NVIDIA’s market share in China falls below 60%; domestic AI chips deliver 1.65 million units annually to capture market share
TechFlow Selected TechFlow Selected
NVIDIA’s market share in China falls below 60%; domestic AI chips deliver 1.65 million units annually to capture market share
Last November, Beijing ordered state-owned data centers to fully adopt domestically produced alternatives, accelerating the reshaping of the market landscape.
Navigating Web3 tides with focused insightsAuthor: TechFlow
TechFlow Intro: According to IDC data, China’s total shipments of AI accelerators in 2025 are expected to reach approximately 4 million units. Domestic vendors collectively delivered 1.65 million units—accounting for 41% of the market—while NVIDIA’s share dropped from roughly 95% before U.S. sanctions to 55%.
Huawei leads the domestic camp with 812,000 chips shipped and recently launched its new Atlas 350 accelerator, claiming inference performance 2.87x that of NVIDIA’s H20.
Last November, Beijing ordered state-owned data centers to fully adopt domestically produced chips—a directive accelerating a fundamental reshaping of the market landscape.
Three years ago, NVIDIA nearly monopolized China’s AI chip market. Today, that landscape is unrecognizable.
According to Reuters’ citation of market research firm IDC, China’s total shipments of AI accelerators—specialized computing chips used in AI servers—are projected at ~4 million units in 2025. NVIDIA remains the largest single supplier, shipping ~2.2 million units and holding a 55% market share. Yet this figure represents a sharp decline of nearly 40 percentage points from its pre-sanction share of ~95%. Meanwhile, Chinese domestic vendors collectively shipped ~1.65 million units, capturing 41% of the market. AMD ranked third with ~160,000 units shipped, representing 4% of the total.
The rise of domestic vendors is both a passive consequence of U.S. export controls and an active outcome of China’s “domestic substitution” policy.
Huawei Leads Domestic Vendors; Atlas 350 Targets NVIDIA H20
Among domestic AI chip vendors, Huawei is the biggest winner.
IDC data shows Huawei shipped ~812,000 AI chips in 2025—representing ~20% of the overall market and nearly half of all domestic vendor shipments. Alibaba’s chip design unit, T-Head, ranked second with ~265,000 units shipped. Baidu’s Kunlunxin and Cambricon each shipped ~116,000 units, tying for third place. Hygon, MetaX, and Iluvatar CoreX accounted for 5%, 4%, and 3% respectively of domestic vendor shipments.
Last month, at its China Partner Conference 2026 in Shenzhen, Huawei unveiled its next-generation AI accelerator, the Atlas 350, powered by its self-developed Ascend 950PR chip. Zhang DIXuan, Head of Huawei’s Ascend Computing Business, stated at the launch event that the Atlas 350 delivers 1.56 PFLOPS (petaflops) of compute power under FP4 low-precision arithmetic—2.87x the performance of NVIDIA’s China-specific H20. The card features 112GB of proprietary high-bandwidth memory (HiBL 1.0), with memory bandwidth of 1.4 TB/s and power consumption of 600W.
However, this performance comparison involves methodological inconsistencies. NVIDIA’s Hopper-architecture GPUs do not natively support FP4 precision, whereas the Atlas 350 is the first domestically developed accelerator optimized specifically for FP4—making direct comparisons at identical precision levels invalid. Huawei’s true competitive edge lies in inference: the Atlas 350 targets inference workloads—the deployment phase of AI models—not large-model training.
Seven Huawei partners have already launched complete server systems based on the Atlas 350. iFLYTEK has also announced that its next-generation Spark large language model will be compatible with the Ascend 910/950 compute platform.
Dual Drivers: Export Controls and Domestic Substitution Mandates
NVIDIA’s collapsing market share in China reflects the combined pressure of escalating U.S. export controls and Beijing’s aggressive domestic substitution policies.
The timeline is roughly as follows: Starting in October 2022, the U.S. imposed restrictions on AI chip exports to China. In response, NVIDIA introduced compliant, downgraded versions such as the H20 and A800/H800. In April 2025, the Trump administration banned all AI GPU exports to China outright; in July, it reinstated export licenses for the H20 and AMD’s MI308; in October, NVIDIA CEO Jensen Huang publicly stated that NVIDIA’s market share in China’s advanced AI accelerator segment had “fallen from 95% to zero”; and in December, the administration permitted NVIDIA to export the H200 to China—but Chinese enterprises were instructed to suspend orders for NVIDIA chips.
The policy push from the other side has been equally forceful. According to a Reuters report from November 2025, Beijing issued guidance directing newly built, state-funded data centers to exclusively adopt domestically produced AI chips. Projects less than 30% complete were required to remove already-installed foreign chips or cancel procurement plans altogether.
Reuters estimates that since 2021, over $100 billion in state funding has flowed into China’s AI data center projects—and most Chinese data centers have received some form of state financial support during construction, meaning this policy applies broadly across the sector.
A large-scale data center being built by China Unicom in Qinghai was cited by Reuters as a flagship case of this strategy: valued at $390 million, the project relies entirely on domestically produced AI chips—including those from T-Head.
Real Technical Gaps Remain—But Inference Performance Has Reached the “Good Enough” Threshold
The rising market share of domestic chips does not imply that technical gaps have disappeared.
Most industry analysts estimate that Chinese domestic AI chips still lag behind NVIDIA by five to ten years in datacenter-scale training performance. NVIDIA’s high-end GPUs remain the preferred choice for training trillion-parameter large language models (LLMs). DeepSeek’s use of a 50,000-GPU cluster based on NVIDIA’s Hopper series to train its R1 model serves as a concrete illustration.
In contrast, the situation differs significantly on the inference side. Industry observers contend that for 90% of commercial applications—including image recognition, chatbots, and autonomous driving—domestic chips have already reached the “good enough” threshold, making migration from NVIDIA to domestic alternatives a viable business decision. Anticipated further tightening of sanctions has only accelerated this shift.
The real bottleneck lies in software ecosystems. NVIDIA’s CUDA platform, refined over more than a decade, has become the de facto standard for AI development. Domestic chipmakers are investing heavily in compatibility: MetaX announced its C500 series will support CUDA compatibility; Huawei fully open-sourced its CANN platform in 2025 to expand its developer ecosystem; and both Cambricon and Moore Threads have developed translation tools bridging CUDA to their proprietary programming languages. The pace of ecosystem catch-up will ultimately determine the ceiling for domestic chips’ market share.
Domestic AI Chip Firms Rush Toward Capital Markets
This market-share shift is being mirrored in capital markets.
Since early 2026, China’s GPU sector has seen a wave of IPOs. Biren Technology and MetaX have listed on the STAR Market; Iluvatar CoreX debuted on the Main Board of the Hong Kong Stock Exchange; and Tianjin-based燧原科技 (Enflame) has had its STAR Market listing application accepted. Baidu announced plans to spin off Kunlunxin as an independent listed entity, and according to informed sources, Alibaba is also considering a similar spin-off for T-Head.
Huawei’s R&D investment in 2025 totaled RMB 192.3 billion—22% of its revenue—with key focus areas including chips, software, and manufacturing tools, all aimed at reducing reliance on U.S. technology. At MWC 2026, Huawei’s Rotating Chairman Xu Zhijun stated Huawei would become “an alternative option ensuring uninterrupted global AI compute supply.” According to Reuters, Huawei’s next-generation Ascend 950PR chip has already attracted purchase interest from tech giants including ByteDance and Alibaba, with a 2026 shipment target of ~750,000 units and mass production scheduled to begin in the second half of the year.
For NVIDIA, even though the H200 has regained export approval, trust has already eroded. Beijing’s “self-reliance and controllability” policy is no longer aspirational—it is now an established reality embodied in every domestically produced chip running inside Chinese data centers. When 2026’s market share figures are released, whether the current 55% will rebound or continue falling will hinge on whether Washington’s export policy shifts again—and how quickly domestic chips close the gap in training performance.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
深潮TechFlow
