertainment and other products are different, the underlying strategic orientations are highly consistent, and they are all committed to capturing more traffic through free strategies and achieving winner-take-all.
The reason is that mobile Internet products can easily achieve economies of scale. As the number of users increases, R&D and operating costs can be diluted. And higher traffic scale can in turn build a data flywheel and multilateral market effects, establishing stronger competition barriers.
Following the experience of the mobile Internet era, after the advent of the AI era, a number of Internet giants have launched aggressive marketing strategies to promote their free AI products to ordinary users. For example, during the 2026 Spring Festival, Tencent Yuanbao distributed 1 billion yuan in cash red envelopes, Alibaba Qianwen launched a 3 billion yuan "Spring Festival Treat Plan", and Baidu Wenxin launched a 500 million yuan Spring Festival red envelope event.
Under the huge subsidies, many AI products have indeed captured a large number of users. The problem is that Internet giants have not used this to gain outstanding performance. Doubao, which leads the user scale, is an example.
Data disclosed by QuestMobile shows that in June 2026, Doubao’s monthly active users were 382 million, a year-on-year increase of 172.1%, ranking first in AI native apps. However, "LatePost" combined the calculations of the Volcano Engine's public API price and found that in May 2026, the computing power cost consumed by Doubao reached tens of millions of yuan per day. In the first half of the year, the average daily revenue of the product was less than one million yuan, mainly from e-commerce commissions.
The reason why the scale effect of the Internet industry fails in the AI era is mainly because the reasoning cost of AI products increases linearly or even superlinearly with the amount of token calls. The greater the number of users, the greater the demand, and the longer the context, the heavier the computing bill will be for enterprises.
In this regard, in March 2026, Zhou Hongyi, founder of 360 Group, said that the business model of Internet products in the AI era will usher in a qualitative change. "Internet traffic is the transfer of information, and the marginal cost is close to zero, so it can be done for free; but Token is an intellectual output, supported by computing power, chips, and electricity. The more users, the greater the consumption, the higher the cost, the marginal cost increases, and the free model simply cannot work."
From this point of view, DeepSeek has a strong inevitability for free user limits. Unlike traditional Internet companies, DeepSeek has no obsession with controlling user access. Its long-term goal is to achieve AGI, which naturally requires reducing non-core investments.
AI business focus shifts, betting on productivity to pay
After giving free users a limit, DeepSeek did not completely block the user's dialog box, but released the Family Bucket application API. Developers can enable the corresponding parameters in the original API and pay to obtain dialogue functions such as deep thinking search and file input, and can even copy the DeepSeek App 1:1.
Previously, DeepSeek has launched a model payment API. Whether it is Family Bucket application API or model API, they all follow the same set of logic: billing is based on the number of calls, and you pay for what you use.
It should be noted that although DeepSeek has tightened the free entrance and opened up paid channels, DeepSeek has not "forced" C-end users. Because the paid API is not aimed at ordinary users chatting in dialog boxes, but developers or corporate customers who are willing to pay for the calls.
In fact, not only DeepSeek, but also the focus of the world's leading AI companies is shifting this year, from relying on free strategies to attract C-side traffic to deeply cultivating Agent and B-side APIs to compete for high-net-worth customers in productivity scenarios.
A typical example is Anthropic, which has taken advantage of its advantages in the programming and enterprise markets and has delivered outstanding performance results.
The prospectus shows that in the first three quarters of 2026, Anthropic’s revenue was US$33.242 billion, a year-on-year increase of 1,206%. Among them, consumption and usage-based billing revenue was US$26.306 billion, accounting for 79.1% of total revenue.
As of the end of September 2026, Anthropic had approximately 11,900 customers that contributed revenue of US$100,000 and above in the past 12 months, compared with just over 950 a year ago. In the first three quarters, these customers contributed approximately US$24.5 billion in revenue, accounting for approximately 74% of total revenue.
OpenAI is also tightening subsidies. At the end of October, OpenAI announced that the quotas of ChatGPT and Codex will be interoperable and measured by Token. At the same time, the head of Codex "Tibo", known as the "Cyber Foster Father" for frequently pressing the reset button, has been transferred from his post, and the frequency of Codex's quota resets has dropped sharply.
In the Chinese market, ByteDance’s “double-sided ledger” is more contrasting. As mentioned earlier, in the first half of 2026, the daily revenue of bean bags was less than one million yuan. According to "LatePost", during the same period, the MaaS business revenue of ByteDance's cloud service platform Volcano Engine reached 4.8 billion yuan, with a gross profit margin of over 55%, which is far better than the C-end business.
Although Anthropic, OpenAI, ByteDance and other companies have different focuses on deploying AI business, they all see a trend. In the AI era, what can really cover the cost of computing power is not the size of users, but whether customers are willing to pay for each call.
When free conversations and paid calls compete for the same computing power, companies naturally need to invest resources in the latter.
Trade-offs under computing power constraints, first guarantee payment and model iteration
Compared with their overseas counterparts, Chinese AI companies not only need to solve commercialization problems, but also face the challenge of tight supply of high-end computing power due to the influence of geopolitical games.
The increasing popularity of agents with higher computing power consumption is significantly enlarging this gap. Taking DeepSeek's AI agent platform DeepSeek Harness as an example, official data shows that based on users using the official API, the platform's daily activity increased from an average of 271,000 in the week of September 30 to 1.54 million on November 4, and the daily consumption of tokens increased from approximately 24 trillion to approximately 92 trillion.
As the volume of API calls increases, it becomes increasingly difficult for DeepSeek to ensure service stability. According to incomplete statistics, DeepSeek has experienced 8 service interruptions since October 2026, involving free conversation services and paid APIs. The malfunction lasted for nearly five hours on October 15. Afterwards, DeepSeek launched a bonus compensation plan for API users and stated that it "resolutely prevents similar problems from happening again."
The computing power of the inference end is stretched, and the same is true for the training end. Liang Wenfeng said at the aforementioned closed-door meeting that the biggest problem in AI today is computing power. "Many times, we know the direction and how to train. What really limits us are resources, especially computing resources. Many technical routes are temporarily put aside because of resources."
As a startup, DeepSeek's computing power reserves are not as good as those of industry giants such as Tencent, Alibaba, and ByteDance, but it has to cope with three types of needs: free conversations, paid API calls, and next-generation model training. The first category is cost burden, and the latter two correspond to current revenue and long-term competitiveness respectively.
There is currently no clear evidence that DeepSeek limits the usage quota of C-side users because of frequent failures. However, under the premise that the total amount of computing power is limited, the choice is obvious. The stability of paid services must be maintained, model iteration cannot be stopped, and the first thing to be compressed can only be free conversations.
The dilemma of Chinese AI companies’ tight computing power is expected to be alleviated next year. With the release of Huawei's Shengteng production capacity, the supply of domestic computing power will gradually become abundant. In September 2026, according to Bloomberg, DeepSeek plans to build a data center in Ulanqab, Inner Mongolia, and deploy at least 160,000 Ascend 950DTs for model inference.
However, not all the new computing power will flow to free C-side users. Liang Wenfeng once revealed to investors that training next-generation models with domestic chips is one of the company's highest priorities. This means that even if the supply improves, free dialogue is still relatively low in the ranking of computing power distribution.
Therefore, for DeepSeek, targeting free user limits is not a temporary measure, but more like a long-term strategic arrangement.
Going back to the original question, are free C-side users a "stumbling block" for AI companies on their way to AGI?
Farsight believes that what hinders the evolution of AGI is not ordinary users themselves, but technology companies' attempts to forcibly apply the "free traffic thinking" of the mobile Internet era to the industrial laws of the computing economy era.
Over the past two decades, China's Internet industry has generally believed that "those who get traffic will win the world" because the marginal cost of software is close to zero. After accumulating a large free user base, it can achieve commercialization through advertising and e-commerce. But in the AI era, every chat consumes graphics card depreciation and electricity bills. The lack of a commercial closed loop with hundreds of millions of daily users is no longer a moat, but a wound that continues to bleed.
The chat box has fulfilled the mission of AI enlightenment for all, but it cannot carry the business end of AGI. Divorcing limited computing power from free chatting and investing it in the research and development of B-side Agent ecology and next-generation basic models is not the arrogance of AI companies, but a rational choice to stay at the poker table.
Respect for the original work, if you are qualified, you should be appreciated.
According to the real-name requirement of the Cybersecurity Law, please bind your mobile phone number before posting a comment.
Express
11:48
Breaking the record, China achieves wireless communication verification with the highest code rate in the 6G frequency band
Malaysian Prime Minister Anwar arrived at the meeting
11:47
Xi Jinping meets with Thai Prime Minister Anutin
Qiyuan Robot signed a strategic cooperation with the Postal and Telecommunications Equipment Consumer Electronics Division
11:46
A-share lunch review: Shenzhen Component Index and ChiNext Index both fell by more than 2%, and the precious metals and liquor sectors performed sluggishly.
11:35
The relevant person in charge of Byte refutes rumors that Xie Xin and Zhang Nan have resigned: the information is inaccurate
11:32
The main domestic futures contracts rose or fell, with palladium and Shanghai silver falling nearly 9%.
11:28
Meta plans to launch Muse Charm, priced at $300
11:27
Spot gold fell below $3,520 per ounce, down nearly 3% on the day
11:23
Over 60% of China's sustainable development indicators are close to or achieving the 2030 goal
11:14
The concept of AI comic drama rebounded during the session, and Guomai Culture rose by more than 10%
11:13
Procter & Gamble plans to cut manufacturing and logistics employees by up to 30% by 2030
11:04
National Energy Group established a hydrogen energy technology company in Ordos with a registered capital of approximately 350 million
11:00
Domestic gold jewelry price falls below 1,100 yuan per gram
10:55
Korean exchange launches SIDECAR mechanism and suspends KOSPI programmed selling
10:54
Pork concept rebounded in intraday shock, Jin Xinnong went out of 3 boards in 6 days
10:51
22 private equity firms appeared in the first batch of 15 active ETFs, holding a total of 524 million shares
10:48
Alibaba Q2 of fiscal year 2027: Xianyu daily activity increased by 23% year-on-year, and order volume increased by 29%
10:47
Thai Prime Minister Anutin arrived at the meeting
10:46
Five departments issued the 2026 version of the "Guidance Catalog" for marine industry investment and financing
Scan to download App




