1. DeepSeek has begun restricting conversations and document reading on its website and app because of insufficient computing capacity [para. 1]. Since Nov. 15, long conversations inactive for 72 hours can no longer be processed, and document reading is capped at the first 20 pages or 10,000 tokens [para. 2]. Users complained on Xiaohongshu and offered to pay for a subscription, but DeepSeek charges only via API and has no consumer paid tier [para. 3].
2. The squeeze comes as Chinese AI labs slow foundational-model advances and emphasize agents, software tools and enterprise partnerships [para. 4]. Yao Shuqiao of Haitong International said DeepSeek has fewer computing resources than major cloud providers and must allocate them efficiently, giving paying customers priority [para. 6]; he expects computing power to remain tight short- and medium-term, with demand growing exponentially by 2027 and a broader cutoff risk tied to Nvidia-related developments [para. 7].
3. Share declines reflect the concern: Z.AI closed at HK$335 on Nov. 17, down 32.9% over four trading days to a six-month low; MiniMax closed at HK$116, down 35.6% [para. 8]. Z.AI's last flagship model dates to mid-August, and MiniMax has yet to release its roughly 3-trillion-parameter Pro model [para. 9]. Moonshot AI suspended new Kimi K3 subscriptions in July and apologized for rate limits after launching K3.1 Pro in late October [para. 10]; DeepSeek released V4.1-Flash on Sept. 10 but no Pro version [para. 11]. With differences among models narrowing, developers increasingly compete on ecosystem completeness [para. 13].
4. A key focus is the "harness"—the operating and control layer around a model that lets AI agents complete complex, multistep tasks [para. 15]. AI expert Xiao Rui compared the model to a horse and the harness to guiding equipment; an agent is a model plus a harness [para. 16][para. 17]. DeepSeek introduced DeepSeek Harness on Aug. 13 with an "everything is a plugin" approach and MIT-licensed source code, making it more accommodating of third-party models and lowering barriers to participation [para. 18][para. 19][para. 20].
5. A desktop version launched on Sept. 29; by Nov. 4, daily active users reached about 1.5 million, a 545% increase, with daily token consumption around 92 trillion [para. 21][para. 22]. A harness-optimized smaller model can deliver close to 90% of a larger model's capabilities at about 4% of the cost in routine agent tasks [para. 23]. Xiao characterized harness development as an engineering challenge rather than a computing-capacity challenge, making it easier for Chinese developers to catch up with U.S. competitors [para. 24].
6. DeepSeek is also widening its paid API services [para. 26]. On Nov. 13 it announced application APIs covering online search, deeper-reasoning search, file input and model safety, plus underlying tools such as text and image search, webpage retrieval, text recognition, speech recognition and synthesis [para. 26]. Fees are 5 yuan per 1,000 calls for text/image searches and webpage retrieval and 1.5 yuan per 1,000 pages for text recognition [para. 28]. The APIs attract traffic through the open-source ecosystem and support monetization [para. 29].
7. Enterprise business is another central growth route: Chinese models accounted for 89% of U.S. enterprise spending on open-source models, with Kimi most popular, according to a Menlo Ventures report [para. 31]. Moonshot AI's Kimi K3, trained with 2.8 trillion parameters, is focused on expanding its enterprise customer base [para. 32][para. 33]; vertical solutions now cover legal, financial, game-development and creative work, and enterprise customers generate more than 70% of Kimi's revenue [para. 34][para. 35]. On Nov. 5, Moonshot launched Kimi Hosted Agents, a cloud-hosted environment for running agents [para. 36]. Z.AI announced cooperation with China Mobile on token technology, model services, product development and industry solutions [para. 37]. As Yao put it, customers are paying not for the model itself, but for the results it produces [para. 38].
AI generated, for reference only