Reports indicate that DeepSeek and Zhipu AI are beginning the development of their own specialized chips for inference operations, following the example set by giants such as OpenAI and Anthropic.
Reports indicate that DeepSeek and Zhipu AI are beginning the development of their own specialized chips for inference operations, following the example set by giants such as OpenAI and Anthropic.
There is a growing trend among companies creating AI models to quietly invest in custom semiconductors. According to sources cited by Reuters and The Information, DeepSeek and Zhipu AI are working on proprietary chips tailored to the inference workloads of their models.
The DeepSeek chip project reportedly started about a year ago. The company recently expanded its chip design team and began collaborating with chip designers, foundries, and memory suppliers. These efforts resemble OpenAI's Jalapeño project, which was co-developed with Broadcom and is already undergoing lab testing with plans for deployment by the end of the year.
Zhipu AI, whose GLM series models are seeing increasing demand, is also evaluating the possibility of using custom silicon to reduce inference costs as user token consumption grows.
The economic arguments for this approach are quite compelling. The cost of running an advanced model accumulates with every API call, every agent step, and every processed token. Specialized inference chips, designed specifically for fixed model architectures, can significantly lower the cost per token compared to general-purpose GPUs.
Although Nvidia H100 and B200 chips remain the industry standard for training, for a company operating continuously with the same large model at scale, the mathematics increasingly favors specialization.
A broader trend is evident: Google has its TPUs, Amazon uses Trainium and Inferentia, Microsoft uses Maia, and Meta uses MTIA. In China, players like Huawei with Ascend, Baidu with Kunlun, and Alibaba with Zhenwu also have their own solutions. Now, the model developers themselves—who are the actual builders and users of this infrastructure—are striving for direct control over the silicon underpinning their models.
For Chinese model developers, the calculation extends beyond pure economics. US export controls restrict access to advanced Nvidia chips, and modern AI chips depend on complex global supply chains, including foundry capacity and HBM memory. Self-developed silicon provides a path toward greater infrastructure independence, even if the most advanced manufacturing processes remain restricted.
However, not all custom chips are the same. A full-stack development approach, where a company defines its own architecture, compiler, software stack, and data center systems, requires years of investment. Lighter options for custom chips, where the model company defines the requirements and core architecture but engages an existing chip firm for implementation, offer a faster route to silicon with fewer risks.
A new artificial intelligence model developed in China is attracting significant attention in the technology sector. According to a Reuters report, the GLM-5.2 model from the startup Z.ai demonstrates performance comparable to systems from the United States, but at a significantly lower cost.
Following the emergence of DeepSeek, which showed the competitiveness of cheaper models, other Chinese companies have accelerated their launches. GLM-5.2 is the latest example of this trend. The model, created in Beijing, gained recognition for its ability to perform complex tasks with minimal commands and function as an AI agent, making intermediate decisions in long processes, which reduces the need for constant user intervention.
In Silicon Valley, the system has begun appearing actively on platforms such as OpenRouter and has reached a level of developer usage exceeding that of models from Anthropic. There is now a Chinese open-weight model comparable to current OpenAI and Anthropic models.
David Sachs, a former head of AI policy in the US government during the Donald Trump administration, told Reuters that GLM-5.2 is 'just slightly below Opus 4.8 (from Anthropic) and on par with GPT 5.5 (from OpenAI)'.
The interest in GLM-5.2 is directly linked to the combination of technical power and reduced price. Compared to advanced closed models, operational costs can be about one-sixteenth of the cost.
Key features of this model include: high efficiency in programming and logical reasoning; the ability to act as an AI agent based on simple commands; an open architecture that is easy to implement; and a substantially lower cost compared to direct competitors.
According to Artificial Analysis rankings, GLM-5.2 ranks fifth among large language models. In Code Arena, the model also secured second place in frontend development tasks, such as creating websites and applications.
Despite the technological progress, global adoption has not yet been uniform. The main point of caution remains data security, especially for companies in the US and in regulated sectors.
Nevertheless, part of the market is already testing alternatives deployed on private cloud or internal servers, seeking a balance between risk and cost. Simultaneously, the development of these models expands China's presence in the global AI market and puts pressure on North American competitors, as cost, performance, and open access begin to matter as much as the origin of the technology.
DeepSeek is changing its strategic focus: instead of remaining a developer of lightweight artificial intelligence models, it is moving towards a more intensive approach based on heavy infrastructure. This shift is confirmed by the active recruitment in the newly established Harness division and plans to build its own computing power.
The DeepSeek Harness team, led by Qiu Tianyi, began online recruitment of researchers, engineers, and product managers for Harness on June 21. This underscores the company's commitment to developing agent infrastructure. Harness, which represents auxiliary systems that define the tools available to AI models, the resources it can use, and the methods of information flow between sub-agents, has become one of the most sought-after areas in the AI sector in 2026.
The DeepSeek Harness team, formed in March 2025, operates on the core formula: Model plus Harness equals Agent, focusing on transforming model capabilities into ready-made agent products. Since late 2025, the team has been continuously expanding, especially after DeepSeek V3.2 emphasized improved reasoning and agent capabilities.
In addition to attracting talent, DeepSeek is considering building its own data centers in Inner Mongolia. This is a rare move among Chinese companies, which mostly prefer to rent computing power. Zhou Zhengan, a board member of IDC China, noted that leading model developers are increasingly favoring their own clusters to ensure long-term cost control and gain the benefits of full-stack customization, despite significant initial investments. This strategy will allow DeepSeek to optimize cooling and planning systems specifically for the architecture of its proprietary model.
This strategic pivot came after DeepSeek's Series A funding round of 51 billion yuan, which included strategic investors such as Tencent, JD.com, NetEase, and CATL. Industry observers suggest that the inclusion of CATL, the world's largest battery manufacturer, may indicate DeepSeek's intention to achieve synergy between computing and electricity, potentially using solar and energy storage solutions for its data center operations.
As major Chinese AI model developers diverge in their strategic directions—Zhipu AI betting on AI coding, and Alibaba focusing on the full AI stack—DeepSeek's focus on heavy infrastructure and agent engineering represents a distinct third path capable of reshaping the competitive landscape.
