AI News

DeepSeek Imposes Usage Limits on App and Web Versions Due to Compute Shortage

36Kr · AIUpdated Nov 16 21:10

AI summary

Due to tight supply of compute resources, DeepSeek has introduced usage limits for its App and web-based chat services for ordinary users, including no longer supporting long conversations that have existed for more than 7 days without active interaction, a cap of about 40,000 characters per single input, and limits on daily conversation quotas. Industry analysis suggests the move is intended to reclaim KV cache resources occupied by idle sessions and prioritize the inference needs of active users. According to DeepSeek's official status page, there have been multiple service interruptions since October, and from August to October the availability of the V4.1 Flash API was 99.46%, while the V4 Pro API was 99.74%.

Why read this

AI score 65

The original article outlines the specific rules of the limits and the compute pressure behind them, so readers can understand how the strain on inference resources is passed on to ordinary users.

Topics

Tags

Body · AI translation

On November 16, Science and Technology Innovation Board Daily reported that due to tight supply of computing power resources, DeepSeek has introduced usage quota measures for the App and web-based chat services for ordinary users.

According to feedback from multiple users, DeepSeek has shown pop-up notices stating that, constrained by computing power, it will no longer support long conversations that have existed for more than 7 days without active interaction; such idle sessions will no longer be able to continue. At the same time, the upper limit for a single text input has been lowered, with the current single-input text limit at approximately 40,000 characters. In addition, the daily conversation quota will be limited; when a user's daily usage reaches the cap, a notice will appear saying "Due to server congestion, the usage quota has reached its limit."

Industry analysis believes the adjustment is intended to reclaim KV cache resources occupied by idle sessions and prioritize inference needs of active users. KV cache is intermediate data retained in GPU memory by large models to accelerate conversations and preserve context; long conversations continuously accumulate KV cache, and long-idle sessions continuously occupy scarce GPU memory resources.

DeepSeek's computing power has remained under pressure, with multiple service interruptions since October and 8 incidents recorded on its status page. According to the incident records on DeepSeek's official status page, they occurred on October 2, October 8, October 12 (twice that day), October 15, October 20, October 27, and November 3, involving two types of anomalies of different severity, and not all were complete outages: most were performance degradation, manifested as slower responses, high generation latency, and occasional request failures, while users could still submit conversations with some probability; a few were partial interruptions.

Among them, the October 15 incident lasted more than 4 hours, with the consumer web and App chat services and the V4 Pro and V4.1 Flash APIs simultaneously affected, making it the incident with the largest impact and longest duration in October. Afterwards, DeepSeek launched a credit compensation plan for API users and said it would "resolutely prevent similar problems from happening again." Most of the other incidents lasted from several minutes to tens of minutes, and some affected only auxiliary search tools without impacting basic conversation capabilities.

According to DeepSeek's official status page, from August to October, the availability of the V4.1 Flash API was 99.46%, and that of the V4 Pro API was 99.74%. Some media converted this into time for an intuitive estimate: with 8,760 hours in a year, 99.46% means users on average cannot use Flash for 47.3 hours per year; for V4 Pro, the corresponding figure is about 22.8 hours.

On November 13, DeepSeek officially opened the full-suite application API. By enabling the corresponding parameters in the existing API, users can obtain conversation features such as deep-thinking search and file input; voice and related underlying capabilities have also been opened as independent interfaces (including OCR, ASR, TTS, etc.), supporting separate or combined calls. Developers can easily implement complex conversation features in their own applications, or even replicate a complete DeepSeek App 1:1.

Against the backdrop of pressure on computing power supply, DeepSeek has successively introduced multiple technical, pricing, and scheduling measures to actively regulate computing load.

In addition to reducing KV cache pressure and saving expensive HBM memory, DeepSeek has enabled peak-valley time-of-use pricing for API services in pricing and scheduling, with off-peak prices at half of peak-period prices, guiding developers to make calls during off-peak hours and smoothing peak traffic. In terms of scheduling mechanisms, the API side maintains independent concurrency limits, with deepseek-flash and v4-pro set at upper concurrency limits of 2,500 and 500 accounts respectively, isolated from the consumer service resource pool.

Science and Technology Daily reported in October that DeepSeek plans to build a GW-level data center in Ulanqab, deploying at least 160,000 domestic AI chips. If completed, it will become one of China's leading domestic AI inference computing clusters, with the entire cluster expected to gradually go into production between the end of 2027 and early 2028.

Separately, according to China Newsweek, in October DeepSeek completed a second round of financing of 105.5 billion yuan, with a post-investment valuation of 615.5 billion yuan, with nearly 30 institutions including Tencent and CATL participating. DeepSeek has not issued an official announcement confirming the financing amount, valuation, or use of funds. The market believes the primary use of this round of funds is computing power infrastructure construction, while also being used for large model training, expansion of the R&D team, and support for model iteration.

On the weekly call rankings of API aggregation platform OpenRouter, for the statistical period from October 19 to 25, the V4.1-Flash-1017 version updated by DeepSeek on October 17 ranked first with weekly call volume of 33.6 trillion tokens, an increase of 801% month-on-month; the old version V4.1 Flash simultaneously ranked second with 18 trillion tokens, and all top five positions on the ranking for that period were occupied by domestic large models.

Tight computing power supply is not a problem unique to DeepSeek; leading AI vendors at home and abroad generally face pressure on inference and training computing power resources, and the industry has also stepped up investment in intelligent computing centers and computing hardware.

Executives of multiple domestic technology companies mentioned at earnings communication meetings that computing power utilization is at a high level and hardware supply is tight, with capital expenditure mainly directed toward AI infrastructure, server and chip procurement, for their own large model training, inference, and cloud customers' AI needs.

Overseas, companies such as OpenAI and Anthropic are likewise continuously securing long-term computing power resources, locking in chip and data center resources for the coming years through long-term large computing contracts to support model iteration and growth in user calls. OpenAI signed a 7-year computing power cooperation agreement with AWS, committing to invest $38 billion and acquire hundreds of thousands of Nvidia GPUs in batches, with the cluster deployment period extending beyond 2027; Anthropic disclosed in its IPO prospectus that it has reached a computing power procurement commitment of at least $114.5 billion with Google, with the agreement period running through July 2033.

This article is from the WeChat public account "Jiemian News," author: Shen Xiaoge, published by 36Kr with authorization.

Source

36Kr · AI