11/20 2026
366
Author | Lucky
"Go play somewhere else—don't hold up AGI training." What started as a joke at DeepSeek's expense is now becoming reality.
Recently, DeepSeek tightened usage quotas for free users of its web version and App. Long conversations left inactive for more than seven days can no longer be continued; the single-input text limit for new conversations has been compressed to around 40,000 characters; and continuous use for about two hours triggers the prompt "Server busy, usage quota reached."
For a long time, the core strategic pursuit of internet products has been to satisfy user needs in exhaustive detail and thereby capture more traffic. DeepSeek's quota strategy departs from this trend, revealing the distinctly different challenges facing AI products.
According to a report by Southern Metropolis Daily, at a closed-door investor meeting in July 2026, DeepSeek founder Liang Wenfeng said, "We haven't put much effort into the consumer side; at one point we didn't even want to keep those users, but they just wouldn't leave." The recent quota strategy continues DeepSeek's cold attitude toward free users.
When a star AI company with over 100 million monthly active users begins proactively imposing quotas on consumer users, a sharply contrasting question surfaces: in 2026, as AI becomes increasingly widespread, why has the massive free consumer user base—once regarded as sacrosanct—become the burden that large model companies are the first to discard on the road to AGI?
01 Economies of Scale Fail: The More Free Users, the Worse the Losses
In the mobile internet era, although products such as social networking, gaming, and entertainment differ in function, their underlying strategic orientation is highly convergent: all are committed to capturing more traffic through free strategies and achieving winner-takes-all.
The reason is that mobile internet products can easily achieve economies of scale; as user numbers grow, R&D and operating costs can be diluted. Moreover, a larger traffic scale can in turn build a data flywheel and multi-sided market effects, establishing stronger competitive barriers.

Image source: Tencent
Following the experience of the mobile internet era, after the AI era arrived, a host of internet giants rolled out large-scale 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 Qwen launched a 3 billion yuan "Spring Festival treat plan," and Baidu Ernie rolled out a 500 million yuan Spring Festival red envelope campaign.
Under massive subsidies, many AI products did capture huge user bases. The problem is that the internet giants did not thereby secure impressive performance—Doubao, which leads in user scale, is one example.

Image source: QuestMobile
Data disclosed by QuestMobile shows that in June 2026, Doubao had 382 million monthly active users, up 172.1% year-on-year, ranking first among AI-native Apps. However, LatePost, based on calculations using Volcano Engine's publicly disclosed API prices, estimated that in May 2026, Doubao consumed tens of millions of yuan in compute costs per day, while in the first half of the year the product averaged less than 1 million yuan in daily revenue, mainly from e-commerce commissions.
The reason the internet industry's economies of scale fail in the AI era is mainly that the inference cost of AI products grows linearly or even superlinearly with token call volume. The more users, the greater the demand, and the longer the context, the heavier the compute bill the company bears.
In this regard, in March 2026, 360 Group founder Zhou Hongyi said that the business model of internet products in the AI era will undergo a qualitative change. "Internet traffic is information to carry (information transfer), with marginal costs approaching zero, so it can be free; but tokens are intellectual output, supported by compute, chips, and electricity—the more users, the greater the consumption, the higher the cost, with increasing marginal costs, and the free model simply doesn't work."
From this perspective, DeepSeek's quota limits on free users have a strong inevitability. Unlike traditional internet companies, DeepSeek has no obsession with controlling the user gateway; its long-term goal is to achieve AGI, so it naturally needs to compress non-core investment.
02 The Center of Gravity of AI Business Is Shifting, Betting on Paid Productivity
After imposing quotas on free users, DeepSeek did not completely seal off users' dialogue boxes, but instead released a full-suite application API. Developers who enable the corresponding parameters in the original API and pay can obtain dialogue features such as deep-thinking search and file input, and can even replicate the DeepSeek App 1:1.

Image source: DeepSeek
Previously, DeepSeek had already launched paid model APIs. Whether it is the full-suite application API or the model API, both follow the same logic: billing by call volume, pay as you go.
It should be noted that although DeepSeek opened paid channels while tightening free access, it did not "force charges" on consumer users. Because the paid API is not aimed at ordinary users chatting in a dialogue box, but at developers or enterprise customers willing to pay for call volume.
In fact, not only DeepSeek—since the beginning of this year, the focus of leading global AI companies has been shifting, from relying on free strategies to attract consumer traffic toward deepening Agents and B2B APIs, competing for high-net-worth customers in productivity scenarios.
A typical example is Anthropic, which has already stolen a march by leveraging its advantages in programming and the enterprise market, delivering impressive results.
The prospectus shows that in the first three quarters of 2026, Anthropic's revenue was 33.242 billion USD, up 1,206% year-on-year. Of this, consumer and usage-based revenue was 26.306 billion USD, accounting for 79.1% of total revenue.
As of the end of September 2026, Anthropic had about 11,900 customers who contributed revenue of 100,000 USD or more over the past 12 months, compared with only more than 950 a year earlier. In the first three quarters, these customers contributed about 24.5 billion USD in revenue, accounting for about 74% of total revenue.
OpenAI is also tightening subsidies. At the end of October, OpenAI announced that ChatGPT and Codex quotas would be interconnected and uniformly measured by tokens. Meanwhile, Codex head "Tibo," known as the "cyber godfather" for frequently pressing the reset button, has been reassigned, and the frequency of Codex quota resets has dropped sharply.
In the Chinese market, ByteDance's "two-sided ledger" is even more contrasting. As mentioned earlier, in the first half of 2026, Doubao's daily revenue was less than 1 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 margin exceeding 55%, far better than the consumer business.
Although companies such as Anthropic, OpenAI, and ByteDance have different focuses in laying out AI businesses, they all see a trend: in the AI era, what can truly cover compute costs is not user scale, but whether customers are willing to pay for each call.
When free conversations and paid calls compete for the same batch of compute, companies naturally need to direct resources to the latter.
03 Trade-offs Under Compute Constraints: Prioritize Paid Services and Model Iteration
Compared with overseas peers, Chinese AI companies not only need to solve commercialization challenges but, affected by geopolitical gaming, also face the challenge of tight high-end compute supply.

Image source: DeepSeek
The growing popularity of Agents, which consume more compute, is significantly amplifying this gap. Taking DeepSeek's AI agent platform DeepSeek Harness as an example, official data shows that, among users of the official API, the platform's daily active users grew from an average of 271,000 in the week of September 30 to 1.54 million on November 4, and daily token consumption grew from about 24 trillion to about 92 trillion.
As API call volume surges, DeepSeek finds it increasingly difficult to guarantee service stability. According to incomplete statistics, since October 2026, DeepSeek has experienced 8 service interruptions, involving both free dialogue services and paid APIs. Among them, the outage on October 15 lasted nearly five hours. Afterwards, DeepSeek launched a compensation plan of bonus credits for API users and stated it would "resolutely prevent similar problems from happening again."
Compute on the inference side is stretched thin, and so is the training side. Liang Wenfeng said at the aforementioned closed-door meeting that the biggest problem in AI today is compute. "Many times, we know the direction and we know how we should train. What truly limits us is resources, especially compute resources. Many technical routes have been temporarily set aside because of resources."
As a startup, DeepSeek's compute reserves are inferior to those of industry giants such as Tencent, Alibaba, and ByteDance, yet it must simultaneously handle three types of demand: free conversations, paid API calls, and next-generation model training. The first is a cost burden, while the latter two correspond to current revenue and long-term competitiveness, respectively.
At present, there is no clear evidence showing that DeepSeek limited consumer users' usage quotas because of frequent outages. But under the premise of limited total compute, the trade-offs are obvious: the stability of paid services must be ensured, model iteration cannot stop, and the first thing to be compressed can only be free conversations.

Image source: Huawei
The dilemma of tight compute for Chinese AI companies is expected to ease next year. With the release of Huawei Ascend production capacity, domestic compute supply will gradually become more abundant. In September 2026, according to Bloomberg, DeepSeek plans to build a data center in Ulanqab, Inner Mongolia, deploying at least 160,000 Ascend 950DT chips for model inference.
However, the newly added compute will not all flow to free consumer users. Liang Wenfeng once told investors that training next-generation models with domestic chips is one of the company's highest priorities. This means that even if supply improves, free conversations will still rank relatively low in compute allocation.
Therefore, for DeepSeek, limiting quotas for free users is not a stopgap measure but more of a long-term strategic arrangement.
Returning to the initial question: are consumer free users really the "roadblock" on the path of AI companies toward AGI?
Farsight believes that what hinders the evolution of AGI is not ordinary users themselves, but tech companies' attempt to forcibly apply the "free traffic thinking" of the mobile internet era to the industrial laws of the compute economy era.
Over the past two decades, China's internet industry has generally believed that "whoever wins traffic wins the world," because the marginal cost of software approaches zero, and after accumulating a huge free user base, commercialization can be achieved through advertising and e-commerce. But in the AI era, every casual chat consumes GPU depreciation and electricity; hundreds of millions of daily active users without a closed commercial loop are no longer a moat but a continuously bleeding wound.
The chat box has completed the mission of nationwide AI enlightenment, but it cannot carry the commercial endgame of AGI. Extracting limited compute from free casual chat and directing it toward the B2B Agent ecosystem and the R&D of next-generation foundation models is not the arrogance of AI companies, but a rational choice to stay at the table.
Interactive Topic
What AI tools do you use daily? What do you think of DeepSeek's quota limits?
This article is original content of Farsight. Reproduction is prohibited without authorization.