Huawei Ascend 910C Chips Complete Full-Parameter Post-Training of DeepSeek V4-Pro (1.6 Trillion Parameters) — China's Domestic AI Hardware Stack Validated at Frontier Training Scale

A research consortium comprising Huawei, Shenzhen Loop Area Institute, Harbin Institute of Technology Shenzhen, and Shenzhen Research Institute of Big Data completed full-parameter post-training of DeepSeek's V4-Pro model entirely on Huawei Ascend 910C chips, the South China Morning Post reported on June 5. The model contains 1.6 trillion parameters — DeepSeek's largest to date — and the post-training run deployed at least 1,000 Huawei Ascend 910C chips. The technical significance is substantial: full-parameter post-training is the most computationally demanding phase of the AI development lifecycle, requiring synchronous gradient updates across the entire model architecture — a fundamentally harder task than inference (running an already-trained model). Until this milestone, Huawei's Ascend chip ecosystem had been validated primarily for AI inference at scale (ByteDance Doubao inference on Ascend; DeepSeek V4-Pro inference on Ascend from May 2026), but not for frontier-model post-training. The completion of full-parameter post-training on domestically produced silicon demonstrates that China now possesses the complete AI development pipeline on domestic hardware: data curation → pre-training → fine-tuning → post-training → inference — all executable without a single Nvidia GPU or Western AI chip. This is the technical closure of the DeepSeek-on-Ascend thesis: the model can now be reproduced, refined, and specialized entirely within China's domestic chip ecosystem. The milestone arrives as DeepSeek finalizes its $7.4B funding round (Bloomberg, June 3) and as the Huawei Ascend 950PR (H2 2026 delivery, ~750K units planned) targets mass-volume AI inference at data center scale. SCMP framed the achievement as 'a major leap for China's AI self-reliance' — the US export control strategy's primary lever (GPU access for frontier model training) has demonstrably lost effectiveness at the level that matters most: cutting-edge model development.