DeepSeek orders 160,000 Huawei chips, and still trains on Nvidia

DeepSeek is preparing to order at least 160,000 Huawei Ascend 950DT accelerators for a new data centre in Inner Mongolia, which would make the largest known cluster of Huawei AI chips anywhere. The sentence that changes the headline sits a paragraph later: the company plans to run its models on them, not train them.

What the DeepSeek Huawei chips order buys

Training stays on Nvidia. DeepSeek has tried training on Huawei silicon before and went back, and Huawei built the 950DT for exactly that harder job. The published specification lists 144 GB of the company's own HBM, 4 TB/s of memory bandwidth and a 2 TB/s interconnect, and Huawei markets the part for training and decoding both.

The scale is easier to picture in Huawei's own units. It sells these chips in SuperPoD blocks of 8,192, each filling 160 cabinets across roughly 1,000 square metres. An order of 160,000 is about twenty of those blocks.

“But we have the capability to build SuperPoDs and SuperClusters, which are the source of our confidence.”

— Eric Xu, Rotating Chairman of Huawei, Huawei Connect, September 2025

Quote source: Huawei Connect, September 2025

Memory is the bottleneck

Delivery is the part nobody controls. Bloomberg first reported the order, and filling it could take more than a year, because shortages of top-end memory hold 950DT output to the low hundreds of thousands of units this year. China's CXMT is producing HBM3E in small batches and sits three to five years behind Samsung, SK Hynix and Micron.

That is the whole story of Chinese AI hardware in one component. The logic die is solvable at home. The memory stacked on top of it is not yet, and a training run punishes memory bandwidth in a way that serving a finished model does not. Splitting the work by hardware is the rational response to that gap rather than a statement of preference.

Capacity, not independence

So the order buys capacity, not independence. Every query a Chinese user sends to a DeepSeek model could soon run on domestic silicon, while the model those queries hit was still shaped on hardware Washington controls the export of. The dependency moves upstream and gets harder to see.

It also fits the pattern we have been tracking. DeepSeek turned its founder into the richest AI creator in the world by shipping models cheaply, and free Chinese models are already the argument some investors use when they question US valuations.

The date to watch is the first frontier model trained end to end on domestic chips. Until that happens, a 160,000-chip cluster in Inner Mongolia is a very large serving fleet with an imported engine.

This article is for informational purposes only and does not constitute investment advice.