Three years ago, the idea that Chinese domestic chipmakers could claim more than four in ten AI accelerator cards sold in China was dismissed as fantasy. This week, IDC made it fact. New market data reported by Reuters confirms that Chinese vendors collectively shipped 1.65 million AI accelerator cards in 2025 — capturing 41% of China's total AI hardware market — while Nvidia's once-dominant grip slipped to 55%. The numbers represent a structural shift that no amount of catch-up product design or geopolitical negotiation is likely to reverse quickly.
The Numbers: 4 Million Cards, One Clear Trend
According to IDC data reviewed by Reuters, approximately 4 million AI accelerator cards were shipped in China in 2025 — a figure that itself signals how aggressively the country is building out AI compute capacity. Of those, Nvidia shipped roughly 2.2 million units, holding a 55% share. AMD carved out a modest 4% with approximately 160,000 cards. Chinese domestic vendors accounted for the remaining 41%, delivering 1.65 million cards across a growing field of chipmakers.
The scale of the shift becomes stark when set against recent history. Nvidia once commanded market share approaching 90% in China's AI accelerator segment — a position that reflected its near-monopoly on the CUDA software ecosystem, manufacturing advantages through TSMC, and the absence of any credible domestic alternative. As recently as early 2025, Bernstein analysts estimated Huawei and Nvidia were roughly tied at around 40% each. IDC now confirms domestic vendors have collectively surpassed even that updated estimate, reaching 41% when smaller players are counted alongside Huawei.
For Nvidia, which halted China-bound H200 production in early 2026 and shifted TSMC capacity toward the upcoming Vera Rubin architecture, the trajectory carries consequences that extend well beyond a single market. China once represented roughly one-fifth of Nvidia's data center revenue. A continued erosion of that position — at a time when Vera Rubin supply will be constrained and priced at the high end — is not a gap the company can easily absorb.
Huawei's Machine: Volume, Yield, and Software
Within the domestic field, Huawei stands in a category of its own. The Shenzhen company shipped approximately 812,000 AI chips in 2025 — nearly half of all domestic shipments, and more than three times as many as the second-ranked domestic vendor. Alibaba's chip design unit T-Head came in second with roughly 265,000 cards. Baidu's Kunlunxin and Cambricon each shipped around 116,000 units. GPU startups MetaX, Iluvatar CoreX, and Hygon accounted for smaller but growing slices.
What makes Huawei's position durable isn't just shipment volume — it's the manufacturing progress underlying it. The company has doubled its AI chip yield to nearly 40%, up from approximately 20% a year ago, on older lithography processes at SMIC, China's leading chip foundry. Industry analysts have noted that SMIC appears to be prioritizing Huawei production — likely under direction from Beijing's industrial planners — giving the company a structural manufacturing advantage that smaller domestic rivals like Cambricon and Baidu's Kunlunxin cannot easily replicate without similar policy support.
The Huawei Ascend 910C, the chip powering approximately half of domestic Chinese data centers, delivers an estimated 60% of Nvidia's H100 inference performance. That gap matters significantly for large-scale model training, where raw throughput and precision are everything. For inference — running pre-trained models to generate responses, classify images, or power recommendation systems — the performance shortfall is increasingly acceptable, particularly when the alternative (Nvidia hardware) is either legally unavailable, subject to export license uncertainty, or simply not prioritized by state-owned data center operators under government buy-domestic directives.
Software ecosystem development is following the hardware gains. Chinese AI firm Zhipu trained its GLM-5 model entirely on Huawei Ascend hardware using the MindSpore framework, without any US-manufactured semiconductors in the stack. Shenzhen has deployed a 10,000-card AI cluster built exclusively on Huawei chips, demonstrating that government-backed deployment is moving from pilot programs to production scale. The CANN (Compute Architecture for Neural Networks) software framework is advancing to reduce CUDA migration friction, lowering the barrier for developers accustomed to the Nvidia toolchain.
Export Controls as Market Accelerant
The IDC data does not exist in a policy vacuum. The market shift traces directly to successive waves of US export controls that progressively closed off China's access to Nvidia's most advanced products. Controls initiated under the first Trump administration targeted the H100 and A100. The Biden administration extended restrictions to cover downgraded variants designed specifically to comply with the prior thresholds — including what would have been the H20. The cumulative effect was to strand Chinese data center operators in a buyers' market with only one practical option: go domestic or do without.
Washington is now reportedly considering an even broader mechanism: universal approval requirements for all semiconductor exports, shifting from targeting specific chip performance thresholds to requiring federal license review for every sale. If implemented, this would eliminate the workaround space that export control regimes typically create — the space that has historically produced products like the H20, designed specifically to thread the needle between commercial viability and regulatory compliance. Allied nations' semiconductor equipment export controls have simultaneously closed supply chain workarounds through third countries, narrowing options for Chinese buyers seeking foreign hardware through indirect channels.
Bernstein analysts outlined a three-year scenario in which Huawei overtakes Nvidia entirely in China — a projection that looked ambitious when published but now appears to be tracking ahead of schedule against the IDC full-year data. The same analysts estimate that under sustained export restrictions, Nvidia's China share could fall to as low as 8%. That floor projection assumes continued regulatory escalation and no meaningful political resolution that would restore Nvidia's ability to sell competitive products in the market.
The Concentration Problem Beijing Doesn't Want to Talk About
The domestic chip story is not uniformly positive for China's strategic planners, even as the headline market share numbers improve. Huawei's dominant position — shipping more than three times as many cards as the second-ranked domestic vendor — illustrates a structural concentration risk that Beijing has not resolved. Government policy is creating a national champion at scale faster than it is creating a competitive domestic ecosystem.
A chip supply chain that depends on a single company, however capable, carries different risk characteristics than one with four or five credible manufacturers competing on performance, price, and reliability. The gap between Huawei at 812,000 units and Alibaba's T-Head at 265,000 is not closing quickly. Baidu's Kunlunxin and Cambricon are tied at 116,000 — real volume, but not the kind of scale that breaks Huawei's structural advantage when SMIC production is being explicitly directed toward Huawei output.
Building software ecosystems to compete with CUDA is the other long game. The CANN framework is advancing, and the Zhipu training result demonstrates that major model development on domestic hardware is feasible. But CUDA's moat is not just about APIs — it is about a decade of optimization libraries, toolchain integrations, debugging infrastructure, and developer muscle memory. Closing that gap on a timeline that matters for commercial AI development is a different challenge than shipping more units.
What the 41% Number Actually Means
The 41% figure is significant not primarily as a market share number but as a signal about trajectory and irreversibility. Once a country's AI infrastructure is built on a particular chip ecosystem, the migration costs are high. Data centers running Huawei Ascend clusters, with software stacks and developer tooling oriented toward CANN, will not convert back to CUDA-based systems even if Nvidia's market access were fully restored tomorrow. The installed base is accumulating — and so is the organizational knowledge required to operate and optimize it.
For the global AI hardware market, the IDC data represents the most concrete evidence yet that the semiconductor technology war between the US and China is producing a bifurcated world — not just in terms of which chips are available where, but in terms of which software ecosystems, cloud APIs, and model training paradigms become dominant in each region. The long-term implications for AI model development, safety research coordination, and technology standards are profound, and they are being set not in diplomatic communiqués but in shipment data and supply chain decisions made today.
Whether Nvidia can arrest the decline — through political negotiation for renewed market access, through products designed for the remaining available slice of the Chinese market, or through growth in other regions that offsets China revenue losses — remains the central question for the company's medium-term strategic position. The IDC data answers a different question, and its answer is unambiguous: China's domestic AI chip industry has crossed a threshold that is unlikely to be walked back.
