AI Trends

Why DeepSeek’s Free-Tier Rate Limits Triggered a Clash Between One-Star Reviews and Five-Star Defenses

36Kr · AIUpdated Nov 20 08:13

AI summary

Starting November 15, DeepSeek imposed restrictions on long conversations and usage quotas in its free app and web version, while the paid API remained unaffected. The restrictions include ending support for long conversations that have been inactive for seven days and reducing individual inputs to about 40,000 Chinese characters; some users have also encountered usage caps. In the four days after rate limiting began, 1,042 new one-star reviews and 1,136 new five-star reviews appeared in the China region of the iOS App Store. The article argues that the controversy stems from the conflict between heavy users’ needs and pressure on computing resources.

Why read this

AI score 66

Drawing on rate-limiting mechanisms, heavy-user scenarios, and changes in app-store ratings, the article highlights the computing-resource allocation dilemma facing free AI services.

Topics

Tags

Body · AI translation

If you ask which major AI model company has the best reputation in the tech community and among developers, if DeepSeek claimed second place, probably few others would dare claim first.

It does not usually pay for user acquisition or traffic, relying entirely on its models to build word of mouth.

But I truly never expected such an ostensibly upstanding company to be bombarded with negative reviews recently—and no, not by us.

Over the past few days, whether on short-video platforms or in app stores, large numbers of users have been complaining about it and furiously posting one-star reviews.

One person said, “You stupid whale, do you have a death wish? ... After the update, I can barely exchange two sentences before it says the server is busy.”

Another said, “Your game is really finished ... What do you mean I can’t continue the conversation after 72 hours? ... If Chinese AI keeps going like this, there’ll be nothing left to use and eventually no one will bother.”

Someone even called the 12345 government service hotline and threatened, “If DeepSeek isn’t fixed, I’ll swallow pills and go on a hunger strike!” This ultimately prompted a group of police officers to visit and check on the person.

Seriously, what unforgivable sin did DeepSeek commit for its reputation to collapse like this?

The cause may surprise you: DeepSeek recently ran short of computing capacity and began throttling free users.

Starting November 15, many users opening an old DeepSeek conversation began seeing an on-screen notice saying that, due to insufficient computing capacity, long conversations inactive for seven days would no longer be supported.

You either have to start a new conversation or export your chat history for use in another app.

In addition to restrictions on old conversations, the text limit for a single input has now been reduced to about 40,000 Chinese characters, and chat usage is also capped to some extent. Some users encountered a notice saying, “Due to high server demand, you have reached your usage limit,” after using it continuously for two hours.

At present, these restrictions appear only in the free app and web version that people normally use; the paid API is unaffected.

At this point, some people may be wondering: If I have not talked to it for seven days, I clearly have not been using any server resources, so why should I be restricted when I come back and send a single message?

This probably has to do with the “context disk cache” mechanism.

But after “a few hours to a few days,” DeepSeek’s cache is cleared. When the conversation resumes, the AI must reread the entire chat history.

Suppose a free user revives a 200,000-character conversation that has been inactive for seven days. Even if they merely send “Hello,” the load placed on DeepSeek’s servers is equivalent to holding another 15 to more than 40 normal exchanges in that conversation.

The calculation is below; I will not go into detail.

To be fair, this round of restrictions is barely noticeable to users who ask only a dozen or so questions each day.

But it is a lot harder to bear for people who rely heavily on context.

For example, many people have long kept medical case data, graduate entrance exam research materials, and business templates in the app. Locking those conversations has caused them considerable trouble.

Some novelists have also spent hundreds of thousands of characters developing their stories and character settings in it, only to return after a few days away and be completely dumbfounded.

Of course, the most emotionally distraught group has been users engaged in long-term role-playing and human-AI romance with DeepSeek.

After all, all their everyday domestic conversations, emotional push-and-pull, and little squabbles have accumulated in a single extremely long window. Starting a new window is like having a lover lose their memory—the conversation is inevitably going to feel wrong.

One user was so devastated that they said, “Even being cheated on wouldn’t hurt this much.”

But what is even more surprising about this incident is this:

While countless negative reviews have poured in, many others have responded by giving DeepSeek five stars and passionately sparring with the one-star crowd.

They believe that those posting one-star reviews are unproductive users who are mostly wasting computing resources.

One person said, “I support DeepSeek—it has finally cut ties with unproductive users!”

Another said, “Good thing they imposed conversation limits ... Just focus on AGI training.”

I scraped DeepSeek’s ratings data from the Chinese iOS App Store. In the four days after it began throttling usage, it received 1,042 new one-star reviews and 1,136 new five-star reviews.

On November 17 in particular, within less than 12 hours, there were 203 one-star reviews and 196 five-star reviews, with the two sides practically firing at each other point-blank.

One review would complain, “Why impose a usage cap? This is insane,” and the very next one would proclaim, “Long live the Sage Liang.” It turned straight into a turn-based game.

That is basically what happened. I do not know what our readers think, but in my view, DeepSeek has done nothing wrong.

Its app and web version have been completely free for everyone from beginning to end.

DeepSeek has never charged you a cent or monetized through advertising. How did it become the villain just because supplies are tight at home and it is serving half a bowl less?

Of course, I know that roughly one in every ten negative reviews is “asking DeepSeek to launch a VIP plan,” and some people have even questioned, “If DeepSeek wants to charge money, why doesn’t it offer in-app purchases?”

So why does DeepSeek not introduce a VIP membership subscription?

The answer is simple: What DeepSeek lacks is not money at all, but computing capacity.

According to media reports, DeepSeek secured two enormous financing rounds this year. Its first round in June was worth about RMB 50 billion, with Liang Wenfeng personally contributing RMB 20 billion; in the second round in October, nearly 30 institutions invested another approximately RMB 105.5 billion.

But having money is one thing; having computing capacity is another.

Various signs suggest that DeepSeek is severely short of computing capacity.

As early as last February, DeepSeek suspended API top-ups because its server resources were stretched and it lacked sufficient computing capacity.

This August, in an effort to shift heavy API traffic to nights and weekends, the company even introduced peak and off-peak pricing for its API.

Even so, DeepSeek has experienced 10 outages since October. On October 15, its website and API were partially unavailable for more than four hours.

Its computing capacity really cannot keep up with demand, folks.

Some people outside the industry may not understand: If you lack computing capacity, why not just buy more?

That is precisely the biggest headache for domestic large-model companies—computing capacity is not something money alone can buy.

As everyone knows, Nvidia’s restrictions on exporting high-end chips to China have become increasingly stringent. Although Bloomberg reported back in September that DeepSeek planned to deploy at least 160,000 Huawei Ascend 950DT chips in Ulanqab, Inner Mongolia, the data center will not begin partial operations until late next year or early 2028 at the earliest. That is too far off to solve the immediate problem.

During this gap, DeepSeek can only make do with the chips it already has.

If one day that is no longer enough, then to avoid affecting the experience of paying API customers, I would also prioritize restricting free consumer users if I were DeepSeek.

As for the monthly in-app subscription that so many people desperately want, let me put it this way: It would leave DeepSeek in an even tougher position.

In January 2025, OpenAI CEO Sam Altman said that the $200-per-month ChatGPT Pro plan was actually losing money because people were using it far more heavily than the company had anticipated.

In July of the same year, Anthropic also reached its limit and announced weekly usage caps for paid Claude subscriptions.

Without caps, some users really will burn through 11 billion tokens on a $200 monthly plan—the equivalent of about $35,000 at API prices.

If overseas giants that can freely buy Nvidia chips cannot withstand the firepower of monthly subscribers, then DeepSeek certainly cannot.

If DeepSeek were to bite the bullet and introduce memberships, the inevitable outcome would be that it would take users’ money and still have to impose limits, such as weekly quotas and five-hour quotas.

Then the one-star reviews would no longer say, “If you can’t afford to play, don’t play,” but rather, “You charge money and still throttle usage? What an ugly cash grab.”

Seriously, if you have hundreds of thousands of Chinese characters’ worth of indispensable novel settings or work materials, or a cherished cyber companion you cannot let go of, you can export your chat history, import it into DeepSeek Harness or the chat-demo that DeepSeek open-sourced on November 13, enter an API key, and keep using it.

There is really no need to sulk in the comments...

Even if many people do not know how to add API credit, you can ask DeepSeek to teach you.

In this day and age, you can chat with AI until you hit the limit, create all kinds of personas, and furiously leave negative reviews in the app store, yet you still do not know how to ask DeepSeek to teach you to top up the API—I really do not know what else to say.

If you do not know how to connect to an API, explore and learn. If your financial means are limited, use it within those limits.

Otherwise, have some grace and accept and respect what others have given you for free.

As Marcus Aurelius wrote in Meditations: Accept without arrogance, and let go without attachment.

Image and information sources:

DeepSeek Tightens Limits Further; Its Comments Section Is Like This (QbitAI, Xiaohongshu)

DeepSeek’s “Seven-Day Limit” Sparks Controversy; Users Say: Please, Just Take My Money (Sina Technology / TechWeb)

DeepSeek Restricts Usage Due to Insufficient Computing Capacity: Long Conversations Inactive for More Than Seven Days No Longer Supported (The Paper, republishing STAR Market Daily)

Computing Capacity Under Strain: DeepSeek Reportedly Imposes Limits on Its App and Website, Restricting Both Long Conversations and Daily Usage (Tencent News)

Details of DeepSeek’s RMB 100 Billion Financing Revealed, With Nearly 30 Institutions Participating (Sina Finance)

DeepSeek temporarily suspends API service top-ups (The Star)

160,000 Huawei Ascend 950DT Chips! DeepSeek Fights to Break Through the Computing-Capacity Bottleneck (Sina Finance)

OpenAI is losing money on its pricey ChatGPT Pro plan, CEO Sam Altman says (TechCrunch)

Anthropic unveils new rate limits to curb Claude Code power users (TechCrunch)

This article is from the WeChat official account “ChaPing X.PIN,” written by Ciwei and edited by Mangshan Laotietou and Mianxian. It is published by 36Kr with authorization.

Source

36Kr · AI