Nvidia China AI Chip Share Falls to 34% as Huawei Surges to 28%
The Nvidia China AI chip share numbers, published in an IDC report seen by Reuters, describe a market being pulled in two directions at once. Nvidia shipped 678,000 data-centre AI accelerator units into China in the first half of 2026 — still the single largest total of any supplier — but that figure was down 44% from a year earlier, and it translated into a 34% share of a market that totalled 2.008 million units in the half. Huawei, shipping from Shenzhen, delivered 572,000 units, up from just over 300,000 a year earlier, for a 28% share. A gap that once looked unbridgeable is now 106,000 units wide.
Behind the top two, a domestic bench is forming. Alibaba's chip arm T-Head took 7% of the market, Cambricon 5%, Baidu's Kunlunxin unit 4% and Hygon 3%. Nine of the eleven Chinese makers tracked by IDC grew their shipments in the period. The overall market tells the same story from another angle: total accelerator shipments reached just over two million units, yet shipments of GPUs — the category Nvidia dominates worldwide — fell 26% to 991,000 units. China's AI build-out is not shrinking. It is changing suppliers.
How the squeeze works: controls from both capitals
Nvidia's decline in China was not primarily a competitive defeat in the ordinary sense; it was engineered, from both sides of the Pacific. Washington's export controls bar Nvidia from selling its most powerful current accelerators, the B200 and B300, into China at all. The compliant, reduced-performance chips Nvidia is permitted to offer sit in a narrower and politically exposed band — and Beijing has added its own layer, requiring approval before Chinese buyers import even those compliant Nvidia products, while state guidance and procurement preferences steer data-centre builders toward domestic alternatives.
That double squeeze explains the shape of the IDC table. Nvidia did not lose 44% of its unit volume because Chinese engineers suddenly preferred rival silicon on merit; it lost volume because whole tiers of its catalogue became unsellable, and because the permitted tier became harder to buy. Huawei's near-doubling, from roughly 300,000 units to 572,000, happened inside a protected demand pool: every large Chinese cloud builder and state-linked AI project that could not, or would not, navigate approvals for American chips needed a domestic supplier able to deliver at scale. Huawei was the only one already shipping in the hundreds of thousands.
Why this matters
The strategic question behind the IDC table is whether export controls are achieving their stated aim. The controls were sold in Washington as a way to slow the build-out of frontier Chinese AI by denying it the world's best training hardware. On a narrow reading, they work: Chinese labs cannot legally buy B200 or B300 systems, and Nvidia's own shipments have fallen steeply. But the H1 2026 data shows the policy's second-order effect arriving at scale — a domestic substitution industry growing at exactly the pace the controls created demand for. Total Chinese accelerator shipments of 2.008 million units in six months is not a starved market. It is a large market learning to feed itself.
Jensen Huang has made the counter-argument in public for some time: that restricting Nvidia does not stop Chinese AI, it hands China's roughly $50 billion AI chip market to Huawei and removes American technology from the stack Chinese developers build on. The IDC figures are the strongest evidence yet for his framing — and also, read carefully, for its limits. Huawei at 28% is a triumph of industrial policy, but Nvidia at 34% is still the market leader despite being barred from selling its best products, which suggests Chinese buyers still prefer Nvidia silicon whenever they can get it approved. Both things are true at once: the controls are costing Nvidia billions, and they have not yet made Chinese accelerators the first choice of China's own builders — only, increasingly, the available one.
The numbers in context
Put the half-year figures against the market's recent history and the reversal sharpens. A few years ago Nvidia's position in Chinese data-centre AI was close to a monopoly of the high end; a 34% share with shipments down 44% year on year would have been unthinkable before the control regime tightened. The 26% fall in GPU shipments to 991,000 units matters for the same reason: much of the growth in the 2.008 million total is coming from non-GPU accelerators — the ASIC-style designs Chinese firms specialise in — which means the architecture of China's AI fleet is diverging from the CUDA-centred fleets being built in the United States and Europe. Divergence compounds. Software written for one stack does not move cheaply to another, so every half-year of substitution locks in a little more of a separate Chinese AI ecosystem, whether or not controls are later loosened.
For investors, the comparison that counts is with Nvidia's global story, which remains extraordinarily strong — see our coverage of Nvidia's record run in the $6 trillion race. China was once a quarter-scale business for Nvidia's data-centre segment; the IDC data implies it is now a shrinking slice of a growing pie, and that the lost volume is not deferred demand but demand permanently re-homed. The same substitution pressure is showing up in the custom-silicon strategies of American cloud firms, covered in our report on Microsoft's Maia 300 AI chip — Nvidia is being challenged at home by customers and in China by policy.
Who benefits, who loses — and what critics on both sides say
The clearest beneficiaries are China's domestic designers. Huawei gains scale, and scale in semiconductors is self-reinforcing: volume funds the next chip generation and deepens the software ecosystem around its parts. T-Head, Kunlunxin, Cambricon and Hygon gain something almost as valuable — anchor customers with a policy reason to buy Chinese. Nine of eleven domestic makers growing in the same half suggests a broad-based lift, not a single champion effect.
The losers are more complicated. Nvidia loses volume and, arguably more important, loses its position as the default platform inside the world's second-largest AI economy. Chinese cloud builders lose access to the best-performing hardware and the mature CUDA software stack, which raises their training costs and complicates their engineering — a real competitive drag, even if domestic chips keep improving. And American policymakers lose the clean narrative: supporters of the controls argue, with some justification, that denying frontier hardware still slows China's most advanced military-adjacent and frontier-model work, whatever happens in the commercial accelerator market. Critics reply that a policy which produces a self-sufficient rival industry, while the target's total shipments still top two million units a half, may be buying a few years of advantage at the price of permanent market loss. The IDC report does not settle that argument, but it has made it impossible to have the argument without numbers.
Nvidia itself declined to comment to Reuters on the IDC figures — a notable silence from a company whose chief executive has been anything but silent about the policy. One reading is that Nvidia sees no advantage in litigating half-year share data while approval decisions for its compliant chips sit with Beijing. Another is that the numbers speak plainly enough on their own.
What happens next
Three things will decide whether 34% is a floor or a waypoint. First, approvals: if Beijing streamlines import approvals for compliant Nvidia chips, some demand snaps back quickly, because the preference for Nvidia hardware among Chinese builders appears intact. If approvals stay slow and discretionary, Huawei's share converges on Nvidia's within a few halves on current trajectories — Huawei needs to add roughly another hundred thousand units a half at Nvidia's current volume to draw level. Second, performance: domestic substitution holds only if Chinese accelerators keep closing the capability gap, particularly for large-model training; a visible stall would push even state-favoured builders to lobby for imported supply. Third, Washington: any further tightening — or, in the opposite direction, a negotiated loosening tied to a broader trade settlement — reprices every line in the IDC table overnight.
The deeper story, though, is already written into the H1 data. The world's AI chip market is splitting into two ecosystems with different suppliers, different architectures and different software stacks — a pattern our coverage of the chip selloff that hit Asian markets this week shows investors are only beginning to price. Whether that split was the controls' purpose or their accident will be debated for years. That it is happening is no longer debatable at all.
Sources
- Reuters — reporting on the IDC China AI accelerator report, Nvidia and Huawei H1 2026 shipment and share data, October 2026: reuters.com/technology
- IDC — China data-centre AI accelerator shipment data, first half 2026: idc.com
- Signal Post News — related coverage linked in the text above.