The competitive landscape of China's AI compute market is being rewritten by a single dataset. According to IDC, Nvidia's share of China's data-center AI chip market fell to 34% in the first half of 2026, while Huawei climbed to 28% — narrowing the gap to just 6 percentage points, down from 31 a year earlier.

Diverging Shipments

From January to June, Nvidia shipped 678,000 AI accelerators in China, down 44% year over year; Huawei shipped 572,000 units, nearly double the 300,000-plus of a year ago. Nvidia CEO Jensen Huang has repeatedly criticized Washington's export curbs, saying they hand China's roughly $50 billion AI chip market to Huawei — a prediction now materializing.

Key Numbers at a Glance

· Nvidia share: 55% (2025) → 34% (H1 2026); shipments -44% YoY

· Huawei share: 28%; 572K units shipped, nearly doubled

· China took 2.008M AI accelerators in H1: GPUs -26% to 991K; non-GPU accelerators +~80% to 1.016M

· IDC forecasts 2026 full-year China AI accelerator shipments +19%, with GPU shipments -9%

Non-GPU Accelerators Overtake GPUs

The deeper shift lies in product form: non-GPU accelerators — ASICs, FPGAs and custom NPUs — grew nearly 80% to 1.016 million units, overtaking GPUs for the first time. Huawei, T-Head, Kunlunxin and Cambricon together hold 95% of the non-GPU market. China's compute growth engine is shifting from the GPU track dominated by Nvidia to the purpose-built accelerator track where domestic vendors lead.

The competition is moving from single-chip performance to system-level capability — how efficiently processors work together across large clusters.

Ascend's Supply Ceiling

Huawei's 28% was achieved under supply constraints. Reports say Ascend 910C shipped ~812,000 units in 2025 plus 150,000 in Q1 2026, serving 37 leading model companies, with top internet firms exhausting inventory. Bottlenecks sit in advanced-process yields and HBM supply; Bloomberg reports Huawei plans to lift 910C output to ~600,000 units in 2026. Rotating Chairman Xu Zhijun has published a three-year roadmap: Ascend 950PR arrives in 2026 with 112GB in-house HBM at 1.4TB/s, extending supply-chain autonomy from fabrication into memory.

What It Means for Enterprise Procurement

For organizations planning AI infrastructure, two signals stand out: domestic compute has moved from "usable option" to primary supply, making hybrid Nvidia-plus-domestic deployments the norm; and system-level delivery — interconnect, cluster scheduling, software stacks — matters more than peak chip specs. NineZenith applies the same workload-first principle in its full-domain AI platform practice: training, fine-tuning and inference loads demand different silicon, and the energy-efficiency edge of non-GPU accelerators deserves first consideration for inference.

(Synthesized from the IDC report and public coverage by EET-China, 36Kr, MyDrivers, Phoenix News, Cailianshe and others)