Why Huawei’s Ascend AI Chips Command 79% of China’s Compute Market – An In‑Depth Industry Analysis

Market Share and Deployment

Huawei's Ascend AI chips have been deployed in 28 cities across China, accounting for roughly 79% of the domestic smart compute center market in 2022, according to financial news reports.

Huawei AI Chip Portfolio

Huawei's flagship AI chip products are the Ascend 310 and 910B. The 310 focuses on inference, while the 910B targets both training and inference with FP32 and FP16 precision, matching the performance of Nvidia A800/A100 on a single card and server level.

Full‑Stack Compute Infrastructure

The Ascend ecosystem is built on a full‑stack architecture that includes:

Hardware: Da Vinci‑core based AI chips, embedded modules, board cards, small stations, servers, and clusters.

Software: The heterogeneous CANN (Compute Architecture for Neural Networks) stack, debugging and optimization tools, the MindStudio development suite, and management utilities.

AI Frameworks: Open‑source MindSpore and other mainstream frameworks, positioning MindSpore in the first tier of AI frameworks.

Application Enablement: MindX, ModelArts, and HiAI services that support a wide range of industry scenarios such as recommendation, natural language processing, speech recognition, and robotics.

Huawei Cloud Pangu 3.0 Model

Huawei Cloud's Pangu 3.0 model, built on Kunpeng CPUs and Ascend AI chips, offers a series of large‑scale models ranging from 100 billion to 1 trillion parameters, supported by the CANN architecture and the MindSpore framework.

HaiGuang DCU

HaiGuang's DCU (Data Compute Unit) is a GPGPU‑based accelerator compatible with CUDA, delivering 100% performance improvement over its predecessor (Deep Compute II). It uses a 7 nm FinFET process and matches the performance of Nvidia A100 and AMD MI100 in typical AI workloads, while maintaining high compatibility with ROCm and CUDA ecosystems.

Cambricon Overview

Founded in 2016, Cambricon focuses on AI processors and offers cloud, edge, and IP licensing product lines. Its MLU series of cloud‑accelerated cards fully support Baichuan2 models (53B, 13B, 7B) and achieve performance comparable to international mainstream products.

JingJiaWei High‑Performance AI Modules

JingJiaWei introduced the "JingHong" series in March 2024, targeting AI training, inference, and scientific computing. The modules support mixed‑precision (INT8, FP16, FP32, FP64) and multi‑card interconnects, compatible with major CPUs, operating systems, and server vendors, accelerating adoption of mainstream deep‑learning frameworks.

Key Takeaways

The Chinese AI compute market is dominated by domestic vendors, with Huawei leading in market share.

Hardware and software stacks are increasingly integrated, offering end‑to‑end solutions for AI model development and deployment.

Competitive products from HaiGuang, Cambricon, and JingJiaWei provide performance on par with international GPUs while offering better price‑performance ratios in the domestic market.

Industry‑specific large models (e.g., finance, government, manufacturing) are being built on these platforms, indicating a maturing AI ecosystem in China.