Nvidia's Direct AI Chip Sales to China Near Zero as Domestic Substitution Accelerates

On April 30, 2025, Jensen Huang, CEO of NVIDIA, stated that the company’s direct sales share in China’s AI accelerator market has dropped to nearly zero. This shift stems from the continued escalation of U.S. export controls on semiconductor chips to China. In April 2025, new U.S. regulations effectively excluded China-specific chips such as NVIDIA’s H20 and AMD’s MI308. Although NVIDIA received approval in July 2025 to resume H20 sales to China, it was required to remit 15% of its China-generated revenue to the U.S. government. By December, exports of the more advanced H200 chip were permitted, but with an increased remittance rate of 25%. This “transactional containment” model has rendered compliant exports highly uncertain. Although NVIDIA chips can still enter the Chinese market through gray channels or via overseas deployment and re-importation, legal direct sales channels have effectively been severed. Meanwhile, domestic AI chips are rapidly filling the gap. It is projected that by 2026—excluding chips self-developed by major internet companies—domestically produced high-performance AI inference chips will reach a supply volume of approximately 3 million units. Of these, Huawei Ascend, Cambricon, and Hygon together will contribute 1.6–1.7 million units, while second-tier vendors such as MetaX and Biren will account for nearly 1 million units. The automotive sector has also been significantly impacted. Li Bin, chairman of NIO, revealed that each of its vehicles uses over 4,000 chips across more than 1,000 part numbers, with chip costs exceeding half of the total vehicle cost and described as being “out of control.” Over the past year, NIO spent $300 million purchasing NVIDIA chips. Faced with unstable supply and soaring costs, multiple automakers are turning to in-house chip development: NIO has cumulatively mass-produced over 550,000 self-developed chips and aims for a 35%–40% domestic semiconductor localization rate in its vehicles by 2027; XPeng’s Turing chips, in mass production since Q3 2025, have shipped over 200,000 units, with a full-year 2026 target approaching 1 million; Li Auto’s Mach 100 chip is scheduled for vehicle integration starting in Q2 2026; and FAW Group has co-developed the automotive-grade, multi-domain integrated chip “Hongqi No.1.” Capital markets have responded positively. In early May 2026, Hygon Information’s market capitalization surpassed RMB 820 billion, marking a nearly 60% gain year-to-date. Cambricon’s stock price became the highest on China’s A-share market, surging over 62% within the month. Cambricon reported Q1 2026 revenue of RMB 2.885 billion, up 159.56% year-over-year, and net profit of RMB 1.013 billion, a 185.04% increase, achieving consecutive profitability for the first time. Industry analysts note that while domestic chips have narrowed the hardware performance gap to just one or two generations behind global leaders, significant disparities remain in software ecosystems—particularly CUDA compatibility—and effective alternatives for large-model training scenarios have yet to emerge. Currently, domestically developed chips are primarily deployed in inference and edge computing, complementing the in-house development strategies pursued by automakers. Overall, China’s AI chip market is transitioning from reliance on a single imported source toward a diversified supply ecosystem jointly formed by Huawei, Cambricon, Hygon, internet giants, and automaker-led in-house efforts. Although gaps persist in ecosystem maturity and training capabilities, the trend toward structural substitution has become irreversible.