Subscribe to our newsletter

Stay updated with the latest news and trends.

No spam. Unsubscribe anytime.

Browse

Categories, quick links and the newsletter.

Quick Links

Newsletter

Get the latest updates

Subscribe to our newsletter to stay informed about trending topics.

No spam. Unsubscribe anytime.

T
Tech

DeepSeek Says It Is Short of Compute — and Starts Rationing Its Chats

Six weeks after its model became the most-used AI in the world on OpenRouter, DeepSeek is capping input lengths, cutting off long-idle conversations and throttling daily messages — and telling users the reason in a system message: not enough compute.

The chatbot that in late October became the most-used AI model on OpenRouter is now asking its users to use less of it. Over the weekend, screenshots began circulating on Weibo of DeepSeek's own system notices: conversations left idle for more than seven days "are no longer supported due to insufficient compute," single inputs are capped at roughly 40,000 Chinese characters, and once a day's messages run out, the app answers with a "server busy, quota reached" notice.

The financial outlet Caijing compiled the changes on Monday, and Taiwan's Central News Agency reported the same limits with a blunt gloss: DeepSeek itself is admitting its computing power is not enough, "sparking concern over its compute supply." The story climbed to No. 5 on Baidu's trending board on Sunday and turned into one of Zhihu's busiest tech discussions within a day.

Caijing's Weibo card for its report on the rationing. Photo: @财经网
Caijing's Weibo card for its report on the rationing. Photo: @财经网

The rationing lands awkwardly because of what DeepSeek's rise was supposed to prove. The Hangzhou lab built its reputation on efficiency — extracting frontier-grade performance from hardware and budgets its rivals considered impossible — and spent the autumn compounding the story: a model that swept the global top five on OpenRouter, then an agent architecture designed to work continuously without being asked. Its model remains free at the point of use, which is precisely why the new notices read as a turning point. A product that costs nothing to try still costs something to run, and for the first time DeepSeek is telling its users what that cost looks like from the inside.

What the shortage means is now the argument running through the coverage. The Central News Agency framed the limits as evidence of strain on the company's compute supply — and in China's AI industry, questions about compute supply are rarely only about servers. Advanced chips sit behind export controls, domestic alternatives are only now reaching mass production, and a consumer-facing service with a vast free user base is the kind of operation where demand outruns hardware fastest. DeepSeek has not published an explanation beyond the notices themselves, and users are left to read tea leaves from error messages.

There is a practical edge to the annoyance. The seven-day cutoff hits a specific habit: keeping old conversations alive as archives. Users who treated a chat as a notebook — a project they could return to after a week away — now find that context simply gone, and the 40,000-character cap constrains the long-document workflows that made the app a default for students and office workers. The daily quota, meanwhile, converts the casual luxury of "ask it anything" into a budget to be spent.

None of the notices says when things return to normal, and the company has not announced an end date. What is clear is the sequencing: six weeks after topping the global usage charts, the most-watched lab in Chinese AI is in the position of rationing a free product at the moment its audience is largest — asking the users it won over with abundance to live within scarcity, with no promise of when the meter resets.