DeepSeek’s computing crunch pushes China’s AI developers beyond models
As computing bottlenecks force DeepSeek to restrict user usage, China’s AI race is entering a pragmatic new phase: less about building bigger models, and more about monetising agents and enterprise tools.
20 Nov 2026
Economy
(By Caixin journalist Yang Zirui)
DeepSeek has begun restricting conversations and document reading on its website and app because of insufficient computing capacity, exposing a constraint that is pushing China’s leading AI developers to seek growth beyond large models.
Since 15 November, individual users have encountered company notices saying that long conversations inactive for 72 hours can no longer be processed by DeepSeek because of a shortage of computing resources. When processing long-form documents, DeepSeek has also warned that it can read only the first 20 pages of a document or 10,000 tokens, citing the same reason.
The restrictions have drawn complaints in the comments beneath the latest post on DeepSeek’s official Xiaohongshu social media account. Some users asked the company to remove the limits, saying that they would be willing to pay for a subscription. Unlike competing services such as Doubao and Kimi, DeepSeek does not offer a paid membership for individual consumers, charging instead for usage through its application programming interface (API) services.
The computing squeeze comes as Chinese AI laboratories slow the pace of foundational model advances and put more emphasis on agents, software tools and enterprise partnerships to find new sources of growth. Their emerging contest is less about topping performance rankings than about delivering usable results — and generating revenue without relying solely on ever-larger models.
Growth constraint
DeepSeek has fewer computing resources than major cloud providers and needs to allocate them efficiently, giving paying customers priority for the most effective service, Yao Shuqiao, chief AI analyst at Haitong International, told Caixin. The usage restrictions allow the company to differentiate how it supplies those resources, Yao said.
Yao expects computing power for Chinese models to remain tight in the short- and medium-term. As products improve and usage rises, with industrial applications entering a more demanding phase in 2027, demand for computing resources is expected to grow exponentially, Yao said, adding that recent overseas developments involving Nvidia Corp. could also increase the risk of a broader cutoff of China’s access to advanced computing chips.
Concern about that risk has already been reflected, to some extent, in the shares of two Hong Kong-listed Chinese model developers, Z.AI Co. Ltd. and MiniMax Group Inc. Z.AI closed at HK$335 (US$43) on 17 November, down 5.3% that day and 32.9% over four trading days, reaching its lowest level in six months. MiniMax closed at HK$116, falling 7.1% that day and 35.6% over four trading days, after sliding below its previous post-listing low.
Z.AI’s most recent flagship model release dates to mid-August. MiniMax has yet to release the Pro model it has described as being trained with about 3 trillion parameters.
DeepSeek is not alone in facing computing pressures. Moonshot AI abruptly suspended new subscriptions when it introduced Kimi K3 in July because of computing power shortages. After launching Kimi K3.1 Pro in late October, it apologised to users for rate limits and unnecessary consumption of usage allowances associated with the new model.
DeepSeek released DeepSeek-V4.1-Flash on 10 September and updated it on 17 October, but has not introduced a V4.1 Pro version. When Moonshot AI released Kimi K3.1 Pro on 24 October, it described it as the first model in the K3.1 series. So far, Moonshot AI has not released an upgraded version from K3.1 Pro.
As model releases slow, Chinese developers are expanding their revenue opportunities through agent-control frameworks, application tools and enterprise cooperation.
The pace of improvement in Chinese foundational models has slowed considerably, narrowing the differences between them, Yao said. Developers are increasingly competing over the completeness of their ecosystems rather than concentrating primarily on improvements to model capabilities.
Making models work
One focus is the harness: the operating and control layer surrounding a model that helps it reliably complete complex, multistep tasks through AI agents.
Xiao Rui, an AI expert, compared the model to a horse and the harness to the equipment people use to guide and constrain it. In AI, the term encompasses the components outside the model that make it usable, he told Caixin.
The industry often describes an agent as a model plus a harness. Yao said that the harness establishes mechanisms that allow a model to keep executing tasks, turning its capabilities into automated productivity.
DeepSeek introduced its harness on 13 August alongside the release of DeepSeek-V4-Pro. DeepSeek Harness adopts what the company described as an “everything is a plugin” approach: the agent’s capabilities are assembled from interchangeable components that can be reorganised. DeepSeek also made the harness’s source code available under the MIT licence.
That design has attracted attention because, unlike frameworks primarily optimised for their developers’ own models, DeepSeek’s is more accommodating of third-party models, Xiao said.
Yao said that DeepSeek Harness supplies open-source infrastructure for running agents, lowering barriers to participation. It can also be jointly optimised with DeepSeek models for functions such as tool use and context management, strengthening their competitiveness while drawing developer community traffic to the company.

Get the ThinkChina Weekly Newsletter
Insights on China, right in your mailbox. Sign up now.
By subscribing, I agree to SPH Media's Terms and Conditions and Privacy Policy.
DeepSeek has updated its harness to support each new model release. On 29 September, it introduced a desktop version for DeepSeek Harness with functions for office work and software development.
According to the company, DeepSeek Harness’ daily active users reached about 1.5 million by 4 November, a 545% increase from the launch of its desktop version, with daily token consumption standing at about 92 trillion.
Efficiency is an important reason behind Chinese model developers’ increasing bet on harnesses. Yao said that a smaller model optimised through a harness can, in some routine agent tasks, deliver close to 90% of a larger model’s capabilities at about 4% of the cost. Changes to self-verification, execution tracing and context optimisation alone can significantly improve coding agent performance, Yao added.
Xiao characterised harness development principally as an engineering challenge rather than a computing capacity challenge. That makes it easier for Chinese developers to catch up with US competitors than in foundational models, Xiao said.
Building a commercial ecosystem
DeepSeek is also widening the range of its API services it sells to developers. On 13 November, the company announced application APIs covering capabilities including online search, search with deeper reasoning, file input and model safety. It also offered access to underlying functions such as text and image search, webpage retrieval, text recognition, speech recognition and speech synthesis.
The company said that developers could use the suite to build complex conversational functions into their own applications.
The services create additional sources of revenue. Beyond token charges, DeepSeek levies fees for the underlying tools. For example, text or image searches and webpage retrieval each cost 5 RMB (US$0.70) per 1,000 calls, while text recognition costs 1.5 RMB per 1,000 pages.
In the agent era, companies want to secure users not only for foundational models but also for their ecosystems and toolchains, Yao said. DeepSeek’s application APIs are intended to attract traffic through its open-source ecosystem and make developers more likely to stay, supporting both long-term ecosystem development and monetisation.
Pursuing enterprise customers
Enterprise business is becoming another central growth route. According to a Menlo Ventures report released in November, Chinese models accounted for 89% of US enterprise spending on open-source models, with Kimi models being most popular among them.
Kimi K3, released in July and trained with 2.8 trillion parameters, was the largest open-source model by parameter count at the time and became a representative of the frontier capabilities of Chinese models.
Sun Wei, chief AI analyst at market research firm Counterpoint, told Caixin that Kimi K3’s customised open-weight licence, which sets commercial use conditions, and its enormous size mean that it is currently more focused on expanding its enterprise customer base.
Since K3’s release, Moonshot AI has introduced a Kimi model for the legal sector, a financial industry solution and solutions for game development and creative work. According to the company, enterprise users of its models include legal AI startups, law firms, financial institutions, game developers and fintech companies.
On 10 September, Moonshot AI launched its Kimi enterprise partner programme, working with information technology service providers and systems integrators to build teams of forward-deployed engineers who help businesses put AI into their core operations. Caixin learned from sources with knowledge of the matter that enterprise customers currently account for more than 70% of Kimi’s revenue.
Moonshot AI extended the programme on 5 November with Kimi Hosted Agents. The company said that the new service supplies the harness needed to run agents through the cloud, allowing businesses and developers to obtain an around-the-clock hosted environment without building and maintaining the framework themselves. It supports long-running tasks and collaboration among multiple agents.
Z.AI, meanwhile, recently announced a cooperation agreement with China Mobile Ltd. The companies said that they would deepen strategic cooperation in token technology, model services, product development and industry solutions.
For Yao, these moves reflect a more pragmatic phase for China’s AI laboratories. Companies that once competed over benchmark scores are increasingly confronting a different purchasing decision: customers are not paying for the model itself, but for the results it produces.
This article was first published by Caixin Global as “In Depth: DeepSeek’s Computing Crunch Pushes China’s AI Developers Beyond Models”. Caixin Global is one of the most respected sources for macroeconomic, financial and business news and information about China.
Related: China’s AI battle moves from chatbots to agents | Cheap, fast and everywhere: Why China leads AI adoption
Popular This Month

Get the ThinkChina Weekly Newsletter
Insights on China, right in your mailbox. Sign up now.
By subscribing, I agree to SPH Media's Terms and Conditions and Privacy Policy.