The tightening of US chip sanctions has sent shockwaves through the high-end AI GPU export market, potentially resulting in short-term supply shortages and cost increases for China's AI and Large Language Model (LLM) industries. However, the long-term implications might favor the development of China's semiconductor industry, prompting Chinese AI chip manufacturers such as Huawei to strive for import substitution through domestic technological innovation.
Nvidia, a dominant force in the GPU market, currently commands a staggering 80% share of the global discrete graphic card market. This stronghold extends into the high-end AI GPU sector, where it exerts considerable influence over AI algorithm training using models like the H100, A100, and V100.
The latest data from IDC China projects the Chinese accelerated computing market, currently valued at US$3.1 billion, is on track to reach US$16.4 billion by 2027. In addition, IDC reports in the first half of 2023, the 50,000 domestically developed AI accelerator cards used in Chinese AI servers represented approximately 10% of the overall server market in China.
When considering full-year data from 2022, IDC's estimations reveal that approximately 1.09 million units of AI accelerator cards were shipped in China. Of this volume, Nvidia maintained a significant 85% share, with Huawei holding 10%. In contrast, Baidu, Cambricon, and Shanghai Enflame Technology accounted for roughly 2%, 1%, and 1%, respectively.
However, despite Nvidia's near 90% market share in China, the recent US sanctions have raised concerns regarding the company's potential withdrawal from the market, prompting a critical question: who is best positioned to fill this substantial void?
While Chinese AI customers have accumulated a considerable number of high-end AI GPUs over the last 2-3 years, the situation might not be as straightforward as it seems. Jensen Huang, the CEO of Nvidia, emphasized in an interview with Financial Times the necessity for the US to proceed cautiously in its actions. He highlighted that if China's access to US procurement is impeded, it will be compelled to pursue the development of its own technology.
Hardware companies represented by Huawei are actively constructing the foundational ecosystem for AI in China. Among these, the Ascend series is a core part of Huawei's AI capabilities. Huawei aims to be an "alternative option" outside of the Nvidia ecosystem.
In recent years, the Ascend series has expanded its scope and shifted its focus towards collaboration with open-source communities, deviating from the Nvidia approach. Huawei seeks to address hardware supply chain concerns and actively challenges Nvidia's ecosystem dominance.
As per reports from 21st Century Business Herald and Beijing Business Today, the PyTorch Foundation recently announced that Huawei has joined as a premier member. Huawei's membership is not only the first of its kind in China but also the tenth globally. This move offers insights into Huawei's evolving strategies for Ascend development and its ambitions in generative AI.
PyTorch is a well-known AI framework globally and was launched by Meta in 2016. Apart from PyTorch, Google released TensorFlow in 2015, and Huawei introduced MindSpore in 2020. Open-source foundations like these bring together top-tier talent, promoting the construction of open ecosystems.
On the front of computing power, Huawei has developed the Kunpeng server CPU based on Arm architecture, as well as Ascend AI chips. Building around Kunpeng and Ascend, Huawei is constructing a new computing ecosystem. Huawei's target is the AI computing infrastructure, and with the advent of generative AI, the Ascend computing system is experiencing rapid growth.
Insiders from Huawei have revealed to Chinese media that, historically, Ascend followed Nvidia's route, creating an entire ecosystem from scratch. The difficulties of this approach include high costs and the challenge of convincing Nvidia customers to shift to the Ascend ecosystem.
Today, Ascend is emulating the open-source route taken by Kunpeng, fostering closer cooperation with global open-source communities. For instance, many customers who had previously established their business on PyTorch found the transition to Huawei's Ascend ecosystem more feasible and less costly.
Nvidia has built a formidable ecosystem on the CUDA platform. Companies like Intel, AMD, and Huawei are striving to capture a portion of the market in AI computing.
Presently, the soaring demand for generative AI, computing power shortages, US chip export bans, and more present new opportunities for Chinese AI chips like the Ascend series.
To expand massively in the short term, working closely with multiple mature open-source communities seems to be the preferred approach. This strategy could significantly reduce the adoption threshold for customers.
Ascend has always been compatible with various AI frameworks, but among many of its customers more preferred PyTorch to be adapted for their AI solutions. In general, Chinese customers tend to engage with multiple hardware providers, but a significant portion has already adopted Ascend.
Chinese media citing industry sources has pointed out it's a common practice for Chinese AI customers to test local AI chips. Although the migration process might span over a duration of 2-3 years or even more, the growing synergy between Ascend and the PyTorch open-source community is anticipated to notably reduce the expenses and intricacies linked to AI customer development and migration.
