DeepSeek and Zhipu both develop in-house AI chips amid US export controls and rising import costs for Nvidia accelerators.
Chinese tech giants and AI companies are rushing to develop their own AI training and inference semiconductors. U.S. sanctions have made it difficult for Chinese firms to import NVIDIA's AI chips, which are essential for powering AI systems. While domestic alternatives like Huawei's Ascend chips have been deployed, a surge in AI model usage has created a severe shortage of AI semiconductors.
This has prompted major tech firms like Alibaba and Baidu, as well as AI model developers such as DeepSeek, to pursue in-house AI chip development. Following advancements in memory semiconductors, China's push for technological self-reliance is expanding to AI chips and models. Some analysts suggest this development could accelerate China's AI surge.
**DeepSeek and Zhipu AI Develop In-House AI Chips**
According to Reuters and other foreign media reports from July 7th, Chinese AI firm DeepSeek has begun developing its own AI chips, following behind-the-scenes discussions with Chinese chip design companies, foundries, and memory producers over more than a year.
DeepSeek, the leading Chinese AI company that surprised Silicon Valley and Wall Street last year with its low-cost, high-performance model R1, is advancing from model innovation to developing the semiconductors that power its AI systems.
Zhipu AI, emblematic of China's AI self-reliance initiative, is also exploring in-house AI chip development. According to The Information, Zhipu has held preliminary talks with Chinese semiconductor design firms to develop AI chips optimized for its GLM model. The timing coincides with Zhipu's recent unveiling of GLM-5.2, an AI model that matches the performance of Anthropic and OpenAI's latest offerings. The Information reported: "As AI usage surges, computational resource burdens have grown, prompting Zhipu to explore its own AI chips."
Industry insiders attribute this shift to U.S. export controls on AI chips, which have made it nearly impossible for Chinese firms to secure high-performance NVIDIA hardware. The Chinese government has encouraged domestic AI semiconductor adoption, positioning Huawei's Ascend chips as alternatives. Both DeepSeek and Zhipu previously optimized their models for Huawei's chips.
However, Huawei's chips face criticism for lagging NVIDIA in supply capacity and performance. This prompted Chinese AI companies to conclude: "To achieve high-performance AI, we need chips tailored to our models." As Reuters noted, "In-house chips would enhance control over hardware supporting their models and reduce reliance on NVIDIA—a direction aligning with global AI developers."
**Competition Intensifies Across China's AI Ecosystem**
Competition in China's AI chip market is intensifying. According to IDC, NVIDIA's market share in China dropped to 55 percent last year from over 90 percent before 2022, when U.S. export controls began. Huawei partially filled the gap with Ascend chips, while Alibaba and Baidu entered the race through semiconductor subsidiaries. Cambricon and Moore Threads, dubbed "China's NVIDIA," are accelerating development after securing funding through IPOs.
With AI model firms like DeepSeek and Zhipu joining the competition, analysts characterize the landscape as entering a "Warring States era" in China's AI semiconductor market. A semiconductor industry insider remarked: "China is transitioning to a 'self-reliant AI ecosystem' where chip design, production, and software are optimized together."
Momentum is building. Bernstein forecasts that the Chinese AI chip market will grow from $24 billion by year-end to $88 billion by 2028, with the self-reliance rate rising from 79 percent to 93 percent. In memory semiconductors—critical for AI chip performance—Chinese firms like CXMT and YMTC have built self-sufficiency in DRAM and NAND flash. With AI chip companies including Huawei, Cambricon, and Moore Threads expanding, and AI model firms pursuing vertical integration, self-reliance is spreading across the entire AI ecosystem.
That said, many analysts assess that Chinese AI chips will struggle to match NVIDIA's latest products in the near term. A semiconductor industry insider noted: "Chinese AI chips still lag behind NVIDIA in performance, power efficiency, and software ecosystems," adding that "even if performance is slightly lower, building an independent AI infrastructure centered on domestic chips is China's direction."