DeepSeek is planning a 160,000-chip Huawei Ascend-950DT inference cluster at an Inner Mongolia data centre — the largest known Huawei chip deployment if completed. Full order delivery is projected to take more than one year due to production constraints. Training workloads continue to run on Nvidia hardware. The inference/training split is the data point that defines the actual state of Chinese AI chip independence: Huawei's Ascend chips are capable enough for inference — serving existing models to users at scale — but remain insufficient for the compute-intensive forward and backward passes of frontier model training. The primary bottleneck is memory. China's CXMT is producing HBM3E in limited quantities but remains three to five years behind Samsung, SK Hynix, and Micron, which already mass-produce HBM4. Until CXMT closes that gap, Chinese AI labs will continue training on foreign silicon regardless of political pressure or government procurement mandates. The 160,000-chip inference cluster is a meaningful demonstration of domestic capability at scale. It is not evidence of the supply chain independence that the domestic chip programme is nominally designed to achieve.