DeepSeek Plans 160,000-Chip Huawei Cluster in Inner Mongolia

DeepSeek is ordering at least 160,000 Huawei Ascend 950DT accelerators for a new gigawatt-scale data center in Ulanqab, Inner Mongolia, according to a Bloomberg report published September 4. If the plan holds, it would be one of the largest known clusters of Chinese-made AI chips anywhere. The sourcing is anonymous, neither company has commented, and the chip in question has not yet shipped in volume. Treat it as a well-detailed report, not a confirmed deal.

Inference, not training

The detail that matters most is what the chips will do. Huawei designed and markets the 950DT as a training part. DeepSeek, per the report, intends to use them for inference, the work of serving models to users, and will keep training on whatever Nvidia hardware it can still obtain. Two readings are available: DeepSeek does not yet trust domestic silicon for training runs, or its inference demand has grown large enough to justify dedicating a new facility to it. Both can be true at once. Either way, the Huawei order adds to DeepSeek's existing Nvidia H800 and H20 fleet rather than replacing it.

Timeline and the real bottleneck

The facility is expected to come partially online in late 2027 or early 2028, and fulfilling the full order could take more than a year. The constraint is Huawei's production capacity: the 950DT is expected to launch in the fourth quarter of 2026, so DeepSeek is reserving 160,000 units of a chip that does not yet exist at scale. Ulanqab, about 350 kilometers northwest of Beijing, is a cold-climate city with cheap wind power that already anchors part of China's data center belt.

The policy backdrop explains the move. Nvidia's H100 and A100 have been restricted from China since October 2022. The H20 was blocked from April to August 2025, then partially allowed back with a reported 15 percent cut of China revenue going to the US government. Beijing, in turn, barred state-funded data centers from buying foreign AI chips in autumn 2025. Some secondary coverage headlines this as Nvidia "giving up on China"; that is commentary layered on top of Bloomberg's reporting, which does not say so.

What it means outside China

DeepSeek is a popular choice for startups worldwide because its API is cheap. Building inference capacity on domestic chips is how those prices stay sustainable under export controls, even if the payoff is two years out. The second lesson is about supply: one of the fastest-growing model providers in the world has decided that depending on a single chip vendor is a risk worth a gigawatt to hedge. Anyone whose product runs on one cloud and one model API might read that as advice.

Sources: Bloomberg, Seeking Alpha, Tech-ish