DeepSeek has started to impose usage quotas.
On November 16, Science and Technology Innovation Board Daily reported that due to the tight supply of computing power resources, DeepSeek has launched a quota-based usage measure for the conversational services of its App and web version for ordinary users.
According to feedback from multiple users, DeekSeek has popped up a prompt stating that due to computing power constraints, it will no longer support long conversations that have existed for more than 7 days with no active interaction, and such idle sessions cannot be resumed for use. At the same time, the upper limit of single-input text has been lowered, and the current upper limit of single-input text is about 40,000 characters. In addition, the daily conversation quota will also be restricted. When the user's daily usage reaches the upper limit, a prompt will appear saying "Due to busy servers, your usage quota has reached the upper limit".
Industry analysis believes that this adjustment is intended to recycle the KV cache resources occupied by idle sessions, and prioritize ensuring the inference needs of active users. KV cache is the intermediate data that large models reside in GPU video memory to accelerate conversations and retain subsequent context. Long conversations will continuously accumulate KV cache, and long-term idle sessions will continuously occupy scarce video memory resources.
DeepSeek's computing power has been under continuous pressure, and there have been multiple service interruptions since October, with 8 failures recorded on the status page. According to the event records on DeepSeek's official status page, the failures occurred on October 2, October 8, October 12 (twice on that day), October 15, October 20, October 27, and November 3 respectively, including two types of anomalies of different levels, not all of which are complete outages: most of them are performance degradation, manifested as slow response, high generation delay, and occasional request failures, and users still have a probability to submit conversations; a few are partial interruptions.
Among them, the failure on October 15 lasted for more than 4 hours, and the C-end web and App conversation services were restricted synchronously with V4 Pro and V4.1 Flash API. It was also the failure with the largest impact scope and the longest duration in October. Afterwards, DeepSeek launched a credit grant compensation plan for API users, and stated that it will "resolutely put an end to the recurrence of similar problems". Most of the other failures lasted from several minutes to dozens of minutes, and some only affected the search auxiliary tools without affecting the basic conversation capability.
According to DeepSeek's official status page, from August to October, the availability of V4.1 Flash API was 99.46%, and that of V4 Pro API was 99.74%. Some media converted it into an intuitive time estimation. With 8760 hours a year, 99.46% means that users cannot use Flash for an average of 47.3 hours per year; the corresponding figure for V4 Pro is about 22.8 hours.
On November 13, DeepSeek officially opened the full-suite application API. By enabling the corresponding parameters in the original API, you can obtain conversation functions such as deep thinking search and file input; independent interfaces for voice and related underlying capabilities (including OCR, ASR, TTS, etc.) have also been opened, supporting separate or combined calls. Developers can easily implement complex conversation functions in their own applications, and even 1:1 reproduce the complete DeepSeek App.
Against the background of tight computing power supply, DeepSeek has successively launched a number of technical, pricing and scheduling measures to actively adjust the computing power load.
In addition to reducing KV cache pressure and saving expensive HBM video memory, DeepSeek has implemented peak-valley time-of-use pricing for API services in terms of pricing scheduling, with the price in off-peak hours being half of that in peak hours, guiding developers to make off-peak calls to smooth peak traffic. In terms of scheduling mechanism, the API side maintains an independent concurrent limit, with the concurrent upper limits of 2500 and 500 accounts set for deepseek-flash and v4-pro respectively, which are isolated from the C-end service resource pool.
According to a report by Science and Technology Daily in October, DeepSeek plans to build a GW-level data center in Ulanqab, deploying at least 160,000 domestically produced AI chips. If completed, this will become one of the largest domestically produced AI inference computing power clusters in China, and the entire cluster is expected to be put into operation gradually from the end of 2027 to the beginning of 2028.
According to China News Weekly, DeepSeek completed the second round of financing of 1.055 trillion yuan in October, with a post-investment valuation of 6.155 trillion yuan, and nearly 30 institutions including Tencent and CATL participated. DeepSeek's official has not issued an announcement to confirm the financing amount, valuation and capital usage. The market believes that the primary investment of this round of funds is in the construction of computing power infrastructure, and it is also used for large model training, expansion of R&D teams, supporting model iteration, etc.
On the weekly call ranking list of the API aggregation platform OpenRouter, with the statistical period from October 19 to 25, the V4.1-Flash-1017 version updated by DeepSeek on October 17 ranked first on the list with a weekly call volume of 33.6 trillion Tokens, with a month-on-month increase of 801%; the old version V4.1 Flash also ranked second with 18 trillion Tokens, and the top five seats on the list in this period are all occupied by domestic large models.
The tight supply of computing power is not a problem encountered by DeepSeek alone. Leading AI manufacturers at home and abroad are generally facing the situation of pressure on inference and training computing power resources, and the industry has also increased investment in intelligent computing centers and computing power hardware.
The management of many domestic technology companies mentioned in the financial report communication meeting that the utilization rate of computing power is at a high level and the hardware supply is tight, and the capital expenditure is mainly invested in AI infrastructure, servers and chip procurement, which are used for self-owned large model training, inference and AI needs of cloud customers.
Overseas, companies such as OpenAI and Anthropic are also continuously locking in long-term computing power resources, and locking in chip and computer room resources for the next few years through long-term large-value computing power contracts to support model iteration and the growth of user calls. OpenAI has signed a 7-year computing power cooperation agreement with AWS, promising to invest 38 billion US dollars to obtain hundreds of thousands of NVIDIA GPU computing power in batches, and the cluster deployment cycle will cover after 2027; Anthropic disclosed in its IPO prospectus that it has reached a computing power procurement commitment of at least 114.5 billion US dollars with Google, and the agreement period will last until July 2033.
This article is from the WeChat Official Account "Jiemian News", author: Shen Xiaoge, published with authorization from 36Kr.