Every major leap in semiconductor history has been accompanied by mergers and acquisitions, and China is no exception. However, in the context of AI, Chinese players now face a new requirement: moving beyond mere domestic substitution to system-level definition. Instead of simply targeting foreign benchmarks, they must adapt to local conditions and define a technology stack tailored to China’s needs. For domestic AI, decoupling computing power, storage capacity, and data transfer capabilities might be the right direction.
When news broke about Hygon’s strategic restructuring and Dawn, the first reaction was how similar it seemed to AMD’s acquisition of ZT System. Back then, many wondered if ZT could continue its NV-related business. Now, similar questions arise: can Dawn still engage in other domestic GPU businesses? We previously mentioned DeepSeek’s new paper, which emphasized that AI, as a new ‘downstream’ sector, has fundamentally different demands for upstream chips. The ‘independent functionality’ of chips may be significantly weakened, especially at this stage, as chips must be redefined and optimized alongside software and algorithms. NVIDIA’s GB200 and Huawei’s Matrix 384 demonstrate a clear logic: raw computing power (Flops) is no longer the sole constraint for AI hardware infrastructure. Whether in pre-training or post-training and inference scaling, balancing network, latency, computation, and power consumption is now more critical—this is the essence of ‘system definition capability.’ For domestic AI, decoupling computing, storage, and data transfer might be the correct path. NVIDIA’s NVLink Fusion reflects the same industry trend, aiming to position NVLink as the ‘CUDA’ of hardware systems and establish it as an industry standard. CUDA + NVLink continues to lock in the definition rights of system ecosystems. Last week, Reuters reported that new restricted GPUs might lack HBM, relying solely on GDDR7. For the Chinese market, NVIDIA’s strategy seems to be: if GPUs can’t enter, then separate storage, computing, and data transfer, reduce specifications for computing and storage, and redefine data transfer as a standalone product (it remains unclear whether NVLink Fusion can enter China). Returning to Hygon’s restructuring, from an industrial logic perspective, it’s not too late, but certainly not leading either.Today, I came across an image on social media: even Kunlunxin is working on super nodes… There were rumors before that XX’s single-card performance is indeed strong, and CUDA adaptation is smooth, but there seems to be room for improvement when it comes to cluster networking. Now, a solution has emerged.
On the other hand, due to various national conditions, inefficiency and redundancy have historically existed in China’s semiconductor industry. In contrast, the ‘monopoly’ established by overseas tech giants through market competition may, from the perspective of the entire tech ecosystem, be the most efficient. Focusing solely on AI, the phase of ‘chaotic competition’ from scratch may have passed. Defining new demands with higher efficiency might be more critical for China’s semiconductor and tech industry at this stage. Finally, I hope China’s semiconductor industry can find its own ‘system-level answer’ in the upcoming wave of mergers and acquisitions. The market carries risks, and investments require caution. This article does not constitute personal investment advice nor considers individual users’ specific investment goals, financial situations, or needs. Users should evaluate whether any opinions, views, or conclusions in this article align with their particular circumstances. Invest at your own risk.
