Behind DeepSeek’s Usage Limits: Why Free Consumer Traffic Has Become a Computing Burden for AGI
DeepSeek recently tightened usage limits for free users of its website and app. Long conversations that have been inactive for more than seven days can no longer be continued, the single-input limit for new conversations is about 40,000 Chinese characters, and roughly two hours of continuous use triggers a usage-limit notification. The article attributes the limits to inference costs rising with Token usage and constraints on computing resources, while noting that DeepSeek has also opened an application API that charges based on usage. Its central conclusion is that AI companies are shifting from free consumer traffic toward Agents and enterprise APIs to prioritize revenue and model iteration.
Why read this
AI score 64
The article explains DeepSeek’s usage limits from three perspectives—inference costs, computing-resource allocation, and the API business model—providing a framework for understanding the shift in AI products from free traffic to paid usage.
Topics
Body · AI translation
By Shijiao Farsight
“Go play somewhere else and don’t hold up AGI training.” This quip aimed at DeepSeek is now becoming a reality.
Recently, DeepSeek tightened usage limits for free users of its web version and app. Long conversations inactive for more than seven days can no longer be continued, the text limit for a single input in a new conversation has been reduced to around 40,000 Chinese characters, and about two hours of continuous use will trigger a message stating, “The server is busy and your usage limit has been reached.”
Internet products have always pursued a core strategy of satisfying every user need, down to the smallest detail, in order to capture more traffic. DeepSeek’s usage-limit strategy runs counter to this trend, revealing that AI products face an entirely different set of challenges.
According to Southern Metropolis Daily, DeepSeek founder Liang Wenfeng said at a closed-door investor meeting in July 2026, “We haven’t devoted much effort to the consumer market. At one point, we didn’t even want to keep supporting those users, but we couldn’t drive them away even if we tried.” The recent usage-limit policy continues DeepSeek’s cool attitude toward free users.
When a high-profile AI company with more than 100 million monthly active users begins proactively limiting consumer usage, a strikingly counterintuitive question emerges: in 2026, as AI becomes increasingly widespread, why have the vast numbers of free consumer users once regarded as indispensable become the first burden that large-model companies are shedding on their journey toward AGI?
Economies of scale break down: the more free users, the greater the losses
In the mobile internet era, although social networking, gaming, entertainment, and other products differed in functionality, their underlying strategic direction was highly similar: all sought to capture more traffic through free offerings and achieve winner-takes-all dominance.
The reason was that mobile internet products could readily achieve economies of scale. As user numbers grew, research, development, and operating costs could be spread across a larger base. Greater traffic could also help create a data flywheel and multilateral market effects, building stronger competitive barriers.
Image source: Tencent
Following the playbook of the mobile internet era, when the AI era arrived, internet giants rolled out lavish marketing campaigns to promote their free AI products to ordinary users. During the 2026 Spring Festival, for example, Tencent Yuanbao distributed RMB 1 billion in cash red envelopes, Alibaba’s Qwen launched a RMB 3 billion “Spring Festival Treat” campaign, and Baidu ERNIE introduced a RMB 500 million Spring Festival red-envelope promotion.
Image source: QuestMobile
Data disclosed by QuestMobile showed 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 calculated using Volcano Engine’s publicly available API prices that Doubao consumed tens of millions of yuan in computing costs each day in May 2026. During the first half of the year, the product generated less than RMB 1 million in average daily revenue, primarily from e-commerce commissions.
The main reason the internet industry’s economies of scale have broken down in the AI era is that the inference costs of AI products grow linearly, or even superlinearly, with Token usage. The more users there are, the greater the demand and the longer the context, the heavier the company’s computing bill becomes.
Addressing this issue in March 2026, Zhou Hongyi, founder of 360 Group, said the business models of internet products would undergo a fundamental transformation in the AI era. “Internet traffic involves moving information, with marginal costs approaching zero, so it can be free. But Tokens are intellectual output, supported by computing power, chips, and electricity. The more users there are, the greater the consumption, the higher the costs, and the more marginal costs increase. The free model simply cannot work.”
From this perspective, it was largely inevitable that DeepSeek would impose limits on free users. Unlike traditional internet companies, DeepSeek is not obsessed with controlling the user gateway. Its long-term goal is to achieve AGI, so it naturally needs to reduce non-core spending.
After limiting free users, DeepSeek did not completely close off their chat access. Instead, it released an all-in-one application API. By enabling the corresponding parameters in the existing API and paying the associated fees, developers can obtain conversational features such as deep-reasoning search and file input, and can even replicate the DeepSeek app on a one-to-one basis.
Image source: DeepSeek
DeepSeek had previously launched a paid model API. Both the all-in-one application API and the model API follow the same logic: usage-based billing, with users paying for what they consume.
It is worth noting that although DeepSeek has opened paid channels while restricting free access, it is not pressuring consumer users to pay. That is because the paid APIs are aimed not at ordinary users chatting casually in a dialogue box, but at developers or corporate customers willing to pay for their usage.
In fact, DeepSeek is not alone. Since the beginning of this year, leading AI companies worldwide have been shifting their focus away from using free offerings to attract consumer traffic and toward developing Agents and enterprise APIs, as they compete for high-value customers in productivity use cases.
Anthropic is a prime example. It has gained an early lead by leveraging its strengths in programming and the enterprise market, delivering an impressive set of financial results.
Its prospectus shows that Anthropic generated revenue of $33.242 billion in the first three quarters of 2026, up 1,206% year on year. Of that total, consumer and usage-based revenue amounted to $26.306 billion, accounting for 79.1% of total revenue.
As of the end of September 2026, Anthropic had approximately 11,900 customers that had each contributed at least $100,000 in revenue over the preceding 12 months, compared with just over 950 a year earlier. In the first three quarters, these customers contributed approximately $24.5 billion in revenue, accounting for around 74% of the total.
OpenAI is also scaling back subsidies. In late October, OpenAI announced that ChatGPT and Codex quotas would be pooled and measured uniformly by token usage. Meanwhile, Codex chief “Tibo,” nicknamed the “cyber godfather” for frequently hitting the reset button, was reassigned, after which the frequency of Codex quota resets dropped sharply.
In the Chinese market, ByteDance’s “two-sided ledger” presents an even starker contrast. As noted above, Doubao’s daily revenue was less than RMB 1 million in the first half of 2026. According to LatePost, however, revenue from the MaaS business of ByteDance’s cloud services platform, Volcano Engine, reached RMB 4.8 billion during the same period, with a gross margin above 55%, far outperforming its consumer-facing business.
Although companies such as Anthropic, OpenAI and ByteDance have different priorities in their AI strategies, they have all recognized the same trend: in the AI era, what can truly cover computing costs is not the size of the user base, but whether customers are willing to pay for every invocation.
When free conversations and paid invocations compete for the same computing resources, companies naturally need to direct those resources toward the latter.
Trade-offs under computing constraints: prioritize paid services and model iteration
Image source: DeepSeek
The growing adoption of agents, which consume more computing power, is significantly widening this gap. Take DeepSeek Harness, DeepSeek’s AI agent platform. Official data covering users of the official API show that the platform’s daily active users rose from an average of 271,000 during the week of September 30 to 1.54 million on November 4, while daily token consumption increased from approximately 24 trillion to around 92 trillion.
As API call volumes have surged, DeepSeek has found it increasingly difficult to ensure service stability. According to incomplete statistics, DeepSeek has experienced eight service outages since October 2026, affecting both its free conversational service and paid API. The October 15 outage lasted nearly five hours. DeepSeek subsequently introduced a credit compensation plan for API users and said it would “resolutely prevent similar problems from happening again.”
Computing resources are stretched thin on the inference side, and the same is true for training. At the aforementioned closed-door meeting, Liang Wenfeng said that computing power is the biggest problem facing AI today. “Often, we know the direction and how we should train. What truly constrains us is resources, especially computing resources. Many technical approaches have been temporarily shelved because of resource limitations.”
As a startup, DeepSeek has fewer computing resources than industry giants such as Tencent, Alibaba and ByteDance, yet it must simultaneously meet 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 respectively to current revenue and long-term competitiveness.
There is currently no clear evidence that DeepSeek restricted consumer-user quotas because of its frequent outages. But with a finite amount of computing power, the trade-off is obvious: the stability of paid services must be protected and model iteration cannot stop, leaving free conversations as the first area to be curtailed.
Image source: Huawei
The computing-power constraints facing Chinese AI companies are expected to ease next year. As production capacity for Huawei Ascend chips comes online, the domestic supply of computing power will gradually become more abundant. In September 2026, Bloomberg reported that DeepSeek planned to build a data center in Ulanqab, Inner Mongolia, deploying at least 160,000 Ascend 950DT chips for model inference.
However, not all of the additional computing power will go to free consumer users. Liang Wenfeng previously told investors that training the next-generation model on domestically produced chips was one of the company’s highest priorities. This means that even if supply improves, free conversations will remain relatively low in the computing-resource allocation hierarchy.
For DeepSeek, therefore, imposing quotas on free users is not a stopgap measure, but more akin to a long-term strategic arrangement.
Returning to the original question: are free consumer users ultimately an “obstacle” on AI companies’ path toward AGI?
Shijiao Farsight believes that what is impeding the evolution of AGI is not ordinary users themselves, but technology companies’ attempt to forcibly apply the “free-traffic mindset” of the mobile internet era to the industrial realities of the computing economy.
For more than two decades, China’s internet industry broadly subscribed to the belief that “whoever controls traffic controls the market.” Because the marginal cost of software approached zero, companies could amass enormous groups of free users and then monetize them through advertising and e-commerce. But in the AI era, every casual conversation incurs GPU depreciation and electricity costs. Hundreds of millions of daily active users without a closed commercial loop are no longer a moat, but a wound that keeps bleeding.
The chat box has fulfilled its mission of introducing AI to the general public, but it cannot support AGI’s ultimate commercial endgame. Redirecting limited computing resources away from free casual conversations and toward the enterprise agent ecosystem and the development of next-generation foundation models is not arrogance on the part of AI companies, but a rational choice necessary to stay in the game.
AI Trends