Chinese artificial intelligence developers DeepSeek and Zhipu AI are moving to design or customize their own inference chips, sharpening a push by China’s leading AI companies to reduce reliance on foreign semiconductor suppliers and gain tighter control over the cost of running large models.
People familiar with the matter said DeepSeek began work on an AI chip project about a year ago and has held discussions with chip design houses, foundries and memory manufacturers. Zhipu AI, meanwhile, is assessing partnerships with domestic chip companies to develop custom processors based on its own model architecture.
The moves reflect a fast-changing priority across the AI industry: the cost of inference, the process of serving user requests after a model has already been trained. While training a large model requires enormous computing resources, inference can become even more expensive over time because it happens continuously, every time a user asks a chatbot a question, generates code, creates an image or calls an AI service through an application.
For companies with mass-market AI products, that means the bill does not end when the model is built. It grows with usage. DeepSeek and Zhipu are now trying to answer a central business question facing large AI developers worldwide: whether to keep paying rising operating costs to outside hardware suppliers, or to own more of the infrastructure that powers their services.
DeepSeek’s chip effort is focused on inference rather than training, according to people familiar with the project. The goal is to lower the cost of serving each user request and reduce dependence on imported high-end chips. Zhipu is taking a different route, exploring collaboration with specialized domestic manufacturers rather than pursuing a wholly internal design from the start.
Both strategies point in the same direction. China’s AI companies want more control over their computing resources as demand rises, export restrictions tighten and the economics of large-scale AI deployment become more difficult to manage.
Why inference is becoming the cost center
Inference has become the daily operating burden of the AI business. Training a large language model is expensive, but it is usually a concentrated phase. Inference is permanent. Every user query consumes computing power, memory bandwidth and energy. As services scale from research demonstrations to products used by millions of people, inference spending can exceed training costs.
Industry assessments often place inference at the center of lifetime AI system costs, with some estimates suggesting it can account for the large majority of total spending over a model’s useful life. The exact figure depends on model size, user volume, hardware efficiency, energy costs and pricing strategy. But the direction is clear: the company that controls inference costs gains an important advantage.
That is why DeepSeek’s project focuses on chips built for serving models rather than training them. Inference chips do not need to perform every task required during model development. They can be optimized for high-throughput, lower-latency deployment, with particular attention to memory bandwidth, power efficiency and compatibility with a company’s model architecture.
For DeepSeek, lowering per-user service cost has become a strategic priority. The company has drawn attention in China and overseas for developing large models with a focus on efficiency. Its leadership has also placed growing emphasis on reducing reliance on imported high-end processors, especially as access to advanced foreign AI chips remains constrained.
DeepSeek had previously adapted parts of its model training process from NVIDIA’s H800 chips to Huawei’s Ascend processors, according to the information provided. That shift already reflected the company’s effort to work within China’s changing chip environment. Designing its own inference chip would take that effort further, potentially allowing DeepSeek to align hardware more closely with its software and product needs.
DeepSeek’s funding and computing ambitions
In June, DeepSeek completed its first funding round, raising about 51 billion yuan, according to the article’s source material. The financing valued the company at between $52 billion and $59 billion. The company said proceeds would be used to expand local computing centers and advance its in-house chip program.
The size of the financing underlines how capital-intensive the AI race has become. Building frontier models, operating consumer products, maintaining enterprise services and developing chips all require heavy spending. For an AI company, the move into semiconductors is not a side project. It is a long-term infrastructure decision that can shape margins, supply security and competitive position for years.
DeepSeek’s chip work reportedly began roughly one year ago. Since then, the company has spoken with design firms, foundries and memory producers. Those talks are important because even companies designing their own chips still depend on a complex manufacturing chain. A successful inference processor requires more than a chip blueprint. It needs fabrication capacity, packaging, memory integration, software tooling and reliable deployment inside data centers.
Memory is especially important for AI inference. Large models move enormous amounts of data between compute units and memory. Performance often depends not only on raw processing power, but also on the speed and efficiency with which model weights and activations can be accessed. That makes high-bandwidth memory and advanced packaging critical areas for customized AI hardware.
If DeepSeek’s project progresses, it could bring new business to Chinese semiconductor design firms, memory suppliers and packaging specialists. But the longer-term effect could be more complicated. If major AI developers eventually run services on self-designed chips, outside hardware vendors may gain near-term revenue from collaboration while facing reduced dependence over time.
Zhipu seeks partners after demand surge
Zhipu AI is pursuing a more partnership-driven model. The company is assessing cooperation with domestic chip firms to create customized processors suited to its own AI models. Rather than building every part of the chip program internally, Zhipu is looking at how local semiconductor specialists can turn its model architecture into hardware better suited for inference.
The strategy follows a period of pressure and expansion for the company. Zhipu, which was listed in Hong Kong earlier this year, had a market value that once exceeded 1 trillion Hong Kong dollars, according to the supplied material. The company also recorded a 29.58 billion yuan loss in 2024 and an additional 23.58 billion yuan loss in the first half of 2025.
Those losses show the financial strain that can accompany rapid AI growth. Large AI developers must spend heavily on talent, research, model training, data, servers and user access. If services are priced too low, rapid adoption can deepen losses rather than improve financial performance. That is one reason AI companies are increasingly focused on pricing, computing partnerships and infrastructure ownership.
After Zhipu released its GLM-5 model, global demand for its services increased. The company responded by raising prices for its code-generation service by 30 percent and launching a program to recruit computing partners, including local chip suppliers.
The price increase highlights the tension facing AI companies. They want to grow usage, but heavy usage creates rising inference costs. If the service is popular, the cost of serving it can grow quickly. Raising prices can help, but pricing power depends on product quality, customer loyalty and competition. Custom chips offer another possible path: reduce the underlying cost rather than relying only on higher prices.
Zhipu’s search for computing partners may also help broaden China’s domestic AI hardware ecosystem. Potential partners could include chip design companies, memory suppliers and firms specializing in integration and packaging. If Zhipu can align its model architecture with domestic processors, it may reduce exposure to supply limits and improve the economics of serving users at scale.
A wider move toward vertical integration
The trend is not limited to China. Major AI and cloud companies in the United States have also developed proprietary chips to manage costs, improve performance and secure hardware supply. Google has long used its Tensor Processing Units for AI workloads. Amazon has developed its Trainium and Inferentia chips. Microsoft has introduced its Maia AI accelerator. OpenAI has also been reported to be exploring custom chip development as part of a broader effort to secure computing capacity.
The common theme is vertical integration. AI developers are no longer treating chips as interchangeable infrastructure purchased only from outside suppliers. Instead, they increasingly see hardware, software and model architecture as parts of one system. A chip designed for a specific model or workload can reduce wasted computing, improve energy efficiency and provide more predictable supply.
For traders, this shift matters because it changes the competitive map for semiconductor suppliers. Companies that dominate general-purpose AI accelerators may face new pressure if their largest customers build internal alternatives. At the same time, custom chips are difficult to design, expensive to manufacture and risky to deploy. In the near term, outside suppliers can still benefit from rising AI demand, even as customers try to reduce dependence on them.
The prospect of Chinese AI developers moving deeper into chip design has also sharpened attention on NVIDIA’s position in China. The company has been a leading global supplier of AI accelerators, but U.S. export controls have limited the sale of its most advanced products to Chinese customers. Those restrictions have encouraged domestic substitution and created openings for Chinese chipmakers.
Huawei has become one of the main beneficiaries of this shift. Its Ascend processors are increasingly used as Chinese technology companies adapt to a more restricted supply environment. DeepSeek’s previous move to adapt training from NVIDIA’s H800 processors to Huawei’s Ascend platform shows how quickly AI developers have been adjusting.
The change does not mean foreign suppliers are irrelevant. Many AI companies still depend on established software ecosystems, mature hardware and proven performance. But it does mean Chinese AI developers are preparing for a future in which they cannot assume steady access to the most advanced imported chips.
What it means for China’s chip sector
For China’s semiconductor industry, the move by DeepSeek and Zhipu could provide a significant boost. Customized inference chips need advanced design capability, strong memory systems, packaging expertise and data-center integration. Domestic firms that can support those needs may see new revenue opportunities.
The benefits could extend beyond a single chip project. If leading AI developers work closely with local hardware companies, they can help define performance requirements and provide real deployment feedback. That feedback loop can accelerate product improvement. In a rapidly evolving field such as AI, the ability to test chips against large real-world workloads is valuable.
Memory companies may also gain from the trend. Inference at scale requires moving data quickly and efficiently, which increases the importance of high-bandwidth memory and next-generation memory technologies. Even a well-designed processor can be held back if memory performance is insufficient.
Foundries and packaging firms may see additional opportunities as well, although advanced manufacturing remains one of the most difficult parts of the semiconductor chain. Designing a chip is only the beginning. Producing it reliably, at scale and at a competitive cost is a separate challenge.
There is also a possible downside for some suppliers. If AI developers eventually shift fully to self-designed hardware, third-party chipmakers could lose pricing power. The first stage of the trend may create partnerships and orders. The later stage could reduce dependence on outside vendors if internal chips perform well enough.
The risks of building custom chips
The path to chip independence is expensive and uncertain. Semiconductor research and development cycles typically take several years. Projects often require multi-billion-yuan spending before a commercial product is ready. Even then, success is not guaranteed.
AI models also change quickly. A chip designed around one model architecture can become less useful if the company later changes its technical direction. That creates the risk of costly redesigns. Industry observers often point to previous cases in which large technology companies had to slow, cancel or restart internal chip programs after model strategies shifted.
Meta’s canceled self-developed chip effort has been cited as one example of the challenge. When model requirements change, hardware plans can quickly become outdated. Unlike software, chips cannot be rewritten overnight. A design mistake or strategic change can cost years.
Software support is another major hurdle. AI chips need developer tools, model compilers, runtime systems and integration with existing AI frameworks. Even a powerful processor can struggle if it is difficult to use. NVIDIA’s strength has not come only from hardware performance, but also from its software ecosystem. Any company developing custom chips must address that gap.
Data-center reliability is equally important. Inference services need stable performance around the clock. A chip used in a consumer or enterprise AI product must handle large traffic swings, long operating periods and strict service requirements. A less mature chip may create operational risks even if it looks efficient in testing.
The new equation for AI companies
The rapid expansion of AI services has revived a long-running infrastructure debate in technology: whether to continue renting or buying capacity from outside suppliers, or to own more of the stack. Cloud computing raised this question years ago. AI has made it more urgent because the cost of computation is central to the product itself.
For Chinese large-scale AI developers, the equation is now shaped by both economics and geopolitics. Inference costs are rising as usage grows. Access to imported chips is less certain. Domestic chip alternatives are improving but still developing. Traders are watching how these forces affect company margins, hardware demand and the balance of power across the AI supply chain.
DeepSeek and Zhipu are taking different paths, but both are responding to the same pressure. DeepSeek is pursuing a more internal chip design strategy aimed at reducing per-user costs and dependence on imported processors. Zhipu is looking to domestic partners to create processors that fit its model architecture and growing service demand.
Neither approach will deliver immediate independence. Chip development takes time, money and repeated testing. But the direction is clear. AI companies that once competed mainly on model quality and user growth are now competing on infrastructure control.
If these efforts succeed, they could reshape China’s AI hardware market and strengthen local semiconductor firms. If they fall short, DeepSeek and Zhipu may still remain dependent on outside suppliers while carrying the cost of ambitious chip programs. Either way, the strategic pivot shows that the next phase of the AI race will not be decided only by model performance. It will also be decided by who can run those models at the lowest cost, with the most reliable supply, and with the greatest control over the machines underneath.
Explore how AI reshapes trading: master automated strategies with our guide to AI copy trading on Toobit today.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
