Nvidia's China Dominance Under Threat: Beijing Adapts to AI Chip Independence

Nvidia's dominance in China's AI supply chain is being challenged by domestic alternatives. Alibaba is testing domestic accelerators, Cambricon posted blowout earnings, and DeepSeek is training models on Huawei hardware. Beijing has shown little interest in letting Nvidia's H20 chips regain traction. Analysts expect Cambricon to ship 143,000 chips in 2024 and Huawei to produce up to 200,000 advanced units in 2025. Nvidia's long-term hold on the China AI market looks less certain than ever.

Nvidia's long-standing dominance in China's AI supply chain is being challenged by domestic alternatives. Alibaba is testing domestic AI accelerators, Cambricon posted blowout earnings, and DeepSeek is training models on Huawei hardware. Beijing has shown little interest in letting Nvidia's H20 chips regain traction. Analysts expect Cambricon to ship 143,000 chips in 2024 and Huawei to produce up to 200,000 advanced units in 2025, signaling a shift in the market.

Alibaba Group Holding Ltd. is developing a new AI chip to address the gap left by Nvidia in the Chinese market. This new chip, currently in testing, is manufactured by a Chinese company, marking a significant shift from Alibaba's earlier AI processors that were fabricated by Taiwan Semiconductor Manufacturing (TSMC) [1]. The new AI chip is designed to handle a broader range of AI inference tasks compared to Alibaba's previous processors. This development aligns with China's broader push to build a self-sufficient AI supply chain amidst U.S. restrictions on advanced chips [4].

Cambricon Technologies, more commonly known in China by its local name Hanwuji, has seen its stock soar by 130 percent in the past two weeks. The company's AI chip shipment is expected to reach 1.09 million in 2028 and 2.33 million in 2030, according to Goldman Sachs' estimates. Cambricon's recently revamped AI chip, Siyuan 590, offers around 90 percent of the performance that Nvidia's A100 GPU can eke out, with a TPP of 4,493 vs. 4,992 for the A100 [2]. Cambricon's revenue jumped by 4,348 percent to $402 million in the first half of 2025, and its market cap has now eclipsed $90 billion.

DeepSeek, a leading startup, is using chips from Shenzhen-based Huawei Technologies Co. to train some models while still using Nvidia’s for its largest and most powerful ones. Such a paradigm shift has long been Beijing’s priority, even if companies prefer Nvidia’s superior offerings. The more momentum this movement gains, the more the transition accelerates as developers build and deploy AI models on top of domestic hardware.

However, there are risks associated with this shift. Cambricon's top 5 customers made up 94.6% of sales, with one client alone accounting for 79.1%. This customer concentration risk could pose a significant challenge to the company's long-term sustainability. Additionally, the production of advanced AI processors in China is still in its early stages, and building advanced fabs takes time and equipment currently barred from the mainland by US-led export controls.

Despite these challenges, China remains well placed to make further progress in AI. Analysts have estimated that shipments from Cambricon will total some 143,000 units this year, while Huawei is expected to produce up to 200,000 advanced AI chips in 2025. Nvidia's long-term hold on the China AI market looks less certain than ever.

References:
[1] https://www.ainvest.com/news/china-private-sector-advance-military-ai-wsj-2509/
[2] https://wccftech.com/goldman-sachs-forced-to-chase-after-chinas-nvidia-cambricon-as-it-glows-red-hot/
[3] https://www.bloomberg.com/opinion/articles/2025-09-04/over-nvidia-the-ultimate-test-for-china-s-ai-chips