Huawei's Ascend 950 supernode now claims support for DeepSeek V4 inference. [1] The claim moves the story past whether export controls are circumvented and into whether the domestic compute stack can carry real demand at the application layer.

The paper yesterday read V4 on Ascend as breaking the export-control thesis in eight days alongside Kimi. Sunday hands the question to developers. CANN, Huawei's CUDA equivalent, has to handle real traffic, real failure modes, and real tooling expectations from a model that just shipped. [2] [3]

The X-side phrasing — "de-CUDAzation" — captures both the ambition and the test. Chinese AI-stack discourse argues V4 running on Ascend would prove that export controls "survive contact with architectural cleverness." It is also the kind of claim developers can falsify quickly. If CANN drops inference, if tooling lags, if cluster utilization underperforms, the falsification will be public within days, not quarters.

The Reuters and CNBC framing of V4 as Chinese self-reliance gets the symbolism right and the operational question wrong. [2] [3] Sovereignty is a posture; throughput is a measurement. Beijing now needs the second to validate the first.

-- DAVID CHEN, Beijing