HomeArticle

DeepSeek quota limit, Tencent's social empire breathes a sigh of relief

Tech商业2026-11-18 19:03
Costs and Shifts in the Social Landscape in the Era of Free AI

Starting from November 16, after a completely ordinary query, DeepSeek users encountered a line of prompt: "Due to heavy server load, your usage quota has reached the upper limit."

Some users were blocked after using the service for less than two consecutive hours, while others found that long conversations they had maintained for a long time suddenly failed to continue, and the system suggested "starting a new conversation" or "exporting historical records". On the same day, the topic "DeepSeek usage quota has reached the upper limit" became a trending search.

This single line of prompt has pulled two seemingly unrelated things together.

The first is the calculation of computing power: the era of AI conversations that was defaulted to "free, unlimited, and always available" has revealed its cost ceiling to the public for the first time.

The second is the calculation of social interaction: the foundation of the social empire is the conversations between people. Nowadays, an increasing amount of conversation time is being taken up by human-AI interactions.

Nothing that is free comes without a price, it is just that someone else is paying for you. When the payer starts to feel the burden, the account book is laid open.

Three Types of Tightening, One Core Reason

First, let's look at what has been restricted.

According to media reports, DeepSeek has imposed three restrictions on ordinary C-end users of both its App and web version: long conversations that have existed for more than 7 days with no active interactions are no longer supported (some users received a 72-hour limit); the maximum single input text length has been reduced to about 40,000 characters; the total daily conversation volume is capped, and the "quota reached upper limit" prompt pops up once the cap is hit.

The official reason is only four words: insufficient computing power. There are no unified rules, no timetable for lifting the restrictions. As a result, user experiences are highly fragmented — some users are blocked after 7 days, while others lose access after 72 hours. The most frequent comment in the comment section is: It's not that we are unwilling to pay, but that there is no place to make a payment.

There is a key dividing line here: only C-end users are restricted. Developers who use the API follow a different set of rules — independent concurrency limits, pay-as-you-go billing, which is completely different from the free chat window.

The same model has two prices. One side pays with waiting time and quotas, while the other side pays with real money. DeepSeek has presented both paths, and one of them is labeled "insufficient computing power".

Even for users who are willing to pay, this path is not easy to take. Switching from the App to the API requires registration, recharging, obtaining a secret key, and configuring a third-party client, which is far less convenient than opening the dialog box and asking questions directly. This seemingly technical threshold is essentially a price threshold: it divides users into two groups, those who "chat casually" and those who "use the service seriously".

This is not an issue exclusive to DeepSeek. Over the past year, large-volume free traffic has almost become the industry standard, and the logic of subsidizing to acquire users has been pushed to the extreme. However, when the call volume ranks first in the world, the subsidy cost can no longer be hidden — the more users there are, the higher the cost the platform bears for every additional sentence in the conversation. DeepSeek is just the first player to put this arithmetic problem on the table.

The Account Book Behind a Few Cents

Why did DeepSeek suddenly tighten its restrictions? Let's do the math first.

The new pricing that took effect on August 17 this year introduced peak-valley pricing to the AI industry. Taking the high-performance V4 Pro as an example, the price for every million output tokens during peak hours is 27 yuan, and 9 yuan for input (cache miss); the price is halved during off-peak hours, at 13.5 yuan for output and 4.5 yuan for input. For the more affordable Flash model, the peak price is 9 yuan for output and 3 yuan for input, and it is halved again during off-peak hours.

Selling computing power at different prices for peak and off-peak hours is extremely similar to electricity and mobile phone data pricing. AI is gradually becoming a public utility business.

The cost of a few cents for one long conversation review is hidden in this set of figures. The so-called "review" refers to the model reprocessing the accumulated long historical content. Assuming a conversation has a context of 200,000 tokens, running it through V4 Pro during off-peak hours, with the cache miss part calculated at 4.5 yuan per million tokens, one review costs nearly 1 yuan; if the context is longer and the operation runs during peak hours, the price will rise further. The cost ranging from a few cents to one yuan comes from this.

For users who use the service for role-playing, paper reading, contract reviewing, and code base analysis, "chatting" is actually the most expensive category on the computing power bill — dozens or hundreds of rounds a day, and the bill accumulates extremely quickly.

The real key lies in the five words "cache miss". DeepSeek sets a much lower price for "cache hit" input: 0.15 yuan per million tokens for V4 Pro during off-peak hours, which is only 1/30 of the cache miss price. To put it in plain terms: repeatedly calling the same prefix costs very little; once the prefix expires and the historical content needs to be reloaded, it will be charged according to the "cache miss" standard.

This is also why the prompt "it is recommended to start a new conversation" carries a touch of black humor: users think it is a service upgrade, but in fact they are moved out of the cache area. If they want to continue chatting about the old content with hundreds of thousands of tokens, they have to be billed according to the new rules. The saved computing power and the extra cost paid are two sides of the same coin.

Looking at the rule that "long conversations with no activity for 7 days will no longer be supported", the logic becomes clear: the cache of idle sessions has long been released, and resuming the conversation is essentially an expensive cache miss. Instead of letting users burn computing power unconsciously, it is better to draw a line in advance. This is a technical constraint, and more importantly, a cost constraint.

The demand curve is still rising sharply. According to data from OpenRouter, in the week from November 9 to 15, DeepSeek V4.1 Flash completed 59.2 trillion token calls, ranking first in the world; in the week of September 14, the figure was only 15.8 trillion, nearly three times the growth in two months. The expansion of computing power has not kept up, and there have been at least five service interruptions since October, including a 5-hour downtime on October 15. The ceiling users encountered is that the demand curve has finally hit the physical boundary.

The "Sesame" in Liang Wenfeng's Eyes

Viewing the quota limit as merely a sign that "the platform cannot afford the cost" is still an underestimation of DeepSeek.

In July this year, founder Liang Wenfeng talked with investors for nearly 4 hours, repeatedly emphasizing that the company's long-term goal is AGI, not maximizing traffic and revenue. He called the current batch of commercial opportunities "sesame", which is not worth betting all resources on. In the same communication, he said frankly that the biggest problem for AI is still computing power: "The team knows the direction and how to train the model, but what really restricts the team is computing power resources."

Last Spring Festival, DeepSeek gained a huge number of users overnight, but it did not focus on user retention, monetization, or building a super App like other players. This restraint continues to this day: users ask for monthly or quarterly subscriptions, but the platform does not offer them; developers found traces of quota management and the "Coding Plan" in the front-end code of the web version, but the official has remained silent. While peers put membership subscriptions on the shelves, DeepSeek is like an exhibition hall that insists on not charging admission — until there are too many people to fit in the exhibition hall.

What is lacking is not users, but computing power; conversely, it is also true: there is no lack of imagination for monetization, what is lacking is the resources to fulfill the free commitment.

In October this year, DeepSeek completed its second round of financing. According to media reports, the scale is about 1055 billion yuan, with a valuation of about 6.155 trillion yuan. Nearly 30 institutions including Tencent and CATL participated in the round, and the funds are mainly invested in computing power infrastructure. Therefore, the quota limit is most likely a temporary measure: when computing power is supplemented, the ceiling will be raised. However, the promise of "free and unlimited" has already cracked a gap.

Chatting is the Foundation of the Social Empire

Now let's shift our focus back to Tencent.

The commercial empire built by WeChat and QQ has only one foundation: people are willing to talk to other people. Chats, group chats, Moments, comment sections, all commercial closed loops start from this point.

However, this foundation is being gradually diverted. More and more conversation time is shifting from "finding a person to talk to" to "asking AI". A conversation partner that is always patient, never leaves messages unread without replying, and is online 24/7 is taking away the time that originally belongs to human-to-human interactions.

What is interesting is Tencent's own actions. It did not shut AI out, but instead welcomed it in. In its 2026 Q2 financial report, WeChat and WeChat combined monthly active users reached 1.439 billion, and QQ mobile monthly active users reached 520 million, the user base is very stable. Yuanbao can appear in WeChat just like a contact, the platform used 1 billion yuan in red envelopes during the Spring Festival to promote "Yuanbao Party" for multi-person AI social interactions, Pony Ma mentioned in the performance meeting that Agents should have a "real human feel", the AI agent "Xiaowei" in WeChat has been in small-scale gray testing, and Martin Lau said that AI investment in 2026 will at least double. Do not forget that Tencent is also a participant in DeepSeek's latest financing round.

This empire obviously does not intend to hold the boundary of "no human-machine conversations", it wants to be the one that defines "how humans and AI get along".

Pushing People Back to the Crowd

However, there is a real commercial logic hidden in this matter.

AI companionship and interpersonal socialization are being priced by two completely different cost curves. The bottleneck of AI conversations is computing power — physical, expensive, and will be broken through by the demand curve; the bottleneck of human-to-human conversations is attention — psychological, free, and has no server bill. Over the past year, AI companies have used subsidized free services to force the first curve to approach zero: a patient, non-annoying conversation partner is always available at your fingertips. The relative cost-effectiveness of interpersonal socialization has been pushed to a low point.

Social platforms are essentially in the attention business. WeChat, Douyin and other platforms are always competing for users' limited waking time. When AI becomes a new competitor for this time, and is subsidized to be almost free, the impact is real — it does not have to be interpreted as "empire panic", it is just a redistribution of time.

Now the subsidies are starting to ebb. Free AI conversations are starting to have limits, queues, and payment requirements, and the previously flattened curve is returning to its original position. No one is persuaded to go back to the crowd, it is just that "chatting with AI" has regained friction and price, and the relative value of "chatting with people" has risen accordingly.

From a more macro perspective, this force does not only come from one company. At the regulatory level, rules targeting AI anthropomorphic interactions are starting to draw red lines, explicitly prohibiting the induction of emotional dependence; at the research level, more and more studies are pointing out the correlation between long-term AI companionship and loneliness and dependency. Cost, regulation, and research, these three forces point to the same direction: unlimited, frictionless, and cost-free AI companionship is being redefined.

Thus, an interesting division has emerged: for AI to learn to chat with you like a human, it relies on stacking computing power, context, and memory, and every extra bit of "human-like" trait requires extra cost; while human-to-human conversations naturally have memory, context, and emotions, with zero cost. Technology still has a long way to go to catch up with interpersonal socialization on the cost curve. The quota limit is just a small signpost on this road.

Companionship is Finally Being Priced

The line of "quota reached upper limit" marks a watershed for the entire industry.

The era of free and unlimited services will not end with an official announcement. It comes to a quiet end with quota limits, peak-valley pricing, and "cache miss" bills. DeepSeek is not the first, nor will it be the last — all AI products that use "always-online companionship" as their selling point will sooner or later have to answer the same question: who pays for this companionship?

Interpersonal socialization may be the most unexpected beneficiary in this return to normal pricing. It is not because there is less AI, but because companionship itself is finally being priced. When "chatting with AI" becomes expensive again, the phrase "let's hang out and chat when you are free" has become more cost-effective than ever before.

This article is from the WeChat official account "Tech Business", author: Tech Business, published with authorization from 36Kr.