As NVIDIA's China Market Share Plummets to Near Zero, Domestic Chinese AI Chips Rise from 'Inference' to 'Training'

For the past few years, domestically produced AI chips in China have remained on the periphery, primarily used in cost-effective and stable supply 'inference' applications such as government services, finance, and industrial quality inspection.

However, with the escalation of U.S. export restrictions, NVIDIA's (NVDA-US) flagship products—A100, H100, and even the China-specific H20—have been successively cut off from supply. As a result, NVIDIA's market share in China has plummeted from 95% to nearly zero. This sealed-off supply gate has forced China's AI chip industry to undergo a transformation in 2026, officially initiating a systematic leap from 'inference' to the peak of 'training'.

Domestic Substitution Rate in China Exceeds 40%

According to data, of the 4 million GPUs delivered in China's AI server market in 2025, domestically produced chips accounted for 41%. On the demand side, the daily token processing volume of large AI models in China surpassed 1.4 million trillion in March 2026, with demand exploding exponentially.

Under this backdrop, Huawei's Ascend has become the foundational pillar of domestic computing power in China. In 2025, it shipped 812,000 units, ranking first among domestic producers. Its 910 series is currently the only本土solution capable of supporting large-scale training at scale.

Differentiated Strategies by Internet Giants

Facing the computing power gap, major tech companies have adopted varied strategies. ByteDance takes a pragmatic approach, primarily purchasing Huawei's Ascend for training workloads, while deploying Cambricon and TianShu Zhixi for inference tasks. It is also concurrently developing its in-house SeedChip inference processor.

Baidu has pursued a full-stack self-research path. Its subsidiary Kunlun芯 has been independently spun off, with a valuation reaching $50 billion, and over half of its revenue now coming from external clients.

Meanwhile, Alibaba has built an ecosystem through Pingtouge Semiconductor, integrating 'chips + cloud + open-source models'. Its Zhenwu PPU has cumulatively shipped over 600,000 units, ranking second in the market. Tencent has adopted an investment-binding model, deeply supporting Enflame Technology and widely deploying its accelerator cards across hundreds of internal applications, including voice-to-text in WeChat.

Technical Breakthroughs and Key Challenges

Although Chinese domestic chips have achieved breakthroughs in single-card performance and several products have obtained government security and reliability certifications, the transition still faces 'three major pitfalls':

Software Ecosystem Barriers: NVIDIA's CUDA ecosystem remains a deep moat. Domestic firms like Moore Threads have launched MUSA to enable seamless migration, but the completeness of compilers and toolchains remains a weakness.

Supply Chain Risks: Advanced manufacturing processes and packaging technologies (such as CoWoS) remain highly dependent on external suppliers. Under sanctions, the process technology gap cannot be eliminated in the short term.

Lack of a Viable Business Model: Currently, most chipmakers rely on capital expenditures (CapEx) from internet giants to survive, but when AI applications will achieve scalable profitability remains uncertain.