Huida CEO Huang Renxun has been involved in a series of controversies recently.
[Reporter Tu Jumin/Taipei Report] Huida CEO Huang Jen-Hsun has recently been involved in the controversy over China's acquisition of advanced AI chips due to his company's business dealings with Singaporean data center operator Megaspeed, and even an extramarital affair. U.S. intelligence agencies suspect that related companies are involved in circumventing export controls. Zhou Guanzhu, an assistant researcher at the Institute of National Security of the National Defense Academy, analyzed that from this incident to the case of former Google engineers stealing technology, it reflects that in the competition for artificial intelligence, China not only faces restrictions on obtaining advanced chips, but also has technical bottlenecks in the integration of large-scale AI data center systems, software ecology, and computing power operations.
On October 29, 2026, the New York Times reported that U.S. intelligence agencies suspected that a former Chinese senior executive of Singapore data center operator Megaspeed was involved in collecting intelligence for China and helping to circumvent export controls to obtain NVIDIA's advanced AI chips. The senior executive had business dealings with NVIDIA's management. Huida stated that the relevant interactions were normal business transactions, and the board of directors has established an independent committee to investigate. The U.S. House of Representatives also asked Huida’s CEO to explain the company’s safety regulations on November 18. This case can be compared with the case of former Google engineer Ding Linwei, who was sentenced in January this year. According to the prosecution content of the US Department of Justice, the object of the theft was the software and hardware platform that enabled the Google data center to train and serve large models. The subject matter of both cases is not the chip design itself: the former involves the acquisition of computing power, and the latter involves the operational capabilities of the AI data center.
Looking at technical bottlenecks from technology theft actions
China does not lack data center experience. ByteDance, Alibaba and other companies have used Huida chips to operate large AI clusters. What China really lacks is experience in operating AI data centers with non-CUDA systems: after a comprehensive replacement of chips, interconnects, and software ecosystems, scheduling, fault tolerance, and performance tuning must all be accumulated again. With the rise of large language models in 2022, the demand for AI computing power has expanded dramatically. In 2026, the China Academy of Information and Communications Technology admitted that cross-regional computing power deployment still faces multiple constraints. So far, China's AI models are mostly built on Huida chips and the CUDA (Compute Unified Device Architecture) ecosystem. According to reports, DeepSeek switched to using Huawei Ascend to train the R2 model in 2025. Even though Huawei engineers were stationed, it was still unable to complete the training, and eventually returned to Huida for training. Ascend was only used for inference. Since the computing performance of a single chip of China's self-made chips is relatively low, more chips must be used to form a chip cluster to make up for the lack of performance. However, as the cluster expands, the difficulty of scheduling and fault tolerance increases, and the chip disadvantages are eventually transferred to the system operation burden.
Cutting-edge pre-training may still require Huida GPUs and CUDA software development kits, but under U.S. controls, Chinese companies have turned to overseas data centers to obtain computing power. Megaspeed was originally an overseas subsidiary invested and established by China Seventh Avenue Company. After multiple equity transfers, the ownership of shareholders is still controversial. This case shows that Chinese companies may obtain Huida’s high-end chips and computing power through overseas affiliates to solve the problem of short-term computing power shortage.
The Ding Linwei case corresponds to a long-term solution. According to the indictment of the U.S. Department of Justice, the business secrets it stole include Google’s self-developed Tensor Processing Unit (TPU) system, inter-chip communication software, and software for scheduling thousands of chips into supercomputers. When Ding Linwei applied for the China Talent Program, he also stated that he would help China develop computing infrastructure capabilities that are comparable to international standards. This kind of knowledge is highly dispersed within the organization and difficult to reconstruct from documents alone, so it needs to be obtained through multiple channels. US export controls are based on hardware specifications and single entities. What China lacks is knowledge that is dispersed, tacit and dependent on the operational process. If the computing power is dispersed to multiple entities and countries, each of which is below the control threshold, and then integrated on the user end, the current mechanism will be difficult to detect.
Impact on my country’s AI industry
First of all, the focus of the US control may include hardware expansion and cloud user review, technical support services and personnel exchanges. Second, the key to judging whether China can get rid of its dependence on Huida is whether it can pre-train cutting-edge models from scratch with self-made chip clusters; until then, China's demand for overseas computing power and operational knowledge will not decrease. Third, Taiwan’s original design manufacturer (ODM) is responsible for the assembly and integration of the whole cabinet system. Taipei is also an important place for exchanges among industry professionals. Taiwan should strengthen the review of substantial beneficiaries of third-country customers and deepen intelligence sharing with the United States on export control law enforcement.
