huawei's ascend chips challenge nvidia's dominance in ai
Shenzhen, Friday, 20 June 2025.
Huawei’s Ascend AI chips are showing strong performance against Nvidia’s processors. Tests using DeepSeek’s R1 model reveal this. Huawei’s CloudMatrix 384 data center architecture is key. It demonstrates Huawei’s success in overcoming US tech restrictions. This could shake up the AI chip market. There is potential impact on Nvidia’s market position and revenue. Huawei’s Ascend AI chip orders have surged by 300%. The US government is threatening 20 year jail terms for those who violate sanctions on the chips.
Huawei’s ascend ai chips challenge nvidia’s dominance in ai
Huawei’s Ascend AI chips are showing strong performance against Nvidia’s processors. Tests using DeepSeek’s R1 model reveal this. Huawei’s CloudMatrix 384 data center architecture is key. It demonstrates Huawei’s success in overcoming US tech restrictions. This could shake up the AI chip market. There is potential impact on Nvidia’s market position and revenue. Huawei’s Ascend AI chip orders have surged by 300% [4]. The US government is threatening 20 year jail terms for those who violate sanctions on the chips [4].
Cloudmatrix 384 architecture
The CloudMatrix 384, described as a specialized “AI supernode,” is designed for handling extensive AI workloads [1]. Huawei anticipates that CloudMatrix will “reshape the foundation of AI infrastructure” [1]. The system includes 384 Ascend 910C neural processing units and 192 Kunpeng server central processing units [1]. These are interconnected via a unified bus, providing ultra-high bandwidth and low latency [1]. This infrastructure supports an advanced large language model (LLM) serving solution known as CloudMatrix-Infer [1].
Performance and efficiency
Huawei’s Ascend AI compute platform, CloudMatrix 384, achieves 300 PFlops per cluster [2]. It also attains an industry-leading energy efficiency ratio [2]. Huawei’s Ascend platform has made breakthroughs in large model training [2]. For example, it trained the Pangu Ultra MoE model, which has a parameter scale of 718 billion [2]. Its training efficiency increased from the industry average of about 30% to over 41%, and has reached over 45% in the laboratory [2].
Strategic implications for nvidia
The rise of Huawei’s Ascend chips poses a direct challenge to Nvidia’s market dominance [4][7]. Dr. Chen, an AI Chip Analyst, noted that Ascend chips have demonstrated impressive gains in specific workloads [4]. This suggests a strong challenge to Nvidia’s market share [4]. Nvidia’s revenue in China has already experienced a 62% drop [4]. Huawei’s advancements could further impact Nvidia’s revenue potential and competitive advantage [1][4].
Huawei’s ecosystem development
Huawei is actively working to build a robust ecosystem around its Ascend AI chips [2]. This includes providing technology tools to help customers migrate to the Ascend ecosystem within a day [2]. Huawei has over 3.3 million Ascend global developers, with 50,000 core developers [5]. Over 400,000 students in China have been trained on Ascend through university collaborations [5]. 思特奇’s IaaS and PaaS products are adapted to Huawei’s 海思鲲鹏 chips, and its AI 九思 large model systems are adapted to Huawei’s 海思昇腾 chips [3].
Overcoming challenges and us sanctions
Despite facing challenges and US sanctions, Huawei is employing innovative strategies to enhance its computing power [2][7]. Ren Zhengfei stated that Huawei uses mathematics to compensate for physics and cluster computing to compensate for single chips [7]. This approach allows Huawei to achieve comparable results to more advanced chips [7]. The US has imposed strict measures, threatening up to 20 years imprisonment for those who violate rules related to Huawei’s Ascend AI chips [4].