Chinese Chipmakers Seize 48% of Domestic Market as Nvidia Share Collapses
US export controls accelerate $150 billion subsidy blitz that turned domestic alternatives from negligible to dominant in 18 months.
Chinese domestic chipmakers captured 41-48% of the local AI accelerator market in 2025, up from under 30% the prior year, as Nvidia’s share fell from over 60% to 55%, according to IDC data released today. The shift marks the most dramatic redistribution of semiconductor market power in a major economy since Japan’s DRAM ascent in the 1980s, driven by $150 billion in cumulative state subsidies, three years of US export sanctions, and competitive parity in inference workloads that make domestic chips “good enough” for 60-70% of Chinese AI infrastructure demand.
The numbers tell the story. Chinese vendors collectively shipped 1.65 million AI accelerator cards in 2025 versus Nvidia’s 2.2 million, per the IDC report. Huawei Technologies alone moved 812,000 units—roughly half of all domestic shipments and triple its 2024 output. Cambricon Technologies, a specialist AI chip designer, plans to ship 500,000 units in 2026 including 300,000 of its Siyuan 590 and 690 models, more than tripling its 142,000-unit 2025 run, Tom’s Hardware reported in December 2025. Huawei targets 600,000 Ascend 910C chips this year, with total Ascend die production reaching 1.6 million across all models, according to Bloomberg.
Bernstein analysts cited in Chinese trade media expect Nvidia’s China share to crater further—from 66% in 2024 to around 8% in 2026—as Huawei, Cambricon, and smaller domestic players approach 80% collectively, per Tom’s Hardware. TrendForce projects domestic share will hit 50% this year, according to CSIS analysis of China’s localization drive.
The Subsidy Machinery
China’s Big Fund Phase III, established in May 2024 with 344 billion yuan ($47.5 billion) in capitalisation, sits atop a decentralised subsidy architecture that now exceeds $150 billion since 2014 when combined with Phase I ($20 billion) and Phase II ($29 billion), according to the Semiconductor Industry Association. Provincial and municipal governments layer additional funding—Shenzhen alone committed $2.3 billion for local chip ventures in 2025. The model creates competitive redundancy: rather than picking winners, Beijing funds parallel paths across Huawei, SMIC, Cambricon, and four emerging GPU startups (Moore Threads, MetaX, Biren, Enflame) that all went public or filed IPO paperwork within a two-month window between January and March 2026, The Meridiem reported.
Cambricon’s revenue surged fourteen-fold in Q3 2025, with the company now capturing over half of domestic cloud provider orders. ByteDance has standardised on Cambricon’s chips for inference workloads, and Alibaba is expected to follow, Tom’s Hardware reports. The company’s Siyuan 590 and 690 accelerators—built on SMIC’s 7nm process—deliver 80-85% of Nvidia H100 performance on inference tasks at half the power draw, industry sources told The Meridiem. “Chinese chips are still behind, but it’s getting better every generation,” said Naveen Rao, CEO of Unconventional AI. “They’re ramping their ability to produce chips and their ability to have each chip be almost similar performance now to Nvidia chips.”
Manufacturing Bottlenecks Persist
Performance gains mask structural weaknesses. Yield rates for Chinese AI Chips remain around 20%, versus TSMC’s 60-80% commercial standard, according to industry estimates cited by TechRadar. SMIC’s total advanced-node capacity (7nm-class) is projected to reach 60,000 wafers per month by end-2026 and 80,000 by 2027, up from 45,000 in late 2025, but equipment import restrictions—ASML’s EUV lithography tools remain blocked—constrain further scaling. High-bandwidth memory (HBM) presents the most acute chokepoint: domestic producer CXMT will output only 2.2 million HBM stacks in 2026, enough for 250,000-400,000 complete Ascend 910C packages, far below Huawei’s production target, according to Tom’s Hardware.
| Metric | Chinese Fabs (SMIC) | Global Leaders (TSMC) |
|---|---|---|
| Process Node | 7nm (DUV) | 3nm (EUV) |
| Yield Rate | 20% | 60-80% |
| Capacity (2026) | 60k wafers/month | 200k+ wafers/month |
| HBM Supply | 2.2M stacks (CXMT) | Unlimited (SK Hynix, Micron) |
The gap means Chinese chipmakers remain 5-10 years behind on leading-edge process technology and 3-4 generations back on energy efficiency. But in a $182.8 billion domestic semiconductor market growing at 6.5-8.9% annually, Statista data shows, “good enough” chips capture revenue that would otherwise flow to Nvidia. Cloud inference, edge AI, and surveillance workloads—where latency and throughput matter more than peak FLOPs—suit current Chinese capabilities.
Export Control Paradox
US export restrictions, intended to constrain Chinese AI development by denying access to cutting-edge compute, instead accelerated the very self-sufficiency they aimed to prevent. “After being tied up by US export controls and sanctions, it has become apparent that escaping this chokehold is more important in the long run than making any temporary advances,” one Chinese industry source told CSIS researchers. Fan Zhiyuan at Sinolink Securities told The Economy that Washington’s policy “is creating new opportunities for China’s semiconductor industry.”
“This is a pivotal moment for the global AI chip market.”
— Stacy Rasgon, Semiconductor Analyst, Bernstein Research
The Commerce Department approved limited Nvidia H200 exports to China in late February 2026 under a new licensing regime, according to Deloitte‘s 2026 semiconductor outlook, but disclosed neither unit volumes nor revenue projections. The move suggests tacit acknowledgment that blanket denial no longer serves US strategic interests when domestic alternatives have achieved parity in most commercial workloads. By the time H200s reach Chinese data centres, Huawei and Cambricon will have locked in multi-year procurement contracts with Alibaba, Tencent, and ByteDance worth an estimated $8-12 billion combined.
Geopolitical Fragmentation
The China market’s bifurcation into a Nvidia-dominated high-end (training, frontier models) and a Chinese-dominated mid-tier (inference, edge) mirrors broader supply chain decoupling. Taiwan remains the critical dependency—TSMC still fabricates Nvidia’s H100s and H200s, and any kinetic conflict over Taiwan would sever both US and Chinese access to leading-edge production. But China’s progress in 7nm and planned 5nm nodes (SMIC targeting 2027-2028 risk production) reduces reliance on external foundries for everything except the bleeding edge.
China’s semiconductor self-sufficiency drive began in earnest after the 2018 ZTE sanctions, which nearly bankrupted the telecom giant overnight. The 2022 Huawei chip ban and subsequent export controls on AI accelerators transformed what was a defensive hedge into an existential priority. Big Fund Phase I (2014) focused on memory and logic; Phase II (2019) targeted equipment and materials; Phase III (2024) concentrates specifically on advanced packaging, EUV alternatives, and AI chip design—the exact chokepoints US controls aim to exploit.
Korea and Japan face impossible arbitrage. SK Hynix and Samsung supply 75% of global HBM—critical for both Nvidia and Chinese aspirations—but US pressure to restrict HBM exports to China conflicts with $40+ billion in annual Korea-China semiconductor trade. ASML’s EUV monopoly gives the Netherlands veto power over Chinese 3nm ambitions, but blocking all DUV sales (needed for 7nm) would sacrifice €8 billion in annual China revenue. These tensions will sharpen as Chinese demand grows and US containment tightens.
What to Watch
SMIC’s 2026 capacity expansion execution will determine whether Huawei and Cambricon can meet production targets or face allocation shortfalls that hand market share back to Nvidia. HBM remains the binding constraint—any breakthrough in CXMT yield rates or a second domestic HBM producer would accelerate Chinese independence; conversely, supply bottlenecks could leave 300,000-400,000 units of AI accelerator demand unfulfilled, creating an opening for authorised Nvidia sales. Watch for bilateral HBM export restrictions from South Korea or Japan—the next logical escalation in US containment strategy.