Chinese AI company DeepSeek is planning a data center built deliberately without Nvidia hardware: at least 160,000 chips from domestic manufacturer Huawei’s Ascend 950DT line are set to run AI models there in day-to-day operation. The site in Inner Mongolia would become one of the largest known clusters of Chinese AI accelerators — and a visible sign that US export controls are accelerating, rather than blocking, exactly what they were meant to prevent.
Key takeaways
- DeepSeek is ordering at least 160,000 Huawei Ascend 950DT chips for a new data center near Ulanqab, Inner Mongolia, roughly 350 kilometers northwest of Beijing.
- The chips will run inference only — executing already-trained models — while DeepSeek continues to rely on Nvidia hardware for the more compute-intensive training process.
- At full capacity, the data center is expected to draw about one gigawatt, enough power for roughly 750,000 homes, drawing on the region’s cheap wind and solar energy.
- Huawei’s production bottlenecks in memory chips could delay full delivery of the order by more than a year; full operation is not expected before late 2027.
- Nvidia CEO Jensen Huang has acknowledged that his company has largely lost the Chinese market to Huawei, with Nvidia’s local share falling from roughly 40 to about 8 percent.
A calculated split
The division of labor is strategic. For training new models, which requires enormous compute working in concert over weeks, DeepSeek is sticking with Nvidia chips for now, wherever it can obtain them through indirect channels. For inference — answering millions of daily user queries — the Huawei chips are sufficient and easier to parallelize. Each Ascend 950DT chip carries 144 gigabytes of high-bandwidth memory (HBM) and 4 terabytes per second of memory bandwidth, at an estimated cost of about $16,000 per unit. The choice of Ulanqab is no accident: besides cheap wind power, the region offers a cool climate averaging 4.3 degrees Celsius, cutting the cost of cooling server halls — though the project will compete there with the region’s power-hungry bitcoin miners for the same renewable capacity.
Why now
The timing is no coincidence. US export controls have restricted Chinese companies’ access to the most capable Nvidia chips for years. Rather than slowing Chinese AI development, industry observers say this has mainly strengthened the domestic competitor: Huawei’s share of the Chinese market has reportedly climbed to around 50 percent, while Nvidia now holds only a small fraction of its former business there. Much like the debate around Nvidia’s new tool for distributed home-network AI, this shows Nvidia searching for new ways to stay relevant even as it steadily loses ground in one of the world’s largest AI markets.
The bottleneck is memory, not compute
Despite the ambitious plans, delivery capacity is limited. Industry estimates suggest Huawei cannot fully ship the 160,000-chip order for more than a year — the constraint lies not in chip fabrication itself but in the high-bandwidth memory those chips require. China’s leading memory maker, CXMT, has only recently begun producing HBM3E for the first time and remains an estimated three to five years behind established suppliers Samsung, SK Hynix, and Micron. That dependence on an immature domestic memory industry will likely determine how quickly China can actually make its inference infrastructure independent of Western hardware, regardless of how many chips Huawei promises on paper.
What this means for the global AI race
Over the next 18 months, a geographic split is taking shape: US providers continue expanding Nvidia-powered data centers, often exceeding 100,000 GPUs each, while China pursues a hybrid approach — Nvidia hardware where available for training, domestic Huawei chips for production inference, plus the aggressive release of open model weights that has already made DeepSeek’s name internationally. For chip sovereignty, this is a significant test case: if the transition succeeds despite the memory bottleneck, it would show that technological containment through sanctions can be worked around given enough time and state commitment.
Outlook
The planned Inner Mongolia data center is more than a technical footnote in the chip industry — it is a litmus test for how effective US export controls really are in the long run. In the short term, they are clearly slowing China’s access to cutting-edge technology. In the medium term, though, they appear to be strengthening exactly the domestic chip industry they were meant to weaken. Whether Huawei can deliver the promised 160,000 chips before its own memory industry catches up technologically will be the defining question of the next year and a half.