DeepSeek buys 160,000 Huawei chips — but won't train on them

DeepSeek is putting 160,000 Huawei Ascend 950DT accelerators into a roughly 1 GW data center in Ulanqab, Inner Mongolia. At about $16,000 a chip that's ~$2.6 billion of silicon, and one of the largest known Huawei AI clusters anywhere. Bloomberg broke the story on 4 September; DeepSeek wants part of the capacity live by late 2027 or early 2028.

The interesting part is what the chips won't do. Huawei designed and markets the 950DT for training — 144GB HBM, 4TB/s of bandwidth, 2 PFLOPS at FP8 — but DeepSeek plans to run it for inference only. Training stays on Nvidia. The Chinese lab with the strongest possible incentive to go fully domestic is still saying no on the hardest workload.

Supply is the other ceiling. HBM and component shortages hold 950DT output to the low hundreds of thousands this year, and DeepSeek reportedly wants more than Huawei can currently make.

Zoom out and the trend is brutal for Nvidia anyway: Bernstein projects its China AI chip share falling from ~40% in 2025 to ~8% in 2026, and Jensen Huang has largely conceded the market. Huawei's AI chip revenue went $7.5B in 2025 to a projected ~$12B in 2026 — a domestic industry at scale, not a niche supplier.

Why it matters: China can now serve frontier models on home silicon at gigawatt scale — but training is still Nvidia's moat, and everyone knows it.