Nvidia Loses Half China AI Market as Huawei Doubles

Nvidia shipped 678,000 AI accelerators to China in the first half of 2026, a year-over-year decline of 44%, according to new IDC data cited by Reuters on October 13. Over the same period, Huawei nearly doubled its own shipments to 572,000 units. Nvidia's share of China's data-center AI chip market fell from 55% across all of 2025 to 34% in H1 2026. Huawei now holds 28%. The gap between them is six percentage points, down from more than 40 percentage points eighteen months ago. This is not a trade war data point. It is documented proof that the world's second-largest AI infrastructure market is being rewired on a hardware stack the United States does not control.

What Actually Happened

IDC's H1 2026 China AI chip market data was cited in a Reuters report published October 13, 2026, and confirmed across multiple outlets. According to Reuters via Yahoo Finance, Nvidia shipped 678,000 AI accelerators into mainland China in the first half of 2026, down from approximately 1.2 million units a year earlier. That 44% decline is driven almost entirely by US export controls: the H100 and H200 are barred from export to China, and the specifically designed H20, which was approved for Chinese sales as a performance-limited alternative, faces constrained supply and growing substitution by domestic products. IDC projects that Nvidia's GPU shipments to China will fall 9% for the full calendar year 2026, even as total AI accelerator investment in China grows 19%.

Huawei's Ascend-series AI accelerators accounted for 572,000 of the units shipped in H1 2026, up from roughly 300,000 units a year earlier, according to TechBooky citing IDC. IDC's data shows Huawei's share rising from approximately 15% of the China AI chip market in H1 2025 to 28% in H1 2026. Nvidia's market share fell from roughly 55% across full-year 2025 to 34% in H1 2026. The composition of the broader market is shifting in structural ways beyond just Nvidia's decline: GPU shipments overall fell 26% to 991,000 units, while non-GPU AI accelerators, a category dominated by Huawei's Ascend chips and Chinese alternatives from companies such as Cambricon and Biren, rose nearly 80% to 1.016 million units. Non-GPU accelerators now lead GPUs by volume in China's AI infrastructure market, a milestone that has not occurred in any previous reporting period.

Nvidia's stock fell approximately 1% on the news, a relatively muted reaction that reflects the market's existing awareness of export control risk as an ongoing headwind rather than a new development. According to Business Recorder citing Reuters, IDC did not specify which Nvidia products comprise the 34% share figure, and it is unclear whether any B200 or B300 volume is included since those chips are fully barred from Chinese export. IDC also did not disclose what fraction of Nvidia's remaining China sales are attributable to the H20 versus older-generation products. IDC projects total AI accelerator shipments in China will grow 19% for full-year 2026, even as GPU-specific shipments fall 9%, pointing to an AI infrastructure investment trajectory that is decoupling from GPU availability and redirecting toward non-GPU alternatives at scale.

Why This Matters More Than People Think

The 44% decline in Nvidia's China shipments is structurally different from any previous period of chip market disruption. In past episodes of demand softness or supply constraint, market share shifts were temporary: demand recovered, supply normalized, and the incumbent supplier restored its position. The current shift is not temporary because it is not driven by demand softness or supply disruption. It is driven by policy-enforced permanent exclusion of Nvidia's most capable products from a major market, combined with the successful scaling of a domestic alternative that is improving rapidly. The normal mechanisms for market share recovery, price competition, product design adjustments, and distribution investment, are not available to Nvidia in China because the binding constraint is law, not economics.

The structural consequence that most coverage is underweighting is the divergence of AI software ecosystems. Every Ascend chip Huawei ships into a Chinese hyperscaler trains its AI software on Huawei's CANN compute architecture rather than CUDA. CUDA's global dominance in AI is partly technical and partly a function of massive network effects: AI frameworks including PyTorch, JAX, and TensorFlow are optimized for CUDA, AI libraries assume CUDA compatibility, and virtually every AI engineer in the world knows CUDA. China is now large enough, and its domestic Ascend supply reliable enough, that those network effects are being replicated on a competing architecture at national scale. The Chinese companies building AI infrastructure are simultaneously solving the CANN software porting problem that was previously considered a near-insurmountable barrier to Huawei adoption. They are solving it because they have no choice and because the state is providing resources to solve it.

The non-GPU accelerator volume surpassing GPU volume in China deserves more attention than it has received in coverage of the IDC data. For years, the China AI chip narrative centered on GPU smuggling, gray-market H100 imports, and Nvidia's attempts to design export-compliant alternatives. The IDC data reveals that China's AI infrastructure investment has already moved past that framing. The market is no longer primarily constrained by GPU availability. It is actively buying non-GPU accelerators at scale, and the total volume of non-GPU AI accelerators in China now exceeds GPU volume. This is a structural transition, not a temporary substitution. It suggests that Chinese AI labs and hyperscalers have moved from the phase of working around GPU constraints to the phase of building infrastructure that does not require GPU-class hardware for the workloads they are running at scale.

The Competitive Landscape

Nvidia's ability to respond to its declining China position is structurally limited by US export control law. The company cannot sell its most capable products in China, and any new China-specific design requires Commerce Department review and approval. The H20 was Nvidia's attempt at a compliant product with enough performance to retain some China market position. Its declining share within an already-restricted market indicates that the H20's performance ceiling is no longer competitive against Huawei's improving Ascend alternatives for the workloads Chinese hyperscalers prioritize. Nvidia CEO Jensen Huang has acknowledged China as a material revenue risk publicly, and the company has redirected its commercial investment toward the Middle East, Japan, and Southeast Asian markets where export controls do not apply.

Huawei's competitive trajectory is improving on dimensions beyond unit volume. The Ascend 910C, Huawei's current most capable chip, is increasingly benchmarked against the H800 rather than exclusively against other domestic Chinese alternatives. Skeptics point out that Huawei still trails Nvidia on raw peak compute per chip and that manufacturing process limitations at SMIC, Huawei's primary chip fabricator, constrain how rapidly performance can scale. A chip manufactured at an effective 5nm-equivalent process cannot match the performance density of Nvidia's chips built at TSMC's 3nm. The bear case for Huawei's trajectory is that SMIC's process technology is approaching a ceiling that TSMC will continue to extend beyond, and that Huawei's per-chip performance gap with Nvidia will persist or widen even as its unit volume grows. In that scenario, China builds more AI infrastructure but falls behind on capability per unit of infrastructure, which is ultimately what matters for the competitive AI development race.

The historical analogy that fits most cleanly is the China mobile chip market trajectory from 2012 to 2018. HiSilicon and MediaTek began as suppliers of lower-performing, lower-cost chips that competed primarily on price for market segments US suppliers did not prioritize. Over five years, their process access improved, their software ecosystems matured, and performance closed to within competitive range of Qualcomm's Snapdragon on consumer workloads. By 2018, Kirin chips were powering Huawei's flagship smartphones with performance comparable to Qualcomm's best. Whether Huawei can replicate that trajectory with AI accelerators, given harder manufacturing constraints from the SMIC process ceiling and a much higher absolute performance bar, is the central question the IDC data raises but cannot answer on its own.

Hidden Insight: Export Controls May Be Funding the Competitor They Were Meant to Prevent

The US government's stated objective for AI chip export controls is to prevent China from acquiring the compute infrastructure needed to develop frontier AI capabilities. The actual effect documented by IDC is more complicated and potentially counterproductive at the system level. Nvidia's market share is declining in China, which is the direct outcome the controls were designed to produce. But China's total AI accelerator investment is growing 19% in 2026, which is not the intended outcome. The controls created a commercial opportunity for Huawei that Huawei is competently filling, and filling it with a product that is improving across every reporting period. The question policymakers need to ask is not whether export controls are hurting Nvidia's China revenue, they clearly are, but whether they are slowing Chinese AI capability development. The IDC evidence is that they are not. They are redirecting Chinese AI infrastructure spending from Nvidia to Huawei, while total investment and total deployment volume continue to grow.

The CUDA ecosystem moat is being tested at the largest possible scale. Every major AI company outside China treats CUDA compatibility as a fixed constraint and builds entirely within that assumption. Model architectures, training recipes, inference frameworks, and deployment toolchains are written with NVIDIA GPU assumptions as the foundation layer. This is Nvidia's deepest competitive advantage: not just the hardware, but the switching cost embedded in the entire software stack built on top of it. Export controls are forcing China's AI ecosystem to absorb precisely that switching cost, paid for by necessity across multiple state-connected hyperscalers simultaneously. The IDC data is not just a hardware volume story. It is evidence that the largest-ever software migration from one AI compute architecture to another is underway in China, funded by the resources of Alibaba, Tencent, Baidu, Huawei, and ByteDance, and validated by Huawei's near-doubling of unit shipments.

The investment implications for Nvidia extend beyond the China revenue line. China accounted for roughly 25% of Nvidia's data-center revenue in 2023. The company has offset the decline with explosive growth in US, European, and Middle Eastern AI infrastructure spending. But that growth is concentrated among a small number of hyperscaler customers, each of whom has both the resources and the strategic incentive to develop alternative chip supply. Amazon's Trainium, Google's TPU series, and Microsoft's Maia custom silicon are all active alternative procurement strategies. If Huawei demonstrates at IDC-measurable scale that it can serve China's entire AI infrastructure requirement without Nvidia, it validates the thesis that CUDA lock-in can be escaped with sufficient investment. That validation strengthens the internal business case at Amazon, Google, and Microsoft for continuing their custom silicon investments as a Nvidia pricing hedge, even in markets where Nvidia is freely available.

The 19% total accelerator growth figure also deserves a second look in context of the broader narrative about export controls constraining Chinese AI. The argument that controls will slow China's AI development rests on an assumed compute scarcity that the IDC data does not support. China's AI compute deployment is growing at 19%, even after a 44% decline in Nvidia's specifically supply-constrained category. If the compute growth rate holds, and if Huawei's shipments continue doubling annually, China's total installed AI compute base will continue expanding rapidly regardless of what happens with Nvidia's export allocations. The compute bottleneck that US export controls were designed to create is not evident in the IDC data. What is evident is a market restructuring that transfers revenue and ecosystem control from Nvidia to Huawei.

What to Watch Next

The H2 2026 IDC data, expected in late January 2027, is the most important near-term marker. If Huawei's shipments continue growing while Nvidia's China volume stays flat or declines further, the market share gap will narrow to single digits within two to three quarters, with Huawei on a trajectory to match or exceed Nvidia's China volume by H1 2027. A key leading indicator is any public announcement about Huawei's Ascend 920 development and performance specifications. If credible benchmark comparisons between the Ascend 920 and Nvidia's H200 emerge in the next 90 days, they will define the performance narrative that drives subsequent procurement decisions at Chinese hyperscalers.

The 90-day US policy response will be visible at Commerce Department level. The IDC data will trigger renewed scrutiny of whether the H20's continued export permission is producing its intended policy outcome. If Commerce concludes that Huawei has already effectively substituted for the H20 in most Chinese hyperscaler deployments, the policy case for maintaining H20 export permission weakens: the revenue is going to Nvidia without the protective effect of slowing Chinese AI capability. If Commerce instead concludes that further restrictions would accelerate Huawei's development by removing Nvidia's remaining China sales from the market, it may maintain or even relax H20 export terms. Either outcome represents a pivotal policy decision with direct financial consequences for Nvidia's China revenue in the second half of 2026.

Nvidia's Q3 and Q4 2026 earnings calls, both occurring within 90 days, will reveal how management is guiding on China data-center revenue in light of the IDC figures. Any downward revision to China guidance, or an outsized increase in the disclosed percentage of China revenue that is attributable to non-data-center segments, would signal that H2 2026 is tracking worse than H1 for Nvidia's China market position. Watch also for Nvidia's commentary on whether the 34% market share figure it holds is stable, improving, or declining as the H20 faces increasing Ascend competition in the segments where it was previously the only export-compliant option.

Nvidia's China chip decline isn't a trade-war footnote; it's documented proof that export controls are building the domestic AI chip industry the United States was trying to prevent.

---