DeepSeek V4 to run fully on Huawei Ascend AI chips

DeepSeek’s decision to move V4 fully onto Huawei’s Ascend 950PR chips is a watershed moment in AI geopolitics. It’s the clearest real‑world example yet of a top‑tier model lab trying to decouple from the US‑centric Nvidia stack, not just for inference but for training. If they pull it off at scale, China demonstrates that US export controls can slow but not prevent frontier‑scale capability, so long as you’re willing to tolerate some efficiency gap and invest heavily in software tooling around domestic silicon.([albis.news](https://www.albis.news/tech/deepseek-v4-huawei-chips-china-ai-independence-2026))

This also puts pressure on US labs: Nvidia’s near‑monopoly on state‑of‑the‑art training hardware has been a de facto throttle on how many independent frontier projects the world can run. A credible alternative path—Ascend plus a China‑centric software ecosystem—means more global experiments, more model diversity, and potentially faster progress as architectural ideas cross‑pollinate.

For the race to AGI, the main implication is not that Huawei will suddenly outgun Nvidia, but that compute is on a path to becoming more multipolar. That reduces the leverage of any single government to slow things down through export controls alone, and raises the odds that US and Chinese labs iterate in parallel. It also increases the strategic value of techniques that squeeze more capability out of constrained hardware, since those will travel across both ecosystems.