DeepSeek users began reporting new restrictions on the company's app and website on Sunday, as one of China's best-known AI labs started rationing access to its own consumer chat service. Some users received a notice saying that long conversations inactive for more than seven days could no longer be resumed because of insufficient computing capacity, according to TechNode's report on the changes.
The limits are among the most visible signs yet of how hard it is for a breakout AI service to keep up with its own success — and they landed just as China's biggest internet companies made dueling flagship model launches this week. For ongoing coverage of the compute squeeze, follow our latest AI developments.
What users are seeing
The most consequential change affects long-running work. Conversations that have been inactive for more than seven days can no longer be picked up where they left off, with DeepSeek citing insufficient computing capacity as the reason, TechNode reported. Users juggling multiple sessions — the typical pattern for anyone using a chatbot as an ongoing work tool — are the ones most likely to lose access to long-running threads.
Users also reported that individual prompts were capped at roughly 40,000 Chinese characters, a ceiling that affects people pasting in long documents or large code files, according to the TechNode report, which cited Chinese tech outlet IT Home.
DeepSeek has not announced when the limits might be lifted, IT Home reported on Monday. That leaves users with no stated timeline, and it means anyone relying on the free consumer service for archival-style workflows — keeping a months-long conversation as a working memory — now has to plan around losing threads that sit idle for a week.
A status page that shows the strain
The restrictions did not come out of nowhere. DeepSeek's own status page has recorded eight incidents since October affecting its chat service or its V4 Pro and V4.1 Flash APIs, TechNode reported. The pattern suggests an infrastructure layer under sustained pressure rather than a one-off outage: repeated degradation across both the consumer product and the paid API products that developers build on.
The reported limits affect consumer chat most directly. But capacity problems that spill into status-page incidents have a habit of reaching paying customers too, and API reliability is the currency on which developer trust is built. A lab that wants its models embedded in other companies' products cannot let its endpoints degrade eight times in six weeks without consequence.
Why capacity has become the binding constraint
DeepSeek rode a wave of extraordinary attention after its low-cost reasoning models turned the lab into a household name in the global AI race, and its consumer service has been absorbing heavy demand ever since. Serving that demand requires clusters of advanced accelerators, and the economics are unforgiving: every idle-but-preserved conversation consumes storage and scheduling overhead, while every active conversation consumes compute that could serve new requests.
The company's own notice frames the change as a capacity issue, and it has not publicly connected the restrictions to any external supply constraint. What can be said without speculation is that the change optimizes for the present: compute that went to maintaining dormant conversations now goes to serving active ones.
The move also arrives during a moment of maximum competitive pressure. Alibaba launched its flagship Qwen4-Max on Sunday with heavy price-cutting claims, and Tencent followed on Monday with Hy4 flash, a budget multimodal model aimed at the same everyday-work market. Rationing consumer access while rivals are advertising abundance is a delicate look for any lab — and it underlines that model quality is only half of the AI product; the other half is having enough machines to serve it.
What it means for users
For now, the practical guidance is simple. Conversations on DeepSeek's free chat that matter to you should be exported or summarized before they sit idle for a week, because there is currently no guarantee they can be resumed. Heavy users working with long documents should watch the roughly 40,000-character prompt ceiling and split inputs accordingly. And developers using the V4 Pro or V4.1 Flash APIs should treat the recent incident record as a prompt to build retry logic and fallbacks into their integrations.
Whether the limits are a temporary squeeze or a lasting design change will depend on factors DeepSeek has not laid out publicly: new capacity coming online, demand growth, and how the lab chooses to balance free consumers against paying API customers. What the past two days have made clear is that even celebrated, efficiency-focused labs cannot escape the same arithmetic as everyone else — at some point, demand outruns the machines.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →