DeepSeek v4 Will Run Entirely on Huawei Chips — A Landmark Moment for China's AI Sovereignty Push

DeepSeek, the Chinese AI research lab that stunned the global AI community in early 2025 with a series of models that matched or exceeded Western frontier performance at dramatically lower cost, is preparing a v4 model that will reportedly rely entirely on Huawei's Ascend AI chips — eliminating any dependency on Nvidia hardware that is subject to US export controls.

Why This Matters

The ability to train and run frontier AI models without Nvidia A100 or H100-class chips has been the central technical question hanging over China's AI development trajectory since 2022. US export controls have successively tightened the hardware available to Chinese AI developers, with the explicit policy goal of creating a capability ceiling that would prevent China from reaching or maintaining frontier AI performance.

DeepSeek's earlier models — particularly DeepSeek-R1 and the V2 series — were trained on restricted but still-available older Nvidia hardware. The move to Huawei's Ascend 910B and the newer Ascend 910C chips for v4 would represent a clean break from that dependency, suggesting that Huawei's chips have reached a level of maturity and performance that makes them viable for frontier-scale training runs.

Huawei's Ascend Progress

Huawei has been investing heavily in its Ascend AI accelerator line as a strategic priority since 2019, when it was first placed on the US Entity List. The Ascend 910B — which entered broader production in 2023 — delivers performance that independent benchmarks place at roughly 60-70% of an Nvidia A100 on standard transformer workloads. The 910C, which began shipping to select customers in late 2025, is reported to close that gap further.

The challenge has never been single-chip performance alone — it has been software ecosystem maturity, multi-chip interconnect efficiency, and the ability to orchestrate thousands of chips in a training cluster with the reliability that frontier runs require. DeepSeek's reported decision to use Huawei chips for v4 suggests the team believes these infrastructure challenges have been solved sufficiently for a production training run.

The Policy Implications

The irony of the export control regime is increasingly visible: by cutting off Chinese AI developers from Nvidia hardware, US policy has created the commercial pressure that made Huawei's chip program a national priority. A DeepSeek v4 trained entirely on Huawei hardware would be cited by Chinese technology nationalists as proof that the strategy has backfired — and it would provide Huawei with a marquee customer reference that accelerates the chip's adoption across the Chinese AI ecosystem.