China’s AI chip crossroads: GPUs vs ASICs in race to break Nvidia’s dominance

Under sustained US export controls on advanced semiconductors, China’s AI chip industry is accelerating efforts to build a self-reliant silicon ecosystem capable of reducing dependence on Nvidia, according to analysis by South China Morning Post (SCMP).

At the heart of this contest lies a strategic divide: whether China should continue refining general-purpose graphics processing units (GPUs) or pivot decisively toward highly specialised application-specific integrated circuits (ASICs).

SCMP notes that the competition is no longer about replicating Nvidia outright. Instead, it is about constructing a layered domestic ecosystem of AI chips capable of supporting increasingly sophisticated models developed by firms such as DeepSeek and Alibaba.

As leading players including Huawei Technologies, Cambricon Technologies and Moore Threads expand their ambitions, the GPU-versus-ASIC debate has become central to China’s semiconductor roadmap.

GPUs: flexibility in Nvidia mould

Originally designed for video game graphics, GPUs evolved into the backbone of modern AI computing after Nvidia demonstrated their effectiveness in parallel processing for neural networks. Their strength lies in versatility: GPUs can be repeatedly reprogrammed to accommodate rapidly changing AI model architectures, making them particularly suitable for research and general-purpose workloads.

In China, companies such as Moore Threads, Biren Technology, Enflame and Iluvatar CoreX are attempting to replicate this flexible computing model.

Moore Threads, founded by a former Nvidia China executive, has positioned itself as a domestic analogue to Nvidia, focusing on general-purpose GPU architectures such as its MTT S5000 series.

However, despite progress, Chinese GPU developers still face significant performance and software ecosystem gaps compared with Nvidia’s CUDA-driven dominance.

ASICs: efficiency over versatility

By contrast, ASICs are purpose-built chips designed for specific computational tasks. While less flexible than GPUs, they are significantly more efficient for targeted workloads, particularly in AI inference and deployment. This efficiency advantage has triggered a wave of investment in specialised architectures.

Within this category, neural processing units (NPUs), tensor processing units (TPUs) and proprietary parallel processing units (PPUs) are emerging as the main design pathways. Huawei is leading with its Ascend NPU series, including the 910C and upcoming 950 models. Cambricon continues to advance its Siyuan ASIC line, while Alibaba’s semiconductor arm, T-Head, has introduced the Zhenwu M890 PPU, which it claims delivers substantial performance gains over its predecessors in complex “agentic AI” workloads.

This shift mirrors a broader global trend in which major tech companies increasingly design custom silicon to optimise cost and performance for AI inference at scale.

Diverging strategies

Industry analysts cited by SCMP argue that the choice between GPUs and ASICs is ultimately determined by workload requirements and engineering maturity. ASICs offer superior cost efficiency for stable, high-volume AI deployment, while GPUs remain essential for flexibility and model development.

Market expectations suggest ASICs may gain near-term dominance. A Morgan Stanley projection estimates Huawei could command more than 60% of China’s AI accelerator market by 2026, with Cambricon also expanding its share.

Performance comparisons further underscore the shift. Some domestic ASICs are reported to outperform Nvidia’s China-allowed H20 chips by significant margins in inference speed, measured in tokens per second.

Software battleground

Yet SCMP emphasises that China’s semiconductor race extends beyond hardware. The ability to challenge Nvidia’s ecosystem depends on developing competitive software stacks to rival CUDA. Huawei’s CANN platform and Moore Threads’ MUSA architecture are central to this effort, as domestic firms attempt to reduce dependence on foreign developer ecosystems.

By Sabina Mammadli