[News] China AI Model Developers Ramp Up In-house Chips: DeepSeek, Zhipu Explore Custom Silicon

Following the lead of AI giants OpenAI and Anthropic, which are developing custom silicon, Chinese players are making their own chip push. According to Reuters, DeepSeek is developing an in-house inference processor aimed at reducing reliance on both NVIDIA and Huawei hardware. On the other hand, The Information, cited by Investing.com, also notes that Chinese AI startup Zhipu AI is in preliminary talks with domestic chip designers to develop a custom processor for its GLM model family.

DeepSeek Reportedly Expands Chip Design Team

DeepSeek’s chip push may still be in its infancy, but the AI upstart has been laying the groundwork for nearly a year, engaging with chip designers, foundries and memory suppliers on potential partnerships, Reuters reports. Sources also told Reuters that the company has also quietly grown its chip design team in recent months.

DeepSeek has historically relied on both NVIDIA and Huawei hardware, with its R1 reasoning model reportedly trained on NVIDIA’s China-specific H800 before Washington banned the product in late 2023. However, with U.S. export restrictions limiting access to advanced AI chips, the company has since deepened its collaboration with Huawei.

Reuters reports that DeepSeek launched its V4 model optimized for Huawei’s Ascend chips in April, while Huawei said its processors were also used in training V4-Flash. Demand for Huawei’s Ascend 950 chips reportedly surged following the model’s release, Reuters adds.

Zhipu AI Eyes Custom ASICs

Meanwhile, The Information, cited by Investing.com, reports that Chinese AI startup Zhipu AI is exploring a custom processor with domestic chip designers for its GLM model family. The move, as noted by the report, comes as usage of its latest model has surged 27-fold, while U.S. export controls continue to tighten access to advanced AI chips.

According to the reports, Zhipu AI is exploring ASICs—custom chips designed for specific AI workloads—as an alternative to NVIDIA GPUs, and has approached several Chinese chip designers for potential collaboration.

As noted by Investing.com, Chinese chip designers such as Cambricon and Biren have also been developing AI ASIC solutions, although The Information has not identified either company as a potential partner for Zhipu.

Challenges Ahead for China’s AI Chip Push

However, China’s AI model developers’ chip push still face challenges. Reuters, citing analyst, notes that DeepSeek is unlikely to commercialize its chips beyond China without access to cutting-edge manufacturing capabilities.

As noted by Reuters, developing a competitive AI chip requires years of investment, while U.S. export restrictions continue to limit Chinese firms’ access to both leading-edge foundries and high-bandwidth memory—two critical pillars of advanced AI chip production.

For Zhipu AI, the key challenge could be time as well. The Information points out its project could take more than two years to complete, requiring it to build out its semiconductor capabilities, navigate chip design and validation, and adapt its software ecosystem to fully utilize the new hardware.

Read more

• [News] Huawei Brings Forward Ascend 950DT Deployment to August, DeepSeek V4.2 Seen as Potential Early Adopter
• [News] DeepSeek Completes First Funding Round, Raises Over CNY 50 Billion at Valuation Above CNY 330 Billion

(Photo credit: DeepSeek)