China’s AI Chip Sector Hits Scale Milestone as Export Curbs Fuel Domestic Demand
China's push for semiconductor self-sufficiency has reached a critical threshold, with at least nine domestic artificial intelligence chipmakers now surpassing 10,000 units in shipments or orders. This surge marks a pivotal transition for the sector, moving from experimental pilot programs to mass-scale validation within the country's data centers as local technology firms accelerate the adoption of homegrown hardware.
The expanding cohort of suppliers ranges from established tech giants like Huawei and Baidu to emerging startups such as Sunrise. This shift is being driven by stringent U.S. export controls, which have compelled Chinese enterprises to seek alternatives to restricted foreign technology. Industry data indicates that while Nvidia Corp. retains a lead, domestic chips captured approximately 35% of China's $16 billion accelerator market in the first half of 2025, with local growth rates now outpacing their American rival.
This volume breakthrough signals that domestic alternatives are achieving necessary market acceptance regarding performance and cost, with unit prices for inference chips ranging from RMB 30,000 yuan (US4,143) to RMB 200,000yuan(US27,624). However, industry insiders caution that reaching the "10,000-card" mark is merely the beginning of a grueling verification phase. The competitive focus is now shifting from raw hardware specifications to the stability of supply chains, software ecosystems, and commercial service capabilities.
Semiconductor industry sources forecast a further explosion in shipments in 2026 as local foundry capacity improves. While some analysts compare the sector's trajectory to the rapid rise of China's solar industry, significant hurdles remain in software compatibility and production bottlenecks, suggesting the market has yet to determine its ultimate winners in a landscape where "trial and error" is evolving into large-scale deployment.
The "10,000-Card Club" Expands
A diverse group of Chinese chipmakers is rapidly scaling operations to meet the void left by restricted foreign imports. Leading the pack are units backed by major technology conglomerates. Huawei’s Ascend series has secured the largest market share among domestic players, widely deployed in telecom and corporate clusters. Baidu’s Kunlun Chip ranks third domestically, having activated a 10,000-card cluster using its P800 chips in February 2025, with plans to expand to a 30,000-card cluster.
Beyond the tech giants, companies preparing for public listings are reporting significant volumes. According to recent prospectus filings, Enflame Technology recorded cumulative sales of approximately 97,200 AI acceleration cards and modules by September 2025. Iluvatar CoreX delivered 52,000 chips to clients in finance and transportation by mid-2025, while MetaX Integrated Circuits (Shanghai) reported cumulative sales exceeding 25,000 units as of August 2025.
Even early-stage startups are crossing the volume threshold. Sunrise, a spinoff from SenseTime Group, disclosed deliveries exceeding 10,000 units in 2025. Similarly, Tsing Micro, a startup backed by state funds, secured orders for over 20,000 cards by January 2026, primarily from municipal intelligent computing centers.
Strategic Pivot to Inference and Cost Efficiency
Facing limitations in advanced manufacturing, Chinese chipmakers are increasingly focusing on the inference market—the process of running live data through trained models—rather than model training. IDC projects that inference will account for 76.8% of China's generative AI infrastructure market by 2029. To compete, local firms are prioritizing "token throughput" efficiency, aiming to lower the cost of generating AI responses.
Some domestic startups are bypassing the restricted 7nm manufacturing nodes and High Bandwidth Memory (HBM) supply chains entirely. Instead, they are utilizing mature 12nm processes and LPDDR memory, which are readily available within the domestic supply chain. This strategy allows companies to offer chips at prices as low as RMB 30,000 yuan (US$4,143), focusing on availability and controllability over peak theoretical performance.
Despite these hardware constraints, performance in specific applications is proving competitive. Industry tests indicate that for certain models like DeepSeek-R1 and Alibaba’s Qwen, chips from Baidu and Alibaba Group have demonstrated token throughput efficiency superior to Nvidia’s H20—a modified, lower-performance chip designed for the Chinese market to comply with U.S. regulations. Sunrise CEO Xu Bing stated that the industry has driven inference costs for one million tokens down to RMB 1 yuan (US$0.14) in 2025, with targets to reduce this further using specialized architectures.
Ecosystem Barriers and Future Outlook
While hardware shipments are rising, the software ecosystem remains a critical bottleneck. Unlike Nvidia’s CUDA platform, which supports millions of models on open-source communities like Hugging Face, domestic chips often support only a few dozen mainstream models. Adapting new models to domestic hardware can take one to two months, delaying the deployment of the latest AI innovations for local enterprises.
The market is currently split on the long-term trajectory of these chipmakers. Optimists view the surge in shipments as parallel to the early days of China’s photovoltaic industry, predicting that intense competition and state support will eventually produce globally competitive champions. Skeptics, however, argue that unlike standardized solar panels, AI chips require complex software stacks and ecological integration.
With the supply chain constrained upstream by foundry capacity and downstream by software fragmentation, the ability to ensure delivery stability and lower migration costs will determine survival. As the industry moves deeper into 2026, the true test for these companies will be securing repeat orders and sustainable revenue beyond their initial bulk deliveries.