🧔♂️ A friendly human may check it before it goes live. More news here
Global investors eye Chinese AI over Wall Street bubble fears
Global investors are increasing their bets on Chinese AI companies as concerns grow over a potential bubble in the US market.
Interest in Chinese AI is rising due to Beijing’s push for tech independence and recent high-profile listings of chipmakers like Moore Threads and MetaX.
Asset managers such as Ruffer are lowering exposure to US tech giants and looking to firms like Alibaba for more AI-related investments.
ETFs from KraneShares targeting Chinese tech stocks have reported growth in assets under management this year, while Rayliant has launched a new ETF focused on Chinese tech.
The Nasdaq trades at a higher earnings multiple than Hong Kong’s Hang Seng Tech Index, which includes major Chinese AI and chip companies like Alibaba, Baidu, Tencent, and SMIC.
MetaX’s Shanghai debut saw its share price rise 700%, while Moore Threads jumped 400% earlier in the month.
Some fund managers caution that valuations for listed Chinese chipmakers may be driven by hype rather than fundamentals, and advise a balanced approach amid ongoing geopolitical tensions.
🔗 Source: Reuters
🧠 Food for thought
Implications, context, and why it matters.
Listed Chinese chipmakers carry high valuations despite minimal share and ongoing losses
- MetaX trades at about 50x 2024 sales versus Nvidia at 34x, while it holds about 1% of China’s AI chip market; Moore Threads says under 1% 12.
- Across four Chinese GPU makers, combined share stays under 3% as Nvidia has 54% and Huawei’s HiSilicon (Huawei’s in-house chip design unit) has 21.4% 2.
- Over three years, these four logged RMB 15.02 billion in losses, with R&D spend about four times revenue 2.
- From 2022 to 2024, MetaX invested over 2.2 billion yuan in R&D, posted a 1.41 billion yuan net loss in 2024, and expects to break even next year 3.
- Investor interest often hinges on geopolitics and necessity from U.S. export controls, with little current commercial traction 2.
Open-source developers and Machine Learning Operations (MLOps) vendors can build differentiated tooling for Chinese AI accelerators
- Ascend (an AI accelerator chip family) works with PyTorch (an open-source deep learning framework) via torch_npu (a plugin for Neural Processing Units, or NPUs) 45. It also supports torchtune (a library for Large Language Model, or LLM, fine-tuning), yet gaps remain for FlashAttention (an algorithm that speeds up transformer attention) 6.
- FastAttention, a custom attention kernel, hit up to 10.7× speedup on Ascend NPUs by tuning for memory hierarchy differences and instruction sets 6.
- Teams at Cambricon (a Chinese AI chipmaker), Hygon (a Chinese processor vendor), and Ascend updated inference engines to support DeepSeek-V3.2-Exp (a recent large-model variant) that signals demand for framework-level tooling 7.
