Nvidia plans to resume targeted shipments of specialized AI inference chips to China following recent export control adjustments.
Nvidia is preparing to ship small initial batches of a newly adapted artificial intelligence chip tailored for Chinese buyers by year-end. According to The Information, several customers in China have already submitted orders for the hardware, which could help the tech giant re-establish its footprint in the regional market. The product is a specialized Language Processing Unit (LPU) variant designed to co-process AI workloads within U.S. export parameters. The chip requires no reductions in raw processing power to meet export guidelines, though software adjustments were made to ensure compatibility with processors available in China. Nvidia initially designed the LPU to function as part of its Vera Rubin system, paired with Rubin GPUs. Because trade regulations prevent Vera Rubin systems from being sold in China, Nvidia modified the underlying software to enable the LPU to work with other processors available in the country. Nvidia declined to comment.
The LPU is an auxiliary processor developed using technology licensed from Groq. It works alongside primary graphics processing units (GPUs) to accelerate response generation for AI chatbots. The Information reported that the LPU itself naturally complies with U.S. export restrictions without requiring structural reductions in computing power.
This development arrives as Chinese tech firms face an acute shortage of inference processing capacity. Local providers like Moonshot AI recently paused new customer registrations because of capacity bottlenecks, while DeepSeek raised developer pricing amid high demand for its latest models. Surging demand for inference hardware among Chinese AI firms has created market opportunities despite domestic competition. The Information noted that major Chinese companies have locked up domestic chip production capacity through 2027, creating an opening for Nvidia despite local competition. The new LPU variant will directly compete with Huawei Technologies’ Ascend 950DT inference chip. While Nvidia CEO Jensen Huang previously said the company had largely conceded the Chinese market to Huawei, current supply constraints offer a brief window to retain domestic clients before they fully migrate to local suppliers.
Navigating trade friction between Washington and Beijing remains a central hurdle. U.S. policy blocks Nvidia’s highest-end accelerators but permits restricted sales of previous-generation H200 chips. Beijing has historically limited or discouraged imports of certain cleared U.S. processors, though Chinese authorities recently signaled that limited purchases of H200 units may proceed. According to sources cited by The Information, assembly firm Amkor Technology (AMKR) completed packaging for over one million H200 chips in the second quarter, utilizing components manufactured by Taiwan Semiconductor Manufacturing Co (TSM).
Amid these developments, NVDA share price eased 0.5% on Thursday, down for the third straight day. Retail sentiment on Stocktwits was ‘extremely bearish’ with ‘low’ message volumes. Most retail investors were exploring going short on the stock. Despite the recent pullback, NVDA stock has gained 15.2% year-to-date.