White House’s Squeeze on Nvidia Chip Exports Is Turning Into a Win for Huawei

The lower-end AI chips Nvidia Corp. is exporting to China to circumvent U.S. export controls are perceived as inadequate by local tech firms, prompting them to turn to Huawei Technologies Co. Ltd. for their AI chip needs [para. 1][para. 2]. This shift is driven by Chinese tech firms' desperation to source AI chips necessary for training large language models (LLMs) similar to ChatGPT. However, the migration from Nvidia’s established AI ecosystem to Huawei's is daunting [para. 3].

Nvidia reported a strong financial performance with a notable 262% year-on-year increase in revenue to $26 billion for Q1 ending April 28. In its annual earnings report, Nvidia, which commands about 85% of the global AI chip market, identified Huawei as a formidable competitor in four out of five of its primary business categories [para. 4]. In August 2022, Nvidia stopped selling its high-end GPUs, A100 and H100, to the Chinese mainland and Hong Kong under directives from the U.S. government aiming to prevent the technology from reaching the Chinese military[para. 5].

Following additional U.S. export controls in October 2023, Nvidia introduced lower-end alternatives, L20 and H20 GPUs, to maintain its presence in China. Despite this, the H20 chips available in China are unable to meet the local tech firms' demands, compelling companies like Baidu Inc. and Iflytek Co. Ltd. to consider Huawei’s Ascend AI chips instead [para. 6][para. 7].

According to Counterpoint Research, in 2023, Nvidia held 85% of the AI chip market, followed by Intel Corp at 2% and AMD at 1% [para. 8]. AI chips primarily comprise GPUs, field-programmable gate arrays, and application-specific integrated circuits, while general-purpose chips like CPUs are becoming less useful for advanced AI tasks [para. 9].

Despite Huawei's Ascend 910B chip sometimes outperforming Nvidia's H20 chip in specific tests, Nvidia retains an edge with its NVLink interconnect technology, which can combine multiple chips to enhance computing power [para. 10]. Huawei plans to launch a more powerful Ascend 910C chip in the latter half of the year and expects to ship around 400,000 Ascend chips in 2024, a significant increase from last year [para. 11]. The anticipated Ascend 920 series aims to surpass Nvidia's H20 chip in both computing power and interconnect capabilities [para. 12].

Chinese tech giants like Baidu, Iflytek, and Meituan are already incorporating Huawei chips into their LLM projects. However, Iflytek’s Chairman Liu Qingfeng noted the additional time required to migrate to Huawei's platform and re-verify algorithms [para. 13]. The absence of a comprehensive software development ecosystem for Huawei chips poses another challenge. In contrast, Nvidia's CUDA Toolkit provides significant convenience to developers, enhancing its appeal [para. 14].

An executive from a Chinese internet giant mentioned reluctance among his colleagues to switch away from Nvidia's CUDA platforms due to the high costs and inconvenience of rewriting algorithms, alongside the scarcity of supportive development tools [para. 15].