DeepSeek V4 to Run on Huawei Chips — A Milestone for Fully Independent Chinese AI Stack?

I’ve been waiting for this news for a long time.

DeepSeek V4 — the next generation of China’s strongest open-source large model series — is reportedly going to run on Huawei Ascend chips. If true, this is the most heavyweight validation yet on China’s domestic AI chip track.

First, why does this matter?

Over the past two years, the progress of Chinese AI companies at the model layer is obvious to all. DeepSeek, Qwen, Zhipu — these models can compete head-to-head with GPT and Claude in certain dimensions. But everyone knows a lurking concern: these models were almost all trained on NVIDIA A100/H100. At the chip layer, our lifeline is still in others’ hands.

US export controls are like a sword hanging overhead — chips available today may be unavailable tomorrow.

So DeepSeek choosing to run V4 on Huawei Ascend, whether “forced” or “voluntary,” has the same strategic significance: proving China’s AI stack can be independent of NVIDIA.

From my previous research, there is indeed a performance gap between Huawei Ascend 910B and NVIDIA A100 — roughly 20-30%. But this gap has noticeably narrowed over the past year. And the DeepSeek team has always been strong in training efficiency optimization; they demonstrated the ability to train better models with less compute on V3.

The key isn’t “can Huawei chips catch up to NVIDIA” but “is the gap within acceptable range.” If training V4 on Huawei chips only takes 20% more time without compromising model quality — that math works out.

Of course, there are skeptical voices.

Some say this is “sacrificing efficiency for autonomy,” a politically correct technical choice. Others worry about chip stability — Huawei Ascend’s reliability in large-scale cluster training hasn’t had much public data yet.

These concerns aren’t without merit. But from an industry trend perspective, multi-chip adaptation is already an irreversible direction. Previously, Zhipu used Huawei chips to train GLM-5.1 reaching 94.6% of Claude’s performance, initially proving feasibility. If DeepSeek can repeat or exceed this achievement on V4, then it’s not “feasible” anymore — it’s “mature.”

Honestly, from my personal perspective, the biggest significance of this event isn’t the technology itself, but that it breaks a psychological dependency. For a long time, there was an unspoken consensus in China’s AI circle — “NVIDIA is irreplaceable.” Every time chips were mentioned, people would sigh as if it were an unsolvable problem.

But facts are proving: it’s not unsolvable; it’s being solved.

If DeepSeek V4 performs well enough on Huawei chips, other Chinese AI companies will accelerate their shift to domestic chips. This isn’t just a technology event; it may be the inflection point for the entire domestic AI computing ecosystem.

However, before final results are out, I’ll say this: don’t rush, wait for the data.