The global AI arms race has shifted decisively from the realm of algorithmic innovation to the availability of raw compute. Under the weight of stringent sanctions, China is pursuing total independence from U.S. silicon, elevating the search for alternatives to a matter of national security. The massive DeepSeek project in Inner Mongolia has become a focal point of this transition, serving as a litmus test for the capabilities of domestic manufacturing. The deployment of hundreds of thousands of homegrown accelerators is designed to prove that the Nvidia ecosystem is no longer the sole gateway to developing frontier AI models.
DeepSeek’s drive toward technological sovereignty is manifesting in the physical realm through the construction of a monolithic data center in Inner Mongolia. The scale of the company's ambitions is staggering: the infrastructure is slated to house no fewer than 160,000 of Huawei's latest accelerators. This project is poised to become one of the largest clusters powered by Chinese hardware in history, effectively establishing an alternative computational bedrock for the evolution of generative AI.
The centerpiece of this architecture will be the next-generation Ascend 950DT accelerators. In terms of raw technical specifications, these chips are approaching the performance levels of Nvidia’s Hopper architecture, making them a viable tool for handling heavy-duty models. However, DeepSeek’s current strategy remains pragmatic: the company intends to deploy the Ascend 950DT primarily for inference—the stage of processing queries and generating responses. Model training remains a more formidable challenge. DeepSeek’s previous experience suggests that Huawei's hardware cannot yet fully displace Nvidia's solutions, which continue to offer superior stability and efficiency when processing massive-scale datasets.
Despite these ambitions, the project faces severe manufacturing bottlenecks. Even as a technological titan, Huawei is constrained by a critical shortage of specialized high-bandwidth memory. Current production capacities allow for only a few hundred thousand units per year, forcing the company to implement strict allocation quotas among key clients and fulfill existing international obligations. Consequently, the full deployment of DeepSeek’s data center could stretch beyond a year.
Within the broader context of the Chinese market, there is a noticeable divergence in strategies regarding "de-Nvidia-ization." On one hand, we see successes from players like Meituan, which has already trained a model using 50,000 Chinese accelerators, or Z.ai, which established a 1GW data center dedicated exclusively to local silicon. On the other hand, giants like Alibaba and Tencent continue to bet on Nvidia, striving to maximize every available supply channel. Many firms are bypassing sanctions by leasing computational power via offshore cloud services, allowing them to remain at the cutting edge of innovation without formally violating regulatory requirements.
The energy profile of DeepSeek’s new facility underscores the true magnitude of the endeavor. The data center's power consumption will be measured in several gigawatts—comparable to the energy requirements of a city with a population of approximately 750,000 people. Furthermore, the 160,000 Ascend 950DT chips will occupy only a fraction of the available capacity, leaving significant headroom for future expansion.
The central paradox of this situation is that even in the pursuit of strategic autonomy, the temptation to utilize the best available tools remains overwhelming. Unofficial reports suggest that DeepSeek may have acquired advanced Nvidia Blackwell accelerators despite all existing restrictions. This highlights the fundamental conflict defining the modern AI industry: the necessity of political independence clashing with the technological imperative to employ the most powerful hardware humanity has engineered to avoid falling behind in the global intelligence race.
Nvidia N1X and the New Era of Arm PCs
